{"generated": 1794569602.4192011, "objective": "agillm-fit-v1", "weights": {"quality": 0.55, "scale": 0.15, "stability": 0.1, "long_context": 0.1, "serving": 0.1}, "base": {"curve": [{"step": 150, "tokens": 307200, "ar": 6.221307754516602, "sat": 6.236490925153096, "nat": 6.262467126051585, "ctx_gain": -0.03959238529205322, "ctx_full": 6.169127106666565, "ctx_256": 6.129534721374512}, {"step": 300, "tokens": 614400, "ar": 6.005859096844991, "sat": 6.033294936021169, "nat": 6.078488290309906, "ctx_gain": -0.16538389523824057, "ctx_full": 6.109192967414856, "ctx_256": 5.943809072176616}, {"step": 600, "tokens": 1228800, "ar": 5.663718183835347, "sat": 5.773470600446065, "nat": 5.995160281658173, "ctx_gain": -0.2902825673421224, "ctx_full": 5.877588669459025, "ctx_256": 5.587306102116902}, {"step": 1200, "tokens": 2457600, "ar": 5.245898425579071, "sat": 5.46754531065623, "nat": 5.843253115812938, "ctx_gain": -0.30009488264719647, "ctx_full": 5.432854731877645, "ctx_256": 5.132759849230449}], "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 102158627, "kv_bytes_per_token": 24903680}, "ms_per_step": 194.1464344660441, "seeds": 3, "ld_spike": 0.0, "ld_mean": 6.25246}, "table": [{"kind": "control", "id": "parents_only", "status": "done", "title": "Parents only (no copy depth)", "slot": "control", "one_line": "Control: the root blocks alone, one visit each (the 80-block exit): what the copy depth must beat at equal box time", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 3, "ar_sd": 0.010066913061242098, "ar": 5.1731, "sat": 5.4104, "nat": 5.824, "ctx_gain": -0.2751, "ld_spike": 0.0, "ld_mean": 5.937341666666666, "ms_per_step": 65.26656945546468, "touched_frac": 0.9760784313725482, "census": {"latent_bytes": 126701535232, "table_bytes": 1638924288, "n_matrices": 552, "fp_params": 21671492, "kv_bytes_per_token": 5242880}, "flip_pct_end": 0.10742049081508089, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.209012985229492, "sat": 6.2188312609990435, "nat": 6.242824951807658, "ctx_gain": -0.027688980102539062, "ctx_full": 6.175889015197754, "ctx_256": 6.148200035095215}, {"step": 300, "tokens": 614400, "ar": 5.975233693917592, "sat": 6.009052614370982, "nat": 6.071708460648854, "ctx_gain": -0.15725767612457275, "ctx_full": 6.069012641906738, "ctx_256": 5.9117549657821655}, {"step": 600, "tokens": 1228800, "ar": 5.57152924935023, "sat": 5.707779367764791, "nat": 5.96176294485728, "ctx_gain": -0.2533593972524007, "ctx_full": 5.73086420694987, "ctx_256": 5.477504809697469}, {"step": 1200, "tokens": 2457600, "ar": 5.1731487313906355, "sat": 5.410383343696594, "nat": 5.823984483877818, "ctx_gain": -0.27506013711293537, "ctx_full": 5.3479394515355425, "ctx_256": 5.072879314422607}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 461, "fit": 85.13, "quality": 90.1, "scale": 100.0, "stability": 53.2, "long_context": 52.5, "serving": 100.0, "d_ar": 0.4906, "d_sat": 0.3631, "d_nat": 0.1712, "quality_delta": 0.4012, "byte_ratio": 0.229, "ms_ratio": 0.336, "step_ratio": 0.256, "tokens_equal_time": 4802188, "kv_ratio": 0.211, "storage_ratio": 0.211, "gain_gap": 0.025}, {"kind": "control", "id": "fp_dense", "status": "done", "title": "FP32 dense control", "slot": "control", "one_line": "Ablation: fp32 weights and a single dense expert: shows what binary + top-1 cost in this regime", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 2, "ar_sd": 0.009457457777905045, "ar": 4.9509, "sat": 5.2616, "nat": 5.7768, "ctx_gain": -0.2327, "ld_spike": 0.0, "ld_mean": 6.2601175, "ms_per_step": 137.77577877044678, "touched_frac": 0.0, "census": {"latent_bytes": 306016419840, "table_bytes": 1638924288, "n_matrices": 1520, "fp_params": 307712364284, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.0, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 5.858580529689789, "sat": 5.955559223890305, "nat": 6.043617069721222, "ctx_gain": -0.21442127227783203, "ctx_full": 5.918464720249176, "ctx_256": 5.704043447971344}, {"step": 300, "tokens": 614400, "ar": 5.529640555381775, "sat": 5.701066941022873, "nat": 6.006625384092331, "ctx_gain": -0.18388521671295166, "ctx_full": 5.61867618560791, "ctx_256": 5.4347909688949585}, {"step": 600, "tokens": 1228800, "ar": 5.242878347635269, "sat": 5.4810503125190735, "nat": 5.892777442932129, "ctx_gain": -0.2111724615097046, "ctx_full": 5.303431689739227, "ctx_256": 5.092259228229523}, {"step": 1200, "tokens": 2457600, "ar": 4.95093110203743, "sat": 5.261553019285202, "nat": 5.776831150054932, "ctx_gain": -0.23271197080612183, "ctx_full": 5.053897976875305, "ctx_256": 4.821186006069183}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 606, "fit": 81.83, "quality": 100.0, "scale": 74.4, "stability": 49.9, "long_context": 56.7, "serving": 50.0, "d_ar": 0.6342, "d_sat": 0.4528, "d_nat": 0.1871, "quality_delta": 0.5085, "byte_ratio": 0.567, "ms_ratio": 0.71, "step_ratio": 0.603, "tokens_equal_time": 2039088, "kv_ratio": 1.0, "storage_ratio": 0.508, "gain_gap": 0.0674}, {"kind": "native", "id": "markov_corpus_prior", "status": "done", "title": "Counted Markov chain beside the transformer: n-gram log-probabilities added to the logits", "slot": "hybrid", "one_line": "Every readout adds lam * log P1 (unigram); every causal slot and the NAT slots just after the prefix add lam * log(P_chain / back-off) from n-gram and skip-gram counts taken on text the run never trains on; 10 fp parameters, no extra latents.", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 3, "ar_sd": 0.007609442470854705, "ar": 4.5016, "sat": 5.0563, "nat": 5.823, "ctx_gain": -0.0746, "ld_spike": 0.0, "ld_mean": 6.0889016666666675, "ms_per_step": 216.78702036539713, "touched_frac": 0.9067460317460302, "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 102158637, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.037007918676326475, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 4.698184212048848, "sat": 5.255873044331868, "nat": 6.0985014239947, "ctx_gain": -0.0021279454231262207, "ctx_full": 4.524677832921346, "ctx_256": 4.5225498874982195}, {"step": 300, "tokens": 614400, "ar": 4.635007997353871, "sat": 5.220865547657013, "nat": 6.060257116953532, "ctx_gain": -0.003473182519276937, "ctx_full": 4.464013914267222, "ctx_256": 4.460540731747945}, {"step": 600, "tokens": 1228800, "ar": 4.577940781911214, "sat": 5.15868753194809, "nat": 5.9669837554295855, "ctx_gain": -0.038190980752309166, "ctx_full": 4.452535808086395, "ctx_256": 4.414344827334086}, {"step": 1200, "tokens": 2457600, "ar": 4.501637578010559, "sat": 5.056338210900624, "nat": 5.823014338811238, "ctx_gain": -0.07458305358886719, "ctx_full": 4.425067226092021, "ctx_256": 4.350484172503154}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 559, "fit": 79.92, "quality": 100.0, "scale": 50.0, "stability": 51.6, "long_context": 72.6, "serving": 50.0, "d_ar": 1.0828, "d_sat": 0.6115, "d_nat": 0.0232, "quality_delta": 0.7766, "byte_ratio": 1.011, "ms_ratio": 1.117, "step_ratio": 1.037, "tokens_equal_time": 1184480, "kv_ratio": 1.0, "storage_ratio": 1.0, "gain_gap": 0.2255, "rank": 1}, {"kind": "native", "id": "dlgm458_lg", "status": "done", "title": "DualScaleMemoryRouter #458 (DLGM) hybrid, learned 32-dim router projection with per-channel kernels", "slot": "hybrid", "one_line": "DLGM layers (local+mid causal conv, delta memory, identity, token/head softmax router) on 3/4 of layers; attention on every 4th", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 3, "ar_sd": 0.04161016751290842, "ar": 5.0904, "sat": 5.3723, "nat": 5.8069, "ctx_gain": -0.0585, "ld_spike": 0.0, "ld_mean": 6.34036, "ms_per_step": 321.2440808614095, "touched_frac": 0.9080873015873007, "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 299551907, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.07735225324446286, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.204118728637695, "sat": 6.221837997436523, "nat": 6.266959985097249, "ctx_gain": -0.006938457489013672, "ctx_full": 6.118436376253764, "ctx_256": 6.1114979187647505}, {"step": 300, "tokens": 614400, "ar": 5.871767481168111, "sat": 5.9520871837933855, "nat": 6.068025688330333, "ctx_gain": -0.043360511461893715, "ctx_full": 5.821094870567322, "ctx_256": 5.777734359105428}, {"step": 600, "tokens": 1228800, "ar": 5.472393174966176, "sat": 5.660475571950276, "nat": 5.9412545164426165, "ctx_gain": -0.04331545035044352, "ctx_full": 5.398481885592143, "ctx_256": 5.355166435241699}, {"step": 1200, "tokens": 2457600, "ar": 5.090413510799408, "sat": 5.372284630934398, "nat": 5.806870261828105, "ctx_gain": -0.05846289793650309, "ctx_full": 5.018206357955933, "ctx_256": 4.959743460019429}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 545, "fit": 56.38, "quality": 57.4, "scale": 50.0, "stability": 49.1, "long_context": 74.2, "serving": 50.0, "d_ar": 0.0994, "d_sat": 0.0458, "d_nat": 0.0247, "quality_delta": 0.0737, "byte_ratio": 1.013, "ms_ratio": 1.655, "step_ratio": 1.173, "tokens_equal_time": 1047528, "kv_ratio": 1.0, "storage_ratio": 1.0, "gain_gap": 0.2416, "rank": 2}, {"kind": "native", "id": "fusion879_lg", "status": "done", "title": "FusionGatedFIRNet #879 (Fusion) hybrid, learned 32-dim gate projection with per-channel kernels", "slot": "hybrid", "one_line": "Fusion layers (short+long causal FIR, delta memory with per-head retention, value; an independent sigmoid gate per path, token and head) on 3/4 of layers; attention on every 4th", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 3, "ar_sd": 0.027074486603129874, "ar": 5.0977, "sat": 5.3826, "nat": 5.8037, "ctx_gain": -0.0953, "ld_spike": 0.0, "ld_mean": 6.313475, "ms_per_step": 280.0167401631673, "touched_frac": 0.9064841269841265, "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 369629987, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.07698433881585297, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.239956319332123, "sat": 6.244104345639546, "nat": 6.28300021092097, "ctx_gain": -0.0011264880498250325, "ctx_full": 6.171831250190735, "ctx_256": 6.17070476214091}, {"step": 300, "tokens": 614400, "ar": 5.931730846563975, "sat": 6.000961681207021, "nat": 6.100297451019287, "ctx_gain": -0.035134593645731606, "ctx_full": 5.868620594342549, "ctx_256": 5.833486000696818}, {"step": 600, "tokens": 1228800, "ar": 5.49220730861028, "sat": 5.6747239629427595, "nat": 5.959956347942352, "ctx_gain": -0.061307152112325035, "ctx_full": 5.429995536804199, "ctx_256": 5.3686883846918745}, {"step": 1200, "tokens": 2457600, "ar": 5.097694794336955, "sat": 5.382589836915334, "nat": 5.803740322589874, "ctx_gain": -0.09531323115030925, "ctx_full": 5.054284850756328, "ctx_256": 4.958971619606018}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 546, "fit": 55.94, "quality": 57.2, "scale": 50.0, "stability": 49.4, "long_context": 70.5, "serving": 50.0, "d_ar": 0.1004, "d_sat": 0.046, "d_nat": 0.0125, "quality_delta": 0.072, "byte_ratio": 1.011, "ms_ratio": 1.442, "step_ratio": 1.119, "tokens_equal_time": 1098498, "kv_ratio": 1.0, "storage_ratio": 1.0, "gain_gap": 0.2048, "rank": 3}, {"kind": "native", "id": "fusion879_hn", "status": "done", "title": "FusionGatedFIRNet #879 (Fusion) hybrid, per-head output norm", "slot": "hybrid", "one_line": "Fusion layers (short+long causal FIR, delta memory with per-head retention, value; an independent sigmoid gate per path, token and head) on 3/4 of layers; attention on every 4th", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 3, "ar_sd": 0.009057032181963685, "ar": 5.083, "sat": 5.3705, "nat": 5.8011, "ctx_gain": -0.0613, "ld_spike": 0.0, "ld_mean": 6.327065, "ms_per_step": 317.21266110738117, "touched_frac": 0.9078650793650785, "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 41081291747, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.07745162519236859, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.2042485276858015, "sat": 6.221100330352783, "nat": 6.264114260673523, "ctx_gain": -0.01301733652750651, "ctx_full": 6.135317802429199, "ctx_256": 6.122300465901692}, {"step": 300, "tokens": 614400, "ar": 5.891513546307881, "sat": 5.974536240100861, "nat": 6.084248801072438, "ctx_gain": -0.028748154640197754, "ctx_full": 5.807206352551778, "ctx_256": 5.77845819791158}, {"step": 600, "tokens": 1228800, "ar": 5.4885713656743365, "sat": 5.670001606146495, "nat": 5.941321233908336, "ctx_gain": -0.03325875600179037, "ctx_full": 5.3902008930842085, "ctx_256": 5.3569421370824175}, {"step": 1200, "tokens": 2457600, "ar": 5.083030541737874, "sat": 5.370459159215291, "nat": 5.801086604595184, "ctx_gain": -0.06132117907206217, "ctx_full": 5.013983448346456, "ctx_256": 4.952662269274394}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 565, "fit": 55.75, "quality": 56.2, "scale": 50.0, "stability": 49.3, "long_context": 73.9, "serving": 50.0, "d_ar": 0.085, "d_sat": 0.0354, "d_nat": 0.0219, "quality_delta": 0.0625, "byte_ratio": 1.012, "ms_ratio": 1.634, "step_ratio": 1.168, "tokens_equal_time": 1052352, "kv_ratio": 1.0, "storage_ratio": 1.0, "gain_gap": 0.2388, "rank": 4}, {"kind": "native", "id": "fusion879", "status": "done", "title": "FusionGatedFIRNet #879 (Fusion) hybrid", "slot": "hybrid", "one_line": "Fusion layers (short+long causal FIR, delta memory with per-head retention, value; an independent sigmoid gate per path, token and head) on 3/4 of layers; attention on every 4th", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 3, "ar_sd": 0.010029003670142577, "ar": 5.0718, "sat": 5.3597, "nat": 5.7861, "ctx_gain": -0.082, "ld_spike": 0.0, "ld_mean": 6.3521350000000005, "ms_per_step": 370.39339542388916, "touched_frac": 0.9077380952380943, "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 41081291747, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.07852571169033573, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.240574141343434, "sat": 6.25745431582133, "nat": 6.300685842831929, "ctx_gain": -0.00818932056427002, "ctx_full": 6.161823908487956, "ctx_256": 6.153634587923686}, {"step": 300, "tokens": 614400, "ar": 5.894458373387654, "sat": 5.978301425774892, "nat": 6.100055595239003, "ctx_gain": -0.028320590655008953, "ctx_full": 5.825304269790649, "ctx_256": 5.79698367913564}, {"step": 600, "tokens": 1228800, "ar": 5.476177453994751, "sat": 5.666662534077962, "nat": 5.948780755201976, "ctx_gain": -0.040674050649007164, "ctx_full": 5.379034002621968, "ctx_256": 5.338359951972961}, {"step": 1200, "tokens": 2457600, "ar": 5.071844935417175, "sat": 5.359661897023519, "nat": 5.786104261875153, "ctx_gain": -0.08195980389912923, "ctx_full": 5.018953561782837, "ctx_256": 4.936993757883708}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 565, "fit": 54.18, "quality": 53.8, "scale": 50.0, "stability": 49.0, "long_context": 71.8, "serving": 50.0, "d_ar": 0.0597, "d_sat": 0.0115, "d_nat": 0.0001, "quality_delta": 0.0381, "byte_ratio": 1.012, "ms_ratio": 1.908, "step_ratio": 1.236, "tokens_equal_time": 994139, "kv_ratio": 1.0, "storage_ratio": 1.0, "gain_gap": 0.2181, "rank": 5}, {"kind": "native", "id": "dlgm458_hn", "status": "done", "title": "DualScaleMemoryRouter #458 (DLGM) hybrid, per-head output norm", "slot": "hybrid", "one_line": "DLGM layers (local+mid causal conv, delta memory, identity, token/head softmax router) on 3/4 of layers; attention on every 4th", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 3, "ar_sd": 0.005705805855826667, "ar": 5.1091, "sat": 5.3895, "nat": 5.8113, "ctx_gain": -0.0732, "ld_spike": 0.0, "ld_mean": 6.303951666666666, "ms_per_step": 361.5771532058716, "touched_frac": 0.9105476190476184, "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 39820506467, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.07517576195184056, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.193723181883494, "sat": 6.218785762786865, "nat": 6.268034756183624, "ctx_gain": -0.008160948753356934, "ctx_full": 6.128138184547424, "ctx_256": 6.119977235794067}, {"step": 300, "tokens": 614400, "ar": 5.894122282663981, "sat": 5.970696071783702, "nat": 6.083240191141765, "ctx_gain": -0.0381018320719401, "ctx_full": 5.83880631128947, "ctx_256": 5.800704479217529}, {"step": 600, "tokens": 1228800, "ar": 5.512416700522105, "sat": 5.685165425141652, "nat": 5.957382162412007, "ctx_gain": -0.048194567362467446, "ctx_full": 5.440674026807149, "ctx_256": 5.392479459444682}, {"step": 1200, "tokens": 2457600, "ar": 5.109130680561066, "sat": 5.389476478099823, "nat": 5.811307787895203, "ctx_gain": -0.07320030530293782, "ctx_full": 5.054995934168498, "ctx_256": 4.98179562886556}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 563, "fit": 53.54, "quality": 52.4, "scale": 50.0, "stability": 49.5, "long_context": 72.7, "serving": 50.0, "d_ar": 0.0386, "d_sat": 0.004, "d_nat": 0.0006, "quality_delta": 0.0241, "byte_ratio": 1.015, "ms_ratio": 1.862, "step_ratio": 1.227, "tokens_equal_time": 1001433, "kv_ratio": 1.0, "storage_ratio": 1.0, "gain_gap": 0.2269, "rank": 6}, {"kind": "control", "id": "no_local", "status": "done", "title": "No local-denoise steps", "slot": "objective", "one_line": "Ablation: drop the 1-in-9 local-denoise step: tests whether those steps cost held-out quality", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 3, "ar_sd": 0.00959113737897984, "ar": 5.1346, "sat": 5.38, "nat": 5.7931, "ctx_gain": -0.297, "ld_spike": 0.0, "ld_mean": null, "ms_per_step": 192.55503018697104, "touched_frac": 0.9765000000000003, "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 102158627, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.07960939666619604, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.214590032895406, "sat": 6.235280195871989, "nat": 6.256255686283112, "ctx_gain": -0.0393599271774292, "ctx_full": 6.17078709602356, "ctx_256": 6.13142716884613}, {"step": 300, "tokens": 614400, "ar": 5.975306371847789, "sat": 6.014833549658458, "nat": 6.063574453194936, "ctx_gain": -0.18256676197052002, "ctx_full": 6.100670576095581, "ctx_256": 5.918103814125061}, {"step": 600, "tokens": 1228800, "ar": 5.54312260945638, "sat": 5.69139301776886, "nat": 5.974994162718455, "ctx_gain": -0.36762388547261554, "ctx_full": 5.796607216199239, "ctx_256": 5.4289833307266235}, {"step": 1200, "tokens": 2457600, "ar": 5.134624679883321, "sat": 5.3799803256988525, "nat": 5.793127119541168, "ctx_gain": -0.2970230182011922, "ctx_full": 5.3259942928949995, "ctx_256": 5.028971274693807}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 539, "fit": 53.44, "quality": 56.2, "scale": 50.0, "stability": 50.0, "long_context": 50.3, "serving": 50.0, "d_ar": 0.0816, "d_sat": 0.0529, "d_nat": 0.0122, "quality_delta": 0.062, "byte_ratio": 1.089, "ms_ratio": 0.992, "step_ratio": 1.065, "tokens_equal_time": 1154295, "kv_ratio": 1.0, "storage_ratio": 1.0, "gain_gap": 0.0031}, {"kind": "native", "id": "dlgm458_lrk", "status": "done", "title": "DualScaleMemoryRouter #458 (DLGM) hybrid, low-rank router with per-channel kernels", "slot": "hybrid", "one_line": "DLGM layers (local+mid causal conv, delta memory, identity, token/head softmax router) on 3/4 of layers; attention on every 4th", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 3, "ar_sd": 0.015771908089404908, "ar": 5.0851, "sat": 5.3723, "nat": 5.8081, "ctx_gain": -0.0706, "ld_spike": 0.0, "ld_mean": 6.349620000000001, "ms_per_step": 411.5292231241862, "touched_frac": 0.9093809523809515, "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 224840867, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.07680072051365272, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.193865716457367, "sat": 6.221475601196289, "nat": 6.265166640281677, "ctx_gain": -0.006899793942769368, "ctx_full": 6.114007870356242, "ctx_256": 6.107108076413472}, {"step": 300, "tokens": 614400, "ar": 5.895945688088735, "sat": 5.974664688110352, "nat": 6.083378314971924, "ctx_gain": -0.037623206774393715, "ctx_full": 5.8374572197596235, "ctx_256": 5.7998340129852295}, {"step": 600, "tokens": 1228800, "ar": 5.4765321016311646, "sat": 5.661115248998006, "nat": 5.943718572457631, "ctx_gain": -0.04329772790273031, "ctx_full": 5.396002689997355, "ctx_256": 5.3527049620946245}, {"step": 1200, "tokens": 2457600, "ar": 5.085072795550029, "sat": 5.372280518213908, "nat": 5.808118621508281, "ctx_gain": -0.0705954631169637, "ctx_full": 5.024581392606099, "ctx_256": 4.953985929489136}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 544, "fit": 53.25, "quality": 51.9, "scale": 50.0, "stability": 49.0, "long_context": 72.9, "serving": 50.0, "d_ar": 0.0329, "d_sat": -0.003, "d_nat": 0.0001, "quality_delta": 0.0192, "byte_ratio": 1.014, "ms_ratio": 2.12, "step_ratio": 1.29, "tokens_equal_time": 952272, "kv_ratio": 1.0, "storage_ratio": 1.0, "gain_gap": 0.2295, "rank": 7}, {"kind": "native", "id": "grid_local16", "status": "done", "title": "Local attention, 16-token window", "slot": "attention", "one_line": "The ar_local16 idea from the old 200-grid, inside the real block: every attention layer sees only the nearest 16 tokens", "novelty": 0, "novelty_reason": "Local windowed attention, carried over from the old 200-grid so that grid is represented under the real objective.", "prior_art": "Windowed attention", "mechanism": "", "why": "", "known": true, "seeds": 2, "ar_sd": 0.00042686328171833473, "ar": 5.1967, "sat": 5.4286, "nat": 6.0355, "ctx_gain": 0.0052, "ld_spike": 0.0, "ld_mean": 6.288595000000001, "ms_per_step": 200.74737071990967, "touched_frac": 0.9050119047619041, "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 102158627, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.06572889966182802, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.189237475395203, "sat": 6.208711922168732, "nat": 6.251869589090347, "ctx_gain": 0.00020116567611694336, "ctx_full": 6.102792263031006, "ctx_256": 6.102993428707123}, {"step": 300, "tokens": 614400, "ar": 5.9769644141197205, "sat": 6.000387281179428, "nat": 6.138799071311951, "ctx_gain": 0.004379391670227051, "ctx_full": 5.90891569852829, "ctx_256": 5.913295090198517}, {"step": 600, "tokens": 1228800, "ar": 5.6180031299591064, "sat": 5.739280134439468, "nat": 6.120278984308243, "ctx_gain": 0.004207134246826172, "ctx_full": 5.514489769935608, "ctx_256": 5.518696904182434}, {"step": 1200, "tokens": 2457600, "ar": 5.1966681480407715, "sat": 5.428640574216843, "nat": 6.035481512546539, "ctx_gain": 0.005222678184509277, "ctx_full": 5.0678364634513855, "ctx_256": 5.073059141635895}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 540, "fit": 53.22, "quality": 50.3, "scale": 50.0, "stability": 49.6, "long_context": 80.8, "serving": 50.0, "d_ar": 0.0378, "d_sat": 0.0285, "d_nat": -0.1255, "quality_delta": 0.0033, "byte_ratio": 1.009, "ms_ratio": 1.034, "step_ratio": 1.015, "tokens_equal_time": 1210267, "kv_ratio": 1.0, "storage_ratio": 1.0, "gain_gap": 0.3053, "rank": 8}, {"kind": "native", "id": "fusion879_lrk", "status": "done", "title": "FusionGatedFIRNet #879 (Fusion) hybrid, low-rank gate with per-channel kernels", "slot": "hybrid", "one_line": "Fusion layers (short+long causal FIR, delta memory with per-head retention, value; an independent sigmoid gate per path, token and head) on 3/4 of layers; attention on every 4th", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 3, "ar_sd": 0.008327708315132147, "ar": 5.0983, "sat": 5.383, "nat": 5.8, "ctx_gain": -0.0885, "ld_spike": 0.0, "ld_mean": 6.303936666666668, "ms_per_step": 393.4518098831177, "touched_frac": 0.9049682539682532, "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 294918947, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.07649181108438478, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.22428959608078, "sat": 6.240396440029144, "nat": 6.280759354432424, "ctx_gain": -0.007453799247741699, "ctx_full": 6.154914935429891, "ctx_256": 6.147461136182149}, {"step": 300, "tokens": 614400, "ar": 5.927216986815135, "sat": 5.998333712418874, "nat": 6.093812465667725, "ctx_gain": -0.03707468509674072, "ctx_full": 5.877199053764343, "ctx_256": 5.8401243686676025}, {"step": 600, "tokens": 1228800, "ar": 5.4856297969818115, "sat": 5.676308512687683, "nat": 5.953851878643036, "ctx_gain": -0.05166550477345785, "ctx_full": 5.425122618675232, "ctx_256": 5.373457113901774}, {"step": 1200, "tokens": 2457600, "ar": 5.098268588383992, "sat": 5.383000214894612, "nat": 5.799977680047353, "ctx_gain": -0.08848289648691814, "ctx_full": 5.041739463806152, "ctx_256": 4.953256567319234}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 545, "fit": 52.83, "quality": 51.4, "scale": 50.0, "stability": 49.5, "long_context": 71.2, "serving": 50.0, "d_ar": 0.0291, "d_sat": -0.0115, "d_nat": -0.0059, "quality_delta": 0.014, "byte_ratio": 1.009, "ms_ratio": 2.027, "step_ratio": 1.263, "tokens_equal_time": 972598, "kv_ratio": 1.0, "storage_ratio": 1.0, "gain_gap": 0.2116, "rank": 9}, {"kind": "native", "id": "unigram_bias", "status": "done", "title": "Token-frequency bias on the readout (control for the Markov chain candidate)", "slot": "hybrid", "one_line": "Every readout adds lam_u * log P1(v), the counted unigram; 10 fp parameters of which one is used, no extra latents.", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 3, "ar_sd": 0.02287818585151581, "ar": 5.2011, "sat": 5.3995, "nat": 5.7486, "ctx_gain": -0.2724, "ld_spike": 0.0, "ld_mean": 6.07707, "ms_per_step": 196.27122084299722, "touched_frac": 0.9176507936507932, "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 102158637, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.059562634257200466, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.050236463546753, "sat": 6.065548459688823, "nat": 6.090462406476338, "ctx_gain": -0.027752002080281574, "ctx_full": 5.996809005737305, "ctx_256": 5.969057003657023}, {"step": 300, "tokens": 614400, "ar": 5.918447872002919, "sat": 5.932642598946889, "nat": 5.96954878171285, "ctx_gain": -0.12355486551920573, "ctx_full": 5.990011056264241, "ctx_256": 5.866456190745036}, {"step": 600, "tokens": 1228800, "ar": 5.647072434425354, "sat": 5.716896692911784, "nat": 5.881177147229512, "ctx_gain": -0.2024350961049398, "ctx_full": 5.770329276720683, "ctx_256": 5.567894180615743}, {"step": 1200, "tokens": 2457600, "ar": 5.201058864593506, "sat": 5.399466713269551, "nat": 5.748633801937103, "ctx_gain": -0.27242469787597656, "ctx_full": 5.368204832077026, "ctx_256": 5.09578013420105}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 542, "fit": 52.52, "quality": 53.8, "scale": 50.0, "stability": 51.8, "long_context": 52.8, "serving": 50.0, "d_ar": 0.0088, "d_sat": 0.0504, "d_nat": 0.1114, "quality_delta": 0.0377, "byte_ratio": 1.023, "ms_ratio": 1.011, "step_ratio": 1.02, "tokens_equal_time": 1204566, "kv_ratio": 1.0, "storage_ratio": 1.0, "gain_gap": 0.0277, "rank": 10}, {"kind": "native", "id": "dlgm458", "status": "done", "title": "DualScaleMemoryRouter #458 (DLGM) hybrid", "slot": "hybrid", "one_line": "DLGM layers (local+mid causal conv, delta memory, identity, token/head softmax router) on 3/4 of layers; attention on every 4th", "novelty": 4, "novelty_reason": "The upstream ASI-Arch #458 DualScaleMemoryRouter mechanism, ported into your binary block (DLGM layers on 3 of 4 layers, attention on every 4th). Not new as an idea; new to this block.", "prior_art": "ASI-Arch #458 DLGM; Kimi-style 3:1 hybrids", "mechanism": "", "why": "", "known": true, "seeds": 3, "ar_sd": 0.012406891878008693, "ar": 5.0965, "sat": 5.3806, "nat": 5.8133, "ctx_gain": -0.0668, "ld_spike": 0.0, "ld_mean": 6.31711, "ms_per_step": 461.06282869974774, "touched_frac": 0.911746031746031, "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 39820506467, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.07665079951190139, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.196931441624959, "sat": 6.215090235074361, "nat": 6.258260528246562, "ctx_gain": -0.004359761873881022, "ctx_full": 6.126434246699016, "ctx_256": 6.122074484825134}, {"step": 300, "tokens": 614400, "ar": 5.875951965649922, "sat": 5.963218808174133, "nat": 6.086914360523224, "ctx_gain": -0.04133570194244385, "ctx_full": 5.816468715667725, "ctx_256": 5.775133013725281}, {"step": 600, "tokens": 1228800, "ar": 5.46951041618983, "sat": 5.6551589369773865, "nat": 5.938343087832133, "ctx_gain": -0.045294721921284996, "ctx_full": 5.379245599110921, "ctx_256": 5.333950877189636}, {"step": 1200, "tokens": 2457600, "ar": 5.096494356791179, "sat": 5.3806353608767195, "nat": 5.813342034816742, "ctx_gain": -0.06679916381835938, "ctx_full": 5.033629496892293, "ctx_256": 4.966830333073934}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 563, "fit": 52.5, "quality": 50.4, "scale": 50.0, "stability": 49.4, "long_context": 73.3, "serving": 50.0, "d_ar": 0.0156, "d_sat": -0.0171, "d_nat": -0.0085, "quality_delta": 0.0042, "byte_ratio": 1.017, "ms_ratio": 2.375, "step_ratio": 1.356, "tokens_equal_time": 906095, "kv_ratio": 1.0, "storage_ratio": 1.0, "gain_gap": 0.2333, "rank": 11}, {"kind": "native", "id": "ctx_local_ar", "status": "done", "title": "Local step = next-token loss at the group exit (frozen causal prefix)", "slot": "objective", "one_line": "Replaces masked-noise denoising by causal next-token prediction at the selected group exit with a no-grad prefix: deep supervision of the main task, every group becomes an early exit", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 3, "ar_sd": 0.014227411661713568, "ar": 5.1773, "sat": 5.4148, "nat": 5.833, "ctx_gain": -0.4058, "ld_spike": 0.0, "ld_mean": 5.293011666666667, "ms_per_step": 214.03908729553223, "touched_frac": 0.9033571428571424, "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 102158627, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.06930221674575478, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.212895353635152, "sat": 6.231295605500539, "nat": 6.257501125335693, "ctx_gain": -0.026198267936706543, "ctx_full": 6.153112014134725, "ctx_256": 6.126913746198018}, {"step": 300, "tokens": 614400, "ar": 5.993741154670715, "sat": 6.018625815709432, "nat": 6.065460820992787, "ctx_gain": -0.16046305497487387, "ctx_full": 6.080905556678772, "ctx_256": 5.920442501703898}, {"step": 600, "tokens": 1228800, "ar": 5.588974436124166, "sat": 5.720929304758708, "nat": 5.986992339293162, "ctx_gain": -0.33783141771952313, "ctx_full": 5.826760450998942, "ctx_256": 5.488929033279419}, {"step": 1200, "tokens": 2457600, "ar": 5.1773403485616045, "sat": 5.41479367017746, "nat": 5.83304496606191, "ctx_gain": -0.4058397610982259, "ctx_full": 5.459864536921184, "ctx_256": 5.054024775822957}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 540, "fit": 52.26, "quality": 54.3, "scale": 50.0, "stability": 59.6, "long_context": 39.4, "serving": 50.0, "d_ar": 0.0569, "d_sat": 0.0394, "d_nat": 0.0047, "quality_delta": 0.043, "byte_ratio": 1.007, "ms_ratio": 1.102, "step_ratio": 1.031, "tokens_equal_time": 1191801, "kv_ratio": 1.0, "storage_ratio": 1.0, "gain_gap": -0.1057, "rank": 12}, {"kind": "native", "id": "markov_corpus_late", "status": "done", "title": "Counted Markov chain switched on at step 600 of an ordinary run (n-gram lifts, faded in over 100 steps)", "slot": "hybrid", "one_line": "An ordinary base run for 600 steps; then lam * log(P_chain / back-off) from n-gram and skip-gram counts is faded into the logits over 100 steps. Same counts as markov_corpus_prior; no unigram term.", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 3, "ar_sd": 0.002943676456947997, "ar": 4.5366, "sat": 5.0975, "nat": 5.8768, "ctx_gain": -0.111, "ld_spike": 0.0, "ld_mean": 6.213448333333333, "ms_per_step": 199.6704339981079, "touched_frac": 0.8905714285714247, "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 102158637, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.03802398483912204, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.214567124843597, "sat": 6.237511177857717, "nat": 6.282918353875478, "ctx_gain": -0.027675747871398926, "ctx_full": 6.168927192687988, "ctx_256": 6.141251444816589}, {"step": 300, "tokens": 614400, "ar": 6.0108587344487505, "sat": 6.036364992459615, "nat": 6.0824341376622515, "ctx_gain": -0.15506084760030112, "ctx_full": 6.107449809710185, "ctx_256": 5.952388962109883}, {"step": 600, "tokens": 1228800, "ar": 5.656271815299988, "sat": 5.770442028840383, "nat": 5.9896466334660845, "ctx_gain": -0.24194606145222983, "ctx_full": 5.825280785560608, "ctx_256": 5.583334724108378}, {"step": 1200, "tokens": 2457600, "ar": 4.536560396353404, "sat": 5.097496330738068, "nat": 5.876841485500336, "ctx_gain": -0.11101154486338298, "ctx_full": 4.482407828172048, "ctx_256": 4.371396283308665}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 560, "fit": 52.23, "quality": 50.5, "scale": 50.0, "stability": 50.4, "long_context": 68.9, "serving": 50.0, "d_ar": 0.0065, "d_sat": 0.0023, "d_nat": 0.0053, "quality_delta": 0.0054, "byte_ratio": 0.993, "ms_ratio": 1.028, "step_ratio": 1.002, "tokens_equal_time": 1226530, "kv_ratio": 1.0, "storage_ratio": 1.0, "gain_gap": 0.1891, "rank": 13}, {"kind": "native", "id": "nope", "status": "done", "title": "NoPE attention", "slot": "pe", "one_line": "Remove RoPE: unit-normalised q,k with no positional rotation at all (the long-context rule)", "novelty": 1, "novelty_reason": "A known technique (no positional encoding); included as the obvious test of your long-context rule.", "prior_art": "NoPE / position-free attention", "mechanism": "", "why": "", "known": true, "seeds": 3, "ar_sd": 0.013906496659896682, "ar": 5.2896, "sat": 5.5248, "nat": 5.848, "ctx_gain": -0.1319, "ld_spike": 0.0, "ld_mean": 6.248266666666666, "ms_per_step": 182.52877394358316, "touched_frac": 0.9001349206349202, "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 102158627, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.058303712610669356, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.213645339012146, "sat": 6.227912704149882, "nat": 6.254367729028066, "ctx_gain": -0.019681493441263836, "ctx_full": 6.159592390060425, "ctx_256": 6.139910896619161}, {"step": 300, "tokens": 614400, "ar": 6.011946578820546, "sat": 6.039622128009796, "nat": 6.086946129798889, "ctx_gain": -0.11548427740732829, "ctx_full": 6.068542281786601, "ctx_256": 5.9530580043792725}, {"step": 600, "tokens": 1228800, "ar": 5.663377205530803, "sat": 5.7937555114428205, "nat": 5.97933969895045, "ctx_gain": -0.1336879332860311, "ctx_full": 5.723506490389506, "ctx_256": 5.589818557103475}, {"step": 1200, "tokens": 2457600, "ar": 5.289645651976268, "sat": 5.5247887174288435, "nat": 5.848016858100891, "ctx_gain": -0.13190690676371256, "ctx_full": 5.347608923912048, "ctx_256": 5.2157020171483355}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 542, "fit": 51.95, "quality": 50.5, "scale": 50.0, "stability": 50.0, "long_context": 66.8, "serving": 50.0, "d_ar": 0.007, "d_sat": -0.0155, "d_nat": 0.0182, "quality_delta": 0.0047, "byte_ratio": 1.004, "ms_ratio": 0.94, "step_ratio": 0.988, "tokens_equal_time": 1244008, "kv_ratio": 1.0, "storage_ratio": 1.0, "gain_gap": 0.1682, "rank": 14}, {"kind": "native", "id": "dory_loop", "status": "done", "title": "Dory-style loop: Input / Recurrent x2 / Output at the live step cost", "slot": "hybrid", "one_line": "Input 2, Recurrent 4 (shared weights, applied twice, input re-injected each loop), Output 2: the same 12 block visits per step as the live block with 2/3 of the stored latents", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 3, "ar_sd": 0.010135302668405144, "ar": 5.2201, "sat": 5.4385, "nat": 5.8227, "ctx_gain": -0.3183, "ld_spike": 0.0, "ld_mean": 6.167411666666666, "ms_per_step": 212.65808741251627, "touched_frac": 0.9007380952380943, "census": {"latent_bytes": 601295421440, "table_bytes": 1638924288, "n_matrices": 2620, "fp_params": 69324779, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.050683046439955894, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.241829971472423, "sat": 6.248870313167572, "nat": 6.27018556992213, "ctx_gain": -0.023180524508158367, "ctx_full": 6.20233416557312, "ctx_256": 6.1791536410649615}, {"step": 300, "tokens": 614400, "ar": 5.994573970635732, "sat": 6.02027012904485, "nat": 6.065988103548686, "ctx_gain": -0.15792596340179443, "ctx_full": 6.090695818265279, "ctx_256": 5.932769854863484}, {"step": 600, "tokens": 1228800, "ar": 5.614365975062053, "sat": 5.736209412415822, "nat": 5.967003722985585, "ctx_gain": -0.2611609697341919, "ctx_full": 5.791971484820048, "ctx_256": 5.530810515085856}, {"step": 1200, "tokens": 2457600, "ar": 5.2200548847516375, "sat": 5.438479741414388, "nat": 5.822696685791016, "ctx_gain": -0.31828900178273517, "ctx_full": 5.43220051129659, "ctx_256": 5.113911509513855}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 537, "fit": 51.63, "quality": 53.1, "scale": 50.0, "stability": 50.9, "long_context": 48.2, "serving": 50.0, "d_ar": 0.0351, "d_sat": 0.0266, "d_nat": 0.0244, "quality_delta": 0.0312, "byte_ratio": 1.003, "ms_ratio": 1.095, "step_ratio": 1.026, "tokens_equal_time": 1197189, "kv_ratio": 1.0, "storage_ratio": 0.999, "gain_gap": -0.0182, "rank": 15}, {"kind": "native", "id": "sk_local_snr", "status": "done", "title": "Local denoise with noise in data units (sigma_data scaling)", "slot": "objective", "one_line": "Scale the local-denoise noise by the rms of the clean embeddings so that sigma is a real signal-to-noise knob instead of swamping the +-1/sqrt(d) embeddings", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 3, "ar_sd": 0.006796498731007624, "ar": 5.2155, "sat": 5.445, "nat": 5.8269, "ctx_gain": -0.3072, "ld_spike": 0.0, "ld_mean": 5.968516666666666, "ms_per_step": 213.13110987345377, "touched_frac": 0.8957539682539682, "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 102158627, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.05998393711286872, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.238363246122996, "sat": 6.252701620260875, "nat": 6.2712002992630005, "ctx_gain": -0.026214321454366047, "ctx_full": 6.178710341453552, "ctx_256": 6.1524960199991865}, {"step": 300, "tokens": 614400, "ar": 6.012339909871419, "sat": 6.0341372688611346, "nat": 6.074204544226329, "ctx_gain": -0.15765817960103354, "ctx_full": 6.094923257827759, "ctx_256": 5.937265078226726}, {"step": 600, "tokens": 1228800, "ar": 5.642239113648732, "sat": 5.75284747282664, "nat": 5.964650472005208, "ctx_gain": -0.25353793303171795, "ctx_full": 5.818166732788086, "ctx_256": 5.564628799756368}, {"step": 1200, "tokens": 2457600, "ar": 5.2154587507247925, "sat": 5.44502317905426, "nat": 5.826881428559621, "ctx_gain": -0.3072005907694499, "ctx_full": 5.412781000137329, "ctx_256": 5.105580409367879}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 542, "fit": 50.93, "quality": 51.3, "scale": 50.0, "stability": 52.8, "long_context": 49.3, "serving": 50.0, "d_ar": 0.0091, "d_sat": 0.0112, "d_nat": 0.0268, "quality_delta": 0.013, "byte_ratio": 0.999, "ms_ratio": 1.098, "step_ratio": 1.024, "tokens_equal_time": 1200566, "kv_ratio": 1.0, "storage_ratio": 1.0, "gain_gap": -0.0071, "rank": 16}, {"kind": "control", "id": "base_r3", "status": "done", "title": "Live 600B block (control, re-run 12 Nov 2026 alongside the Fusion #879 runs)", "slot": "control", "one_line": "Identical to base: re-run on the same night and GPU as fusion879 so that ms/step and seed noise are compared like for like", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 1, "ar_sd": null, "ar": 5.2433, "sat": 5.4705, "nat": 5.8675, "ctx_gain": -0.3763, "ld_spike": 0.0, "ld_mean": 6.2622100000000005, "ms_per_step": 194.4422721862793, "touched_frac": 0.9001190476190466, "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 102158627, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.06218329862756289, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.2151278257369995, "sat": 6.220139026641846, "nat": 6.24438339471817, "ctx_gain": -0.017903923988342285, "ctx_full": 6.160545468330383, "ctx_256": 6.142641544342041}, {"step": 300, "tokens": 614400, "ar": 5.997397303581238, "sat": 6.0249884724617, "nat": 6.066418349742889, "ctx_gain": -0.15519630908966064, "ctx_full": 6.075271368026733, "ctx_256": 5.920075058937073}, {"step": 600, "tokens": 1228800, "ar": 5.635841369628906, "sat": 5.746695518493652, "nat": 5.971850216388702, "ctx_gain": -0.30987751483917236, "ctx_full": 5.839450240135193, "ctx_256": 5.5295727252960205}, {"step": 1200, "tokens": 2457600, "ar": 5.243287801742554, "sat": 5.47048407793045, "nat": 5.867462158203125, "ctx_gain": -0.3762725591659546, "ctx_full": 5.509885787963867, "ctx_256": 5.133613228797913}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 541, "fit": 50.63, "quality": 52.5, "scale": 50.0, "stability": 49.9, "long_context": 42.4, "serving": 50.0, "d_ar": 0.0263, "d_sat": 0.0255, "d_nat": 0.0229, "quality_delta": 0.0254, "byte_ratio": 1.004, "ms_ratio": 1.002, "step_ratio": 1.003, "tokens_equal_time": 1224999, "kv_ratio": 1.0, "storage_ratio": 1.0, "gain_gap": -0.0762}, {"kind": "native", "id": "ctx_local", "status": "done", "title": "Local step in real context (frozen prefix, trained group)", "slot": "objective", "one_line": "The selected group reads the residual stream of the blocks below it (no-grad prefix) instead of noised raw embeddings; masked-token target and schedule unchanged", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 3, "ar_sd": 0.032213642565704984, "ar": 5.2358, "sat": 5.4518, "nat": 5.8585, "ctx_gain": -0.3132, "ld_spike": 0.0, "ld_mean": 5.728751666666667, "ms_per_step": 212.9455804824829, "touched_frac": 0.8986111111111107, "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 102158627, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.06554453862529115, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.221489508946736, "sat": 6.247809370358785, "nat": 6.282633642355601, "ctx_gain": -0.027466694513956707, "ctx_full": 6.170877973238627, "ctx_256": 6.14341127872467}, {"step": 300, "tokens": 614400, "ar": 6.0160640478134155, "sat": 6.034179588158925, "nat": 6.077534794807434, "ctx_gain": -0.1783308188120524, "ctx_full": 6.132079641024272, "ctx_256": 5.953748822212219}, {"step": 600, "tokens": 1228800, "ar": 5.656287709871928, "sat": 5.759020904699962, "nat": 5.973817547162374, "ctx_gain": -0.2913147608439128, "ctx_full": 5.872315367062886, "ctx_256": 5.581000606218974}, {"step": 1200, "tokens": 2457600, "ar": 5.235835512479146, "sat": 5.451833764712016, "nat": 5.858470340569814, "ctx_gain": -0.31317758560180664, "ctx_full": 5.452280720074971, "ctx_256": 5.139103134473165}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 542, "fit": 50.45, "quality": 50.1, "scale": 50.0, "stability": 55.2, "long_context": 48.7, "serving": 50.0, "d_ar": -0.0057, "d_sat": 0.0044, "d_nat": 0.0176, "quality_delta": 0.001, "byte_ratio": 1.002, "ms_ratio": 1.097, "step_ratio": 1.026, "tokens_equal_time": 1198049, "kv_ratio": 1.0, "storage_ratio": 1.0, "gain_gap": -0.0131, "rank": 17}, {"kind": "native", "id": "ctx_local_noise", "status": "done", "title": "Local step in real context + data-unit input noise", "slot": "objective", "one_line": "ctx_local with the variance-preserving sigma*rms(e) embedding noise kept: does diffusion noise help once the group sees real context?", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 3, "ar_sd": 0.01942793791961509, "ar": 5.229, "sat": 5.4515, "nat": 5.8363, "ctx_gain": -0.3246, "ld_spike": 0.0, "ld_mean": 5.814791666666667, "ms_per_step": 211.85763676961264, "touched_frac": 0.8990158730158729, "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 102158627, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.06628796404612904, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.252036829789479, "sat": 6.2595014572143555, "nat": 6.285690983136495, "ctx_gain": -0.021718541781107586, "ctx_full": 6.188709735870361, "ctx_256": 6.1669911940892534}, {"step": 300, "tokens": 614400, "ar": 6.008329212665558, "sat": 6.03004664182663, "nat": 6.079040706157684, "ctx_gain": -0.1697316567103068, "ctx_full": 6.112627585728963, "ctx_256": 5.942895929018657}, {"step": 600, "tokens": 1228800, "ar": 5.651998162269592, "sat": 5.759242276350657, "nat": 5.973761816819509, "ctx_gain": -0.2923802137374878, "ctx_full": 5.863923192024231, "ctx_256": 5.571542978286743}, {"step": 1200, "tokens": 2457600, "ar": 5.229038933912913, "sat": 5.45150492588679, "nat": 5.836292147636414, "ctx_gain": -0.32455337047576904, "ctx_full": 5.446623007456462, "ctx_256": 5.122069636980693}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 542, "fit": 50.41, "quality": 50.4, "scale": 50.0, "stability": 54.4, "long_context": 47.6, "serving": 50.0, "d_ar": -0.0008, "d_sat": 0.0047, "d_nat": 0.0177, "quality_delta": 0.004, "byte_ratio": 1.002, "ms_ratio": 1.091, "step_ratio": 1.025, "tokens_equal_time": 1199291, "kv_ratio": 1.0, "storage_ratio": 1.0, "gain_gap": -0.0245, "rank": 18}, {"kind": "native", "id": "markov_posthoc", "status": "done", "title": "Plain model with the counted Markov chain mixed in afterwards (scored at the end, not trained with)", "slot": "hybrid", "one_line": "Trains exactly like base; after the last step scores (1-g) p_model + g p_chain on the evaluation windows, g chosen on the other half of the windows. No parameters, no training change.", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 3, "ar_sd": 0.010386921083830313, "ar": 5.2546, "sat": 5.4806, "nat": 5.8601, "ctx_gain": -0.3166, "ld_spike": 0.0, "ld_mean": 6.28491, "ms_per_step": 193.44464937845865, "touched_frac": 0.8960079365079358, "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 102158627, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.05988123468923353, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.24335503578186, "sat": 6.2569997906684875, "nat": 6.279259244600932, "ctx_gain": -0.0415730873743693, "ctx_full": 6.217387557029724, "ctx_256": 6.1758144696553545}, {"step": 300, "tokens": 614400, "ar": 6.015591144561768, "sat": 6.035103380680084, "nat": 6.0770682493845625, "ctx_gain": -0.15579322973887125, "ctx_full": 6.112026770909627, "ctx_256": 5.956233541170756}, {"step": 600, "tokens": 1228800, "ar": 5.6601922909418745, "sat": 5.769629736741384, "nat": 5.981516003608704, "ctx_gain": -0.24464035034179688, "ctx_full": 5.8250546058019, "ctx_256": 5.580414255460103}, {"step": 1200, "tokens": 2457600, "ar": 5.254641175270081, "sat": 5.480567057927449, "nat": 5.860143919785817, "ctx_gain": -0.3166408936182658, "ctx_full": 5.477745294570923, "ctx_256": 5.161104400952657}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 542, "fit": 50.15, "quality": 50.6, "scale": 50.0, "stability": 49.7, "long_context": 48.3, "serving": 50.0, "d_ar": 0.0045, "d_sat": 0.0045, "d_nat": 0.0139, "quality_delta": 0.0064, "byte_ratio": 0.999, "ms_ratio": 0.996, "step_ratio": 0.998, "tokens_equal_time": 1230794, "kv_ratio": 1.0, "storage_ratio": 1.0, "gain_gap": -0.0165, "rank": 19}, {"kind": "control", "id": "base", "status": "done", "title": "Live 600B block (control)", "slot": "control", "one_line": "The exact current block: RoPE cosine attention with T = sqrt(hd) in every layer, binary qkv, top-1 relu^2 MoE", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 3, "ar_sd": 0.012785029831354908, "ar": 5.2459, "sat": 5.4675, "nat": 5.8433, "ctx_gain": -0.3001, "ld_spike": 0.0, "ld_mean": 6.25246, "ms_per_step": 194.1464344660441, "touched_frac": 0.8968650793650786, "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 102158627, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.060210045958673575, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.221307754516602, "sat": 6.236490925153096, "nat": 6.262467126051585, "ctx_gain": -0.03959238529205322, "ctx_full": 6.169127106666565, "ctx_256": 6.129534721374512}, {"step": 300, "tokens": 614400, "ar": 6.005859096844991, "sat": 6.033294936021169, "nat": 6.078488290309906, "ctx_gain": -0.16538389523824057, "ctx_full": 6.109192967414856, "ctx_256": 5.943809072176616}, {"step": 600, "tokens": 1228800, "ar": 5.663718183835347, "sat": 5.773470600446065, "nat": 5.995160281658173, "ctx_gain": -0.2902825673421224, "ctx_full": 5.877588669459025, "ctx_256": 5.587306102116902}, {"step": 1200, "tokens": 2457600, "ar": 5.245898425579071, "sat": 5.46754531065623, "nat": 5.843253115812938, "ctx_gain": -0.30009488264719647, "ctx_full": 5.432854731877645, "ctx_256": 5.132759849230449}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 542, "fit": 50.0, "quality": 50.0, "scale": 50.0, "stability": 50.0, "long_context": 50.0, "serving": 50.0, "d_ar": 0.0, "d_sat": 0.0, "d_nat": 0.0, "quality_delta": 0.0, "byte_ratio": 1.0, "ms_ratio": 1.0, "step_ratio": 1.0, "tokens_equal_time": 1228800, "kv_ratio": 1.0, "storage_ratio": 1.0, "gain_gap": 0.0}, {"kind": "control", "id": "local_noop", "status": "done", "title": "Local-denoise slot kept, update skipped (control)", "slot": "control", "one_line": "Same schedule and anchor-step count as base, but the 1-in-9 local step updates nothing: isolates what the local updates themselves do", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 3, "ar_sd": 0.025008908486393096, "ar": 5.2043, "sat": 5.4328, "nat": 5.8375, "ctx_gain": -0.3387, "ld_spike": 0.0, "ld_mean": 7.16926, "ms_per_step": 213.8365109761556, "touched_frac": 0.8665793650793651, "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 102158627, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.06657061896261905, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.265810529390971, "sat": 6.268050909042358, "nat": 6.2853415211041765, "ctx_gain": -0.01912558078765869, "ctx_full": 6.22090756893158, "ctx_256": 6.201781988143921}, {"step": 300, "tokens": 614400, "ar": 6.021411697069804, "sat": 6.0442877014478045, "nat": 6.0871643622716265, "ctx_gain": -0.151975949605306, "ctx_full": 6.117503205935161, "ctx_256": 5.965527256329854}, {"step": 600, "tokens": 1228800, "ar": 5.638756692409515, "sat": 5.761382162570953, "nat": 5.971237937609355, "ctx_gain": -0.2682838837305705, "ctx_full": 5.802367250124614, "ctx_256": 5.534083366394043}, {"step": 1200, "tokens": 2457600, "ar": 5.204338987668355, "sat": 5.432750662167867, "nat": 5.837476929028829, "ctx_gain": -0.3386569023132324, "ctx_full": 5.448113044102986, "ctx_256": 5.109456141789754}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 542, "fit": 49.92, "quality": 52.2, "scale": 50.0, "stability": 40.8, "long_context": 46.1, "serving": 50.0, "d_ar": 0.0249, "d_sat": 0.0121, "d_nat": 0.0239, "quality_delta": 0.0222, "byte_ratio": 0.966, "ms_ratio": 1.101, "step_ratio": 1.0, "tokens_equal_time": 1228765, "kv_ratio": 1.0, "storage_ratio": 1.0, "gain_gap": -0.0386}, {"kind": "native", "id": "dlgm458_lr", "status": "done", "title": "DualScaleMemoryRouter #458 (DLGM) hybrid low-rank", "slot": "hybrid", "one_line": "DLGM layers (local+mid causal conv, delta memory, identity, token/head softmax router) on 3/4 of layers; attention on every 4th", "novelty": 4, "novelty_reason": "The #458 mechanism in a low-rank form sized to fit at 600B scale.", "prior_art": "ASI-Arch #458 DLGM", "mechanism": "", "why": "", "known": true, "seeds": 3, "ar_sd": 0.00481983887881823, "ar": 5.0901, "sat": 5.3797, "nat": 5.8043, "ctx_gain": -0.0886, "ld_spike": 0.0, "ld_mean": 6.320846666666667, "ms_per_step": 542.5885915756226, "touched_frac": 0.9058412698412691, "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 109016867, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.07909986104272444, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.1923433144887285, "sat": 6.2180571754773455, "nat": 6.269387265046437, "ctx_gain": -0.00769968827565511, "ctx_full": 6.12634293238322, "ctx_256": 6.118643244107564}, {"step": 300, "tokens": 614400, "ar": 5.89665162563324, "sat": 5.968656301498413, "nat": 6.0850677490234375, "ctx_gain": -0.04384382565816244, "ctx_full": 5.860845963160197, "ctx_256": 5.817002137502034}, {"step": 600, "tokens": 1228800, "ar": 5.467918356259664, "sat": 5.661232113838196, "nat": 5.95894608894984, "ctx_gain": -0.06959136327107747, "ctx_full": 5.411595543225606, "ctx_256": 5.342004179954529}, {"step": 1200, "tokens": 2457600, "ar": 5.09007853269577, "sat": 5.379696806271871, "nat": 5.804326971371968, "ctx_gain": -0.0886243184407552, "ctx_full": 5.051960150400798, "ctx_256": 4.963335831960042}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 542, "fit": 49.88, "quality": 46.1, "scale": 50.0, "stability": 49.3, "long_context": 71.1, "serving": 50.0, "d_ar": -0.0367, "d_sat": -0.0544, "d_nat": -0.0322, "quality_delta": -0.0393, "byte_ratio": 1.01, "ms_ratio": 2.795, "step_ratio": 1.456, "tokens_equal_time": 843845, "kv_ratio": 1.0, "storage_ratio": 1.0, "gain_gap": 0.2115, "rank": 20}, {"kind": "control", "id": "base_r2", "status": "done", "title": "Live 600B block (control, re-run 11 Nov 2026 alongside the Sakana candidates)", "slot": "control", "one_line": "Identical to base: re-run on the same day and GPU as the sk_* candidates so that ms/step and seed noise are compared like for like", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 3, "ar_sd": 0.008883435575581291, "ar": 5.2397, "sat": 5.4588, "nat": 5.8382, "ctx_gain": -0.3125, "ld_spike": 0.0, "ld_mean": 6.25629, "ms_per_step": 213.17382653554282, "touched_frac": 0.895619047619047, "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 102158627, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.05976304509583594, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.2215965787569685, "sat": 6.229712247848511, "nat": 6.2611478964487715, "ctx_gain": -0.026232043902079265, "ctx_full": 6.170052568117778, "ctx_256": 6.143820524215698}, {"step": 300, "tokens": 614400, "ar": 6.007381498813629, "sat": 6.030665119489034, "nat": 6.07929124434789, "ctx_gain": -0.15816330909729004, "ctx_full": 6.104330937067668, "ctx_256": 5.946167627970378}, {"step": 600, "tokens": 1228800, "ar": 5.657027463118236, "sat": 5.764114499092102, "nat": 5.979922155539195, "ctx_gain": -0.2691258986790975, "ctx_full": 5.845732569694519, "ctx_256": 5.576606671015422}, {"step": 1200, "tokens": 2457600, "ar": 5.239748895168304, "sat": 5.458752075831096, "nat": 5.838246524333954, "ctx_gain": -0.3124830722808838, "ctx_full": 5.454478184382121, "ctx_256": 5.141995112101237}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 542, "fit": 49.84, "quality": 49.9, "scale": 50.0, "stability": 50.0, "long_context": 48.8, "serving": 50.0, "d_ar": -0.005, "d_sat": 0.0004, "d_nat": 0.0119, "quality_delta": -0.0005, "byte_ratio": 0.999, "ms_ratio": 1.098, "step_ratio": 1.023, "tokens_equal_time": 1200633, "kv_ratio": 1.0, "storage_ratio": 1.0, "gain_gap": -0.0124}, {"kind": "native", "id": "dory_elastic", "status": "done", "title": "Dory-style elastic loop: 1-3 loops per training step, one checkpoint for any loop count", "slot": "hybrid", "one_line": "Same Input/Recurrent/Output split; the loop count is drawn from 1, 2, 3 on every training step (mean cost = the live block), scored at 2 loops, and also measured at 1, 2 and 4 loops", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 3, "ar_sd": 0.03517661948346403, "ar": 5.2291, "sat": 5.4493, "nat": 5.8368, "ctx_gain": -0.3197, "ld_spike": 0.0, "ld_mean": 6.143605, "ms_per_step": 208.89623959859213, "touched_frac": 0.6670153664302562, "census": {"latent_bytes": 802353577984, "table_bytes": 1638924288, "n_matrices": 3496, "fp_params": 69324779, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.04791636284914616, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.228933036327362, "sat": 6.238316396872203, "nat": 6.2647631367047625, "ctx_gain": -0.02647872765858968, "ctx_full": 6.188772360483806, "ctx_256": 6.162293632825215}, {"step": 300, "tokens": 614400, "ar": 6.0018492341041565, "sat": 6.024842182795207, "nat": 6.075345655282338, "ctx_gain": -0.14574321111043295, "ctx_full": 6.079936265945435, "ctx_256": 5.934193054835002}, {"step": 600, "tokens": 1228800, "ar": 5.638753771781921, "sat": 5.753761490186055, "nat": 5.982594569524129, "ctx_gain": -0.26268454392751056, "ctx_full": 5.811582366625468, "ctx_256": 5.548897822697957}, {"step": 1200, "tokens": 2457600, "ar": 5.2290568351745605, "sat": 5.44930084546407, "nat": 5.836841444174449, "ctx_gain": -0.3196527163187663, "ctx_full": 5.430211265881856, "ctx_256": 5.11055854956309}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 613, "fit": 49.25, "quality": 51.6, "scale": 39.6, "stability": 51.1, "long_context": 48.0, "serving": 50.0, "d_ar": 0.0184, "d_sat": 0.0148, "d_nat": 0.0109, "quality_delta": 0.0162, "byte_ratio": 0.992, "ms_ratio": 1.076, "step_ratio": 1.013, "tokens_equal_time": 1213475, "kv_ratio": 1.0, "storage_ratio": 1.333, "gain_gap": -0.0196, "rank": 21}, {"kind": "native", "id": "local_no_table", "status": "done", "title": "Local-denoise without token-table updates", "slot": "objective", "one_line": "Keeps the 1-in-9 local-denoise step but stops it changing the shared token table (bits and row scales); table learning comes only from AR, SAT and NAT", "novelty": 1, "novelty_reason": "A narrow test of one lead: local-denoise steps stop updating the shared token table.", "prior_art": "Selective freezing of shared embeddings", "mechanism": "", "why": "", "known": true, "seeds": 3, "ar_sd": 0.019754885493877165, "ar": 5.2523, "sat": 5.4695, "nat": 5.8569, "ctx_gain": -0.3636, "ld_spike": 0.0, "ld_mean": 6.295273333333333, "ms_per_step": 233.66757233937582, "touched_frac": 0.8868492063492047, "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 102158627, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.05780696737066565, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.241801679134369, "sat": 6.250885347525279, "nat": 6.276377499103546, "ctx_gain": -0.021798372268676758, "ctx_full": 6.195835471153259, "ctx_256": 6.1740370988845825}, {"step": 300, "tokens": 614400, "ar": 6.0270082751909895, "sat": 6.043166855971019, "nat": 6.0849177440007525, "ctx_gain": -0.14139469464619955, "ctx_full": 6.105787952740987, "ctx_256": 5.964393258094788}, {"step": 600, "tokens": 1228800, "ar": 5.69086879491806, "sat": 5.790413459142049, "nat": 5.977022588253021, "ctx_gain": -0.26879676183064777, "ctx_full": 5.866932153701782, "ctx_256": 5.598135391871135}, {"step": 1200, "tokens": 2457600, "ar": 5.252261618773143, "sat": 5.469525933265686, "nat": 5.85688042640686, "ctx_gain": -0.36364181836446124, "ctx_full": 5.499907334645589, "ctx_256": 5.136265516281128}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 542, "fit": 47.53, "quality": 46.8, "scale": 50.0, "stability": 49.6, "long_context": 43.6, "serving": 50.0, "d_ar": -0.0473, "d_sat": -0.0321, "d_nat": 0.0117, "quality_delta": -0.0325, "byte_ratio": 0.989, "ms_ratio": 1.204, "step_ratio": 1.043, "tokens_equal_time": 1178687, "kv_ratio": 1.0, "storage_ratio": 1.0, "gain_gap": -0.0635, "rank": 22}, {"kind": "control", "id": "sk_mode_interference", "status": "done", "title": "Diagnostic: which objective moves which held-out loss (per-step interference + KL drift)", "slot": "control", "one_line": "Live block unchanged; probes held-out AR/SAT/NAT after every step and attributes the change to the step objective", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 2, "ar_sd": 0.002634557353567475, "ar": 5.2393, "sat": 5.4575, "nat": 5.8412, "ctx_gain": -0.3397, "ld_spike": 0.0, "ld_mean": 6.2288625, "ms_per_step": 320.32066583633423, "touched_frac": 0.8991309523809515, "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 102158627, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.06046581246357317, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.210313141345978, "sat": 6.216659128665924, "nat": 6.239502489566803, "ctx_gain": -0.037171006202697754, "ctx_full": 6.15465235710144, "ctx_256": 6.117481350898743}, {"step": 300, "tokens": 614400, "ar": 6.005674719810486, "sat": 6.027657777070999, "nat": 6.076352834701538, "ctx_gain": -0.1868097186088562, "ctx_full": 6.119691371917725, "ctx_256": 5.932881653308868}, {"step": 600, "tokens": 1228800, "ar": 5.664170503616333, "sat": 5.767685204744339, "nat": 5.984972923994064, "ctx_gain": -0.24768799543380737, "ctx_full": 5.837899327278137, "ctx_256": 5.59021133184433}, {"step": 1200, "tokens": 2457600, "ar": 5.2393291890621185, "sat": 5.45749494433403, "nat": 5.841156929731369, "ctx_gain": -0.33971041440963745, "ctx_full": 5.476968228816986, "ctx_256": 5.137257814407349}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 573, "fit": 46.47, "quality": 44.3, "scale": 50.0, "stability": 50.2, "long_context": 46.0, "serving": 50.0, "d_ar": -0.0754, "d_sat": -0.0513, "d_nat": -0.0099, "quality_delta": -0.0575, "byte_ratio": 1.003, "ms_ratio": 1.65, "step_ratio": 1.164, "tokens_equal_time": 1055336, "kv_ratio": 1.0, "storage_ratio": 1.0, "gain_gap": -0.0396}, {"kind": "native", "id": "loop_small_span", "status": "done", "title": "Loop a short middle span twice (all weights kept)", "slot": "hybrid", "one_line": "All 12 blocks kept; the 2 middle blocks are applied twice with input re-injection: 14 visits per step (+17% step cost), no extra stored latents", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 3, "ar_sd": 0.01626237274316471, "ar": 5.2417, "sat": 5.4685, "nat": 5.8447, "ctx_gain": -0.3192, "ld_spike": 0.0, "ld_mean": 6.347090000000001, "ms_per_step": 267.16967423756915, "touched_frac": 0.8903116531165315, "census": {"latent_bytes": 701421846528, "table_bytes": 1638924288, "n_matrices": 3056, "fp_params": 102822179, "kv_bytes_per_token": 29032448}, "flip_pct_end": 0.051623519448462774, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.242535432179769, "sat": 6.245243847370148, "nat": 6.276566882928212, "ctx_gain": -0.02802598476409912, "ctx_full": 6.202270229657491, "ctx_256": 6.174244244893392}, {"step": 300, "tokens": 614400, "ar": 5.995561261971791, "sat": 6.019871095816295, "nat": 6.065175493558248, "ctx_gain": -0.16615025202433267, "ctx_full": 6.10333792368571, "ctx_256": 5.937187671661377}, {"step": 600, "tokens": 1228800, "ar": 5.632730960845947, "sat": 5.749077161153157, "nat": 5.969409108161926, "ctx_gain": -0.2692604462305705, "ctx_full": 5.804511348406474, "ctx_256": 5.535250902175903}, {"step": 1200, "tokens": 2457600, "ar": 5.241664528846741, "sat": 5.468501567840576, "nat": 5.844656626383464, "ctx_gain": -0.31918970743815106, "ctx_full": 5.447556972503662, "ctx_256": 5.128367265065511}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 584, "fit": 45.47, "quality": 44.8, "scale": 44.5, "stability": 49.1, "long_context": 48.1, "serving": 44.5, "d_ar": -0.0696, "d_sat": -0.0506, "d_nat": -0.0008, "quality_delta": -0.052, "byte_ratio": 1.157, "ms_ratio": 1.376, "step_ratio": 1.212, "tokens_equal_time": 1014068, "kv_ratio": 1.166, "storage_ratio": 1.165, "gain_gap": -0.0191, "rank": 23}, {"kind": "control", "id": "t1", "status": "done", "title": "Attention temperature T=1", "slot": "control", "one_line": "Ablation: logit = cos(q,k), the setting before step 1182 of the live run", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 2, "ar_sd": 0.04189915148219551, "ar": 5.3968, "sat": 5.5912, "nat": 5.8826, "ctx_gain": -0.2672, "ld_spike": 0.0, "ld_mean": 6.2902475, "ms_per_step": 193.0609941482544, "touched_frac": 0.8785833333333308, "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 102158627, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.045661904155056485, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.285264551639557, "sat": 6.271162807941437, "nat": 6.294732213020325, "ctx_gain": -0.023960411548614502, "ctx_full": 6.214517533779144, "ctx_256": 6.19055712223053}, {"step": 300, "tokens": 614400, "ar": 6.032492250204086, "sat": 6.057846933603287, "nat": 6.094016909599304, "ctx_gain": -0.13001638650894165, "ctx_full": 6.08850634098053, "ctx_256": 5.958489954471588}, {"step": 600, "tokens": 1228800, "ar": 5.81103190779686, "sat": 5.8661258816719055, "nat": 5.997697591781616, "ctx_gain": -0.2274838089942932, "ctx_full": 5.9860236048698425, "ctx_256": 5.758539795875549}, {"step": 1200, "tokens": 2457600, "ar": 5.396803230047226, "sat": 5.591157466173172, "nat": 5.882622539997101, "ctx_gain": -0.26721280813217163, "ctx_full": 5.597878038883209, "ctx_256": 5.330665230751038}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 540, "fit": 44.82, "quality": 40.1, "scale": 50.0, "stability": 49.6, "long_context": 53.3, "serving": 50.0, "d_ar": -0.1373, "d_sat": -0.086, "d_nat": 0.0003, "quality_delta": -0.0995, "byte_ratio": 0.98, "ms_ratio": 0.994, "step_ratio": 0.983, "tokens_equal_time": 1249651, "kv_ratio": 1.0, "storage_ratio": 1.0, "gain_gap": 0.0329}, {"kind": "native", "id": "fusion879_lr", "status": "done", "title": "FusionGatedFIRNet #879 (Fusion) hybrid low-rank", "slot": "hybrid", "one_line": "Fusion layers (short+long causal FIR, delta memory with per-head retention, value; an independent sigmoid gate per path, token and head) on 3/4 of layers; attention on every 4th", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 3, "ar_sd": 0.028593949657937816, "ar": 5.1074, "sat": 5.3908, "nat": 5.8154, "ctx_gain": -0.118, "ld_spike": 0.0, "ld_mean": 6.294751666666667, "ms_per_step": 743.5529629389445, "touched_frac": 0.9082698412698406, "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 114233507, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.07873930982386622, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.246403952439626, "sat": 6.264572163422902, "nat": 6.304192006587982, "ctx_gain": -0.002152085304260254, "ctx_full": 6.1589955886205034, "ctx_256": 6.156843503316243}, {"step": 300, "tokens": 614400, "ar": 5.922684550285339, "sat": 5.99828455845515, "nat": 6.10121105114619, "ctx_gain": -0.046471238136291504, "ctx_full": 5.861321012179057, "ctx_256": 5.814849774042766}, {"step": 600, "tokens": 1228800, "ar": 5.490259607632955, "sat": 5.677478174368541, "nat": 5.959240575631459, "ctx_gain": -0.08859082063039143, "ctx_full": 5.444842576980591, "ctx_256": 5.3562517563502}, {"step": 1200, "tokens": 2457600, "ar": 5.107352991898854, "sat": 5.390841384728749, "nat": 5.815444668134053, "ctx_gain": -0.117959459622701, "ctx_full": 5.089093844095866, "ctx_256": 4.971134384473165}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 543, "fit": 43.86, "quality": 35.6, "scale": 50.0, "stability": 49.6, "long_context": 68.2, "serving": 50.0, "d_ar": -0.1638, "d_sat": -0.1542, "d_nat": -0.0748, "quality_delta": -0.1441, "byte_ratio": 1.013, "ms_ratio": 3.83, "step_ratio": 1.717, "tokens_equal_time": 715666, "kv_ratio": 1.0, "storage_ratio": 1.0, "gain_gap": 0.1821, "rank": 24}, {"kind": "native", "id": "tied_copies", "status": "done", "title": "Copy depth tied to its parent latents", "slot": "storage", "one_line": "The 600B lineage with every copy block reading the latents of its root parent (block i -> (i % 320) % 80): same visits and step cost, 80 of 380 blocks stored; each visit keeps its own row scales, norms and router", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 3, "ar_sd": 0.014562463033271193, "ar": 5.3715, "sat": 5.5496, "nat": 5.8827, "ctx_gain": -0.5391, "ld_spike": 0.0, "ld_mean": 6.189998333333334, "ms_per_step": 212.0464245478312, "touched_frac": 0.8913807189542514, "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 102158627, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.03137486583703789, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.230621079603831, "sat": 6.240866025288899, "nat": 6.259874939918518, "ctx_gain": -0.03016022841135661, "ctx_full": 6.184254924456279, "ctx_256": 6.154094696044922}, {"step": 300, "tokens": 614400, "ar": 6.03654404481252, "sat": 6.048552254835765, "nat": 6.081683814525604, "ctx_gain": -0.12481339772542317, "ctx_full": 6.097284913063049, "ctx_256": 5.972471515337626}, {"step": 600, "tokens": 1228800, "ar": 5.776838302612305, "sat": 5.824767529964447, "nat": 5.978163739045461, "ctx_gain": -0.308216134707133, "ctx_full": 6.010234157244365, "ctx_256": 5.7020180225372314}, {"step": 1200, "tokens": 2457600, "ar": 5.371455550193787, "sat": 5.549602667490642, "nat": 5.8826602300008135, "ctx_gain": -0.5390644073486328, "ctx_full": 5.812458912531535, "ctx_256": 5.273394505182902}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 522, "fit": 43.24, "quality": 41.9, "scale": 50.0, "stability": 50.6, "long_context": 26.1, "serving": 50.0, "d_ar": -0.12, "d_sat": -0.0572, "d_nat": 0.0143, "quality_delta": -0.0806, "byte_ratio": 0.994, "ms_ratio": 1.092, "step_ratio": 1.018, "tokens_equal_time": 1206523, "kv_ratio": 1.0, "storage_ratio": 1.0, "gain_gap": -0.239, "rank": 25}, {"kind": "native", "id": "grow_experts", "status": "done", "title": "Growth by width: one more expert in every block (+204B)", "slot": "expert", "one_line": "Every block gets one more expert (at 600B: 1311 experts instead of 931, +204B latents): same compute per token, more bytes streamed per step", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 3, "ar_sd": 0.06899936522990589, "ar": 5.3182, "sat": 5.529, "nat": 5.9175, "ctx_gain": -0.2915, "ld_spike": 0.0, "ld_mean": 6.257000000000001, "ms_per_step": 247.97375996907553, "touched_frac": 0.8820271867612308, "census": {"latent_bytes": 805843238912, "table_bytes": 1638924288, "n_matrices": 3382, "fp_params": 133288607, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.048238097977585964, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.256974756717682, "sat": 6.252871851126353, "nat": 6.280542592207591, "ctx_gain": -0.025095780690511067, "ctx_full": 6.1899155378341675, "ctx_256": 6.164819757143657}, {"step": 300, "tokens": 614400, "ar": 6.043026328086853, "sat": 6.058166424433391, "nat": 6.090719064076741, "ctx_gain": -0.12382773558298747, "ctx_full": 6.108907659848531, "ctx_256": 5.985079924265544}, {"step": 600, "tokens": 1228800, "ar": 5.69802447160085, "sat": 5.783527950445811, "nat": 5.981428921222687, "ctx_gain": -0.224247415860494, "ctx_full": 5.855807860692342, "ctx_256": 5.631560444831848}, {"step": 1200, "tokens": 2457600, "ar": 5.3181896805763245, "sat": 5.528996606667836, "nat": 5.917510608832042, "ctx_gain": -0.291500449180603, "ctx_full": 5.503664612770081, "ctx_256": 5.2121641635894775}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 620, "fit": 41.38, "quality": 37.1, "scale": 39.5, "stability": 50.0, "long_context": 50.9, "serving": 50.0, "d_ar": -0.1675, "d_sat": -0.1161, "d_nat": -0.0285, "quality_delta": -0.1294, "byte_ratio": 1.317, "ms_ratio": 1.277, "step_ratio": 1.307, "tokens_equal_time": 940214, "kv_ratio": 1.0, "storage_ratio": 1.339, "gain_gap": 0.0086, "rank": 26}, {"kind": "native", "id": "grow_depth", "status": "done", "title": "Growth by depth: +200B of new blocks", "slot": "storage", "one_line": "16 independent blocks instead of 12 (at 600B: 507 blocks, +200B latents): what one more depth expansion costs per step and returns at equal box time", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 3, "ar_sd": 0.03369978734220644, "ar": 5.2775, "sat": 5.4868, "nat": 5.8641, "ctx_gain": -0.2977, "ld_spike": 0.0, "ld_mean": 6.312721666666666, "ms_per_step": 306.34427070617676, "touched_frac": 0.8865839243498806, "census": {"latent_bytes": 802890448896, "table_bytes": 1638924288, "n_matrices": 3498, "fp_params": 136229466, "kv_bytes_per_token": 33226752}, "flip_pct_end": 0.05171597760004424, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.251921832561493, "sat": 6.237688442071279, "nat": 6.261025269826253, "ctx_gain": -0.03357887268066406, "ctx_full": 6.225348194440206, "ctx_256": 6.1917693217595415}, {"step": 300, "tokens": 614400, "ar": 6.012177387873332, "sat": 6.032998720804851, "nat": 6.0791661739349365, "ctx_gain": -0.1431272824605306, "ctx_full": 6.093534509340922, "ctx_256": 5.950407226880391}, {"step": 600, "tokens": 1228800, "ar": 5.681624352931976, "sat": 5.785883108774821, "nat": 5.987574537595113, "ctx_gain": -0.22938485940297446, "ctx_full": 5.831207950909932, "ctx_256": 5.601823091506958}, {"step": 1200, "tokens": 2457600, "ar": 5.277501503626506, "sat": 5.486751159032186, "nat": 5.864108145236969, "ctx_gain": -0.2976565758387248, "ctx_full": 5.465406735738118, "ctx_256": 5.167750159899394}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 628, "fit": 39.87, "quality": 36.4, "scale": 39.6, "stability": 49.4, "long_context": 50.2, "serving": 39.6, "d_ar": -0.1727, "d_sat": -0.1282, "d_nat": -0.0353, "quality_delta": -0.1363, "byte_ratio": 1.319, "ms_ratio": 1.578, "step_ratio": 1.384, "tokens_equal_time": 888141, "kv_ratio": 1.334, "storage_ratio": 1.334, "gain_gap": 0.0024, "rank": 27}, {"kind": "native", "id": "rot12_trunc", "status": "done", "title": "Anchor rotation with the backward truncated below the updating shard", "slot": "objective", "one_line": "Same rotation as rot12_full, but blocks below the updating shard run forward-only: tests whether the dropped gain/router/embedding-side gradients matter per step", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 3, "ar_sd": 0.019468931815250137, "ar": 5.9414, "sat": 6.0227, "nat": 6.0617, "ctx_gain": -0.1933, "ld_spike": 0.0, "ld_mean": 6.276924999999999, "ms_per_step": 117.59666601816814, "touched_frac": 0.49620238095238145, "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 102158627, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.07988739466799698, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.362269838651021, "sat": 6.332282344500224, "nat": 6.34775843222936, "ctx_gain": -0.015369256337483725, "ctx_full": 6.307417710622151, "ctx_256": 6.292048454284668}, {"step": 300, "tokens": 614400, "ar": 6.117544968922933, "sat": 6.131814281145732, "nat": 6.152203778425853, "ctx_gain": -0.025133371353149414, "ctx_full": 6.0636740525563555, "ctx_256": 6.038540681203206}, {"step": 600, "tokens": 1228800, "ar": 6.062958478927612, "sat": 6.099640687306722, "nat": 6.135173141956329, "ctx_gain": -0.04141612847646078, "ctx_full": 6.042083263397217, "ctx_256": 6.000667134920756}, {"step": 1200, "tokens": 2457600, "ar": 5.941427131493886, "sat": 6.022731999556224, "nat": 6.061684171358745, "ctx_gain": -0.19330827395121256, "ctx_full": 6.091645956039429, "ctx_256": 5.898337682088216}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 493, "fit": 37.39, "quality": 25.2, "scale": 50.0, "stability": 49.8, "long_context": 60.7, "serving": 50.0, "d_ar": -0.2996, "d_sat": -0.2631, "d_nat": -0.0797, "quality_delta": -0.2483, "byte_ratio": 0.553, "ms_ratio": 0.606, "step_ratio": 0.566, "tokens_equal_time": 2169586, "kv_ratio": 1.0, "storage_ratio": 1.0, "gain_gap": 0.1068, "rank": 28}, {"kind": "control", "id": "rot12_full", "status": "done", "title": "Anchor rotation mirror: one block of 12 updated per anchor step, full backward (control)", "slot": "control", "one_line": "Mirrors the live contiguous-shard rotation (1/12 of the body updated per anchor step) with the full-depth backward; reference for rot12_trunc", "novelty": null, "novelty_reason": "", "prior_art": "", "mechanism": "", "why": "", "known": false, "seeds": 3, "ar_sd": 0.017105053258236094, "ar": 5.7857, "sat": 5.8444, "nat": 5.9835, "ctx_gain": -0.2907, "ld_spike": 0.0, "ld_mean": 6.302006666666666, "ms_per_step": 172.34142621358237, "touched_frac": 0.8870873015872971, "census": {"latent_bytes": 601832292352, "table_bytes": 1638924288, "n_matrices": 2622, "fp_params": 102158627, "kv_bytes_per_token": 24903680}, "flip_pct_end": 0.07807444109281804, "causal_ok": true, "curve": [{"step": 150, "tokens": 307200, "ar": 6.3761366208394366, "sat": 6.399401048819224, "nat": 6.42850391070048, "ctx_gain": -0.0011096000671386719, "ctx_full": 6.3004023631413775, "ctx_256": 6.299292763074239}, {"step": 300, "tokens": 614400, "ar": 6.11952942609787, "sat": 6.12773980696996, "nat": 6.156139115492503, "ctx_gain": -0.03404430548350016, "ctx_full": 6.079339504241943, "ctx_256": 6.045295198758443}, {"step": 600, "tokens": 1228800, "ar": 6.027705272038777, "sat": 6.055102149645488, "nat": 6.088212947050731, "ctx_gain": -0.10487711429595947, "ctx_full": 6.057559529940288, "ctx_256": 5.952682415644328}, {"step": 1200, "tokens": 2457600, "ar": 5.7857009172439575, "sat": 5.844401220480601, "nat": 5.983540991942088, "ctx_gain": -0.29073552290598553, "ctx_full": 6.04784909884135, "ctx_256": 5.757113575935364}], "sim_sha": ["dc5c3c28ddbf1853"], "skipped": 0, "max_gpu_mib": 539, "fit": 34.52, "quality": 21.8, "scale": 50.0, "stability": 49.5, "long_context": 50.9, "serving": 50.0, "d_ar": -0.3511, "d_sat": -0.2704, "d_nat": -0.0875, "quality_delta": -0.2822, "byte_ratio": 0.989, "ms_ratio": 0.888, "step_ratio": 0.964, "tokens_equal_time": 1275025, "kv_ratio": 1.0, "storage_ratio": 1.0, "gain_gap": 0.0094}, {"kind": "asi105", "id": "ASI#1735", "index": 1735, "name": "HybridGateFlow", "internal": "delta_net_hhmr", "parent": 1544, "summary": "Introduce hierarchical hybrid gating with rich statistics for adaptive routing, enhancing extraction, reasoning, and specialisation efficiency.", "params_upstream": "495.73M", "upstream_fitness": 2.2847, "upstream_acc_8task": 39.79, "upstream_train_loss": 3.3998, "code_dist_deltanet": 0.1155, "code_dist_knn5": 0.0085, "code_same_as_parent": null, "code_novelty": 5.4, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 38276, "forward_ms": 5673.8, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#1475", "index": 1475, "name": "TempMixAnnealRouter", "internal": "delta_net_len_hgate_mixanneal", "parent": 1385, "summary": "Introduce dynamic temperature annealing and persistent cross-head mixing floor for sharper, diversified routing and improved cooperation.", "params_upstream": "817.01M", "upstream_fitness": 2.6654, "upstream_acc_8task": 39.94, "upstream_train_loss": 3.4014, "code_dist_deltanet": 0.122, "code_dist_knn5": 0.0149, "code_same_as_parent": null, "code_novelty": 9.2, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 86590, "forward_ms": 7575.7, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#518", "index": 518, "name": "TriScale-GatedFusion", "internal": "delta_net_rmsgm", "parent": 364, "summary": "Introduce feedback-aware gating and triple-scale FIR to balance local-global paths, prevent collapse, and enhance reasoning.", "params_upstream": "466.61M", "upstream_fitness": 2.2933, "upstream_acc_8task": 39.58, "upstream_train_loss": 3.4027, "code_dist_deltanet": 0.1068, "code_dist_knn5": 0.0104, "code_same_as_parent": null, "code_novelty": 4.2, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 33475, "forward_ms": 36.2, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#665", "index": 665, "name": "HyperRouteFusion", "internal": "delta_net_hrem", "parent": 556, "summary": "Introduce hierarchically-routed entropic multi-scale memory fusion with convex token gating and adaptive per-head entropy control.", "params_upstream": "474.88M", "upstream_fitness": 2.3414, "upstream_acc_8task": 39.13, "upstream_train_loss": 3.404, "code_dist_deltanet": 0.1245, "code_dist_knn5": 0.0149, "code_same_as_parent": 0.733, "code_novelty": 9.2, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 36608, "forward_ms": 4.7, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#902", "index": 902, "name": "EntropyEnhancedMultiScaleGateNet", "internal": "delta_net_entropy_floor", "parent": 730, "summary": "Introduce persistent diversity enforcement, mid-range FIR kernel, and content-aware gating for stable multi-path reasoning.", "params_upstream": "464.54M", "upstream_fitness": 2.4207, "upstream_acc_8task": 39.35, "upstream_train_loss": 3.4044, "code_dist_deltanet": 0.1042, "code_dist_knn5": 0.0125, "code_same_as_parent": null, "code_novelty": 4.9, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 196022, "forward_ms": 35.6, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#559", "index": 559, "name": "OutputAwareMultiScaleRouter", "internal": "delta_net_mor", "parent": 364, "summary": "Introduce output-aware routing using modulated logits and path statistics to balance multi-scale pathways for improved reasoning.", "params_upstream": "464.64M", "upstream_fitness": 2.3262, "upstream_acc_8task": 40.3, "upstream_train_loss": 3.4046, "code_dist_deltanet": 0.1117, "code_dist_knn5": 0.0078, "code_same_as_parent": null, "code_novelty": 4.6, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 29104, "forward_ms": 9.3, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#458", "index": 458, "name": "DualScaleMemoryRouter", "internal": "delta_net_dlgm", "parent": 364, "summary": "Introduce dual-scale gated memory with causal conv branches and dynamic routing for adaptive local-global token processing.", "params_upstream": "464.64M", "upstream_fitness": 2.4564, "upstream_acc_8task": 39.95, "upstream_train_loss": 3.4047, "code_dist_deltanet": 0.0999, "code_dist_knn5": 0.0071, "code_same_as_parent": null, "code_novelty": 1.4, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 29096, "forward_ms": 4.0, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#1354", "index": 1354, "name": "FusionGate-MS", "internal": "delta_net_omsgf", "parent": 908, "summary": "Introduce output-aware gated fusion combining multi-scale FIR paths, per-head softplus temperature, and entropy regularization.", "params_upstream": "443.87M", "upstream_fitness": 2.362, "upstream_acc_8task": 40.04, "upstream_train_loss": 3.4058, "code_dist_deltanet": 0.1017, "code_dist_knn5": 0.0088, "code_same_as_parent": 0.606, "code_novelty": 2.5, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 28014, "forward_ms": 4.8, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#1763", "index": 1763, "name": "AdaptiveGatedRouter-Hybrid", "internal": "delta_net_cpaghr", "parent": 1544, "summary": "Introduce nonlinear content-positional gating, progressive per-head untying, entropy annealing, and always-on local residual path.", "params_upstream": "446.13M", "upstream_fitness": 2.5671, "upstream_acc_8task": 39.5, "upstream_train_loss": 3.406, "code_dist_deltanet": 0.1076, "code_dist_knn5": 0.0088, "code_same_as_parent": null, "code_novelty": 3.6, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 31226, "forward_ms": 6.2, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#1396", "index": 1396, "name": "FusionGate-CAGT", "internal": "delta_net_cagf_dpaf_eash", "parent": 671, "summary": "Introduce context-aware gated fusion with dual-phase path floor and entropy-annealed sharpening for adaptive specialization.", "params_upstream": "444.54M", "upstream_fitness": 2.353, "upstream_acc_8task": 39.14, "upstream_train_loss": 3.406, "code_dist_deltanet": 0.1244, "code_dist_knn5": 0.0155, "code_same_as_parent": null, "code_novelty": 9.4, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 29738, "forward_ms": 21.2, "error": null, "max_future_leak": 5.781650543212891e-06, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#908", "index": 908, "name": "AdaptiveFusionNet", "internal": "delta_net_aft", "parent": 580, "summary": "Introduce token-adaptive context floor and per-head learnable temperature for dynamic probability allocation in DeltaNet-AFT.", "params_upstream": "615.54M", "upstream_fitness": 2.6619, "upstream_acc_8task": 40.01, "upstream_train_loss": 3.4061, "code_dist_deltanet": 0.1039, "code_dist_knn5": 0.009, "code_same_as_parent": null, "code_novelty": 3.0, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 53418, "forward_ms": 4.7, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#537", "index": 537, "name": "FusionGateNet_v3", "internal": "delta_net_mshmfv2", "parent": 433, "summary": "Introduce short FIR, richer fusion stats, and per-head temperature for improved local precision and dynamic sparsity.", "params_upstream": "490.52M", "upstream_fitness": 2.3945, "upstream_acc_8task": 39.82, "upstream_train_loss": 3.4064, "code_dist_deltanet": 0.0905, "code_dist_knn5": 0.0129, "code_same_as_parent": null, "code_novelty": 3.7, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 47338, "forward_ms": 4.6, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#1730", "index": 1730, "name": "AdaptiveEntropyRouter", "internal": "delta_net_afef", "parent": 1329, "summary": "Introduce adaptive per-head, per-path gating with learnable floors and entropy regularisation to balance routing diversity.", "params_upstream": "468.48M", "upstream_fitness": 2.6612, "upstream_acc_8task": 39.98, "upstream_train_loss": 3.4066, "code_dist_deltanet": 0.1028, "code_dist_knn5": 0.0108, "code_same_as_parent": null, "code_novelty": 3.9, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 32946, "forward_ms": 4.9, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#556", "index": 556, "name": "PathAwareMemoryRouter", "internal": "delta_net_dfpcr", "parent": 370, "summary": "Introduce path-conditioned memory routing with dual-scale convolutions for adaptive, feedback-driven global and local attention allocation.", "params_upstream": "471.08M", "upstream_fitness": 2.6057, "upstream_acc_8task": 39.63, "upstream_train_loss": 3.4085, "code_dist_deltanet": 0.1066, "code_dist_knn5": 0.0076, "code_same_as_parent": null, "code_novelty": 2.9, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 33608, "forward_ms": 4.1, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#927", "index": 927, "name": "AdaptiveFusionRNet", "internal": "delta_net_afrc", "parent": 730, "summary": "Introduce adaptive context floors, richer gate statistics, very-long FIR, hierarchical routing, and efficient causal implementation.", "params_upstream": "472.41M", "upstream_fitness": 2.4534, "upstream_acc_8task": 39.68, "upstream_train_loss": 3.4086, "code_dist_deltanet": 0.1173, "code_dist_knn5": 0.0117, "code_same_as_parent": null, "code_novelty": 7.6, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 37040, "forward_ms": 4.7, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#1434", "index": 1434, "name": "HierarchiMix-Gate", "internal": "delta_net_hafmg", "parent": 1367, "summary": "Introduce hierarchical adaptive-floor mixture gating for robust token-wise routing, preventing path collapse and optimizing context-copy balance.", "params_upstream": "469.26M", "upstream_fitness": 2.2285, "upstream_acc_8task": 39.92, "upstream_train_loss": 3.4096, "code_dist_deltanet": 0.1112, "code_dist_knn5": 0.0084, "code_same_as_parent": null, "code_novelty": 4.3, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 31914, "forward_ms": 4.7, "error": null, "max_future_leak": 2.218782901763916e-05, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#1195", "index": 1195, "name": "BoundedTempAnnealNet", "internal": "delta_net_afbt", "parent": 908, "summary": "Introduce bounded-temperature gating and annealed adaptive floors to stabilize contextual softmax and improve task-specific routing.", "params_upstream": "615.54M", "upstream_fitness": 2.527, "upstream_acc_8task": 39.26, "upstream_train_loss": 3.4098, "code_dist_deltanet": 0.1036, "code_dist_knn5": 0.0088, "code_same_as_parent": 0.82, "code_novelty": 2.7, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 53418, "forward_ms": 4.4, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#814", "index": 814, "name": "GateDivergeTransformer", "internal": "delta_net_htfr", "parent": 471, "summary": "Introduce hierarchical three-way gating with \u03b5-floor, entropy regularisation, and per-head temperature for balanced memory path utilisation.", "params_upstream": "817.77M", "upstream_fitness": 2.6162, "upstream_acc_8task": 40.06, "upstream_train_loss": 3.4103, "code_dist_deltanet": 0.1223, "code_dist_knn5": 0.0149, "code_same_as_parent": null, "code_novelty": 9.2, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 88562, "forward_ms": 5.0, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#1771", "index": 1771, "name": "AdaptiveGateRouter_X", "internal": "delta_net_taigr_xs", "parent": 1598, "summary": "Introduce token-adaptive gated identity paths and hierarchical context routing with learnable annealing for task-flexible mass allocation.", "params_upstream": "477.57M", "upstream_fitness": 2.4496, "upstream_acc_8task": 38.82, "upstream_train_loss": 3.4114, "code_dist_deltanet": 0.1159, "code_dist_knn5": 0.0106, "code_same_as_parent": null, "code_novelty": 6.8, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 32849, "forward_ms": 5.0, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#1167", "index": 1167, "name": "GatedFusionTransformer", "internal": "delta_net_headgated", "parent": 1000, "summary": "Introduce head-specific output gating to enhance token-level control post-fusion, improving local reasoning without global trade-offs.", "params_upstream": "439.52M", "upstream_fitness": 2.6026, "upstream_acc_8task": 39.12, "upstream_train_loss": 3.4126, "code_dist_deltanet": 0.1074, "code_dist_knn5": 0.0049, "code_same_as_parent": null, "code_novelty": 2.3, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 27435, "forward_ms": 4.6, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#1017", "index": 1017, "name": "FusionLogicNet", "internal": "delta_net_hybrid_floor_gt", "parent": 682, "summary": "Introduce group-wise temperature sharing, hybrid static-dynamic residuals, and auto-annealed entropy-KL for adaptive multi-path routing.", "params_upstream": "442.57M", "upstream_fitness": 2.3732, "upstream_acc_8task": 39.6, "upstream_train_loss": 3.4137, "code_dist_deltanet": 0.1054, "code_dist_knn5": 0.0107, "code_same_as_parent": null, "code_novelty": 4.2, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 28467, "forward_ms": 17.6, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#922", "index": 922, "name": "ParallelFusionTransformer", "internal": "delta_net_phfg", "parent": 497, "summary": "Introduce Parallel\u2013Hierarchical Fusion Gate enabling independent local-global routing without zero-sum competition, ensuring stable gradient flow.", "params_upstream": "466.51M", "upstream_fitness": 2.3161, "upstream_acc_8task": 39.11, "upstream_train_loss": 3.4141, "code_dist_deltanet": 0.1174, "code_dist_knn5": 0.0128, "code_same_as_parent": null, "code_novelty": 8.0, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 30124, "forward_ms": 4.9, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#486", "index": 486, "name": "DepthwiseConvPointMixer", "internal": "delta_net_mscmix_pointwise", "parent": 417, "summary": "Introduce point-wise convolution after depth-wise kernels for efficient global channel mixing and improved narrative reasoning tasks.", "params_upstream": "492.04M", "upstream_fitness": 2.2916, "upstream_acc_8task": 39.53, "upstream_train_loss": 3.4143, "code_dist_deltanet": 0.1061, "code_dist_knn5": 0.0122, "code_same_as_parent": null, "code_novelty": 5.1, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 37478, "forward_ms": 4.6, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#657", "index": 657, "name": "QuotaGatedStatNet", "internal": "delta_net_ms_gstat3_quota", "parent": 556, "summary": "Introduce Delta-quota gating to enforce minimal long-memory flow, preserving statistic-aware mechanisms and sub-quadratic complexity.", "params_upstream": "464.64M", "upstream_fitness": 2.3365, "upstream_acc_8task": 38.72, "upstream_train_loss": 3.4163, "code_dist_deltanet": 0.1058, "code_dist_knn5": 0.0097, "code_same_as_parent": 0.513, "code_novelty": 3.6, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 29098, "forward_ms": 4.5, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#1567", "index": 1567, "name": "StreamAwareRouter", "internal": "delta_net_qsr", "parent": 1367, "summary": "Introduce Query-and-Summary Router for parameter-efficient gating, reducing compute while preserving content-aware stream fusion.", "params_upstream": "415.99M", "upstream_fitness": 2.627, "upstream_acc_8task": 39.08, "upstream_train_loss": 3.4179, "code_dist_deltanet": 0.1026, "code_dist_knn5": 0.0085, "code_same_as_parent": null, "code_novelty": 2.4, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 46198, "forward_ms": 4.6, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#712", "index": 712, "name": "AdaptiveGateNet-AFP", "internal": "delta_net_afp", "parent": 471, "summary": "Introduce adaptive per-head gating with learnable floors and temperatures for efficient, specialised routing and stability improvements.", "params_upstream": "415.42M", "upstream_fitness": 2.5283, "upstream_acc_8task": 39.81, "upstream_train_loss": 3.4193, "code_dist_deltanet": 0.1012, "code_dist_knn5": 0.0138, "code_same_as_parent": null, "code_novelty": 4.7, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 25332, "forward_ms": 5.9, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#1543", "index": 1543, "name": "FusionGateBR", "internal": "delta_net_aegf_br", "parent": 965, "summary": "Introduce annealed fusion gate combining entropy regularisation and balanced residuals for variance control and global-task accuracy.", "params_upstream": "439.13M", "upstream_fitness": 2.6608, "upstream_acc_8task": 39.47, "upstream_train_loss": 3.4229, "code_dist_deltanet": 0.0999, "code_dist_knn5": 0.0095, "code_same_as_parent": 0.725, "code_novelty": 2.6, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 27184, "forward_ms": 4.7, "error": null, "max_future_leak": 9.5367431640625e-07, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#703", "index": 703, "name": "ContentSharpRouter", "internal": "delta_net_csm", "parent": 649, "summary": "Introduce ContentSharpGate combining content-aware gating with per-head temperature for adaptive routing, preserving O(N) efficiency.", "params_upstream": "425.44M", "upstream_fitness": 2.4923, "upstream_acc_8task": 39.46, "upstream_train_loss": 3.4229, "code_dist_deltanet": 0.108, "code_dist_knn5": 0.0132, "code_same_as_parent": null, "code_novelty": 6.0, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 21514, "forward_ms": 4.3, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#828", "index": 828, "name": "FusionGate AdaptiveNet", "internal": "delta_net_aggf_v2", "parent": 565, "summary": "Introduce hierarchical adaptive gating with decaying bias, auxiliary delta-path loss, and learnable per-head temperature for robust fusion.", "params_upstream": "439.13M", "upstream_fitness": 2.2262, "upstream_acc_8task": 39.28, "upstream_train_loss": 3.4231, "code_dist_deltanet": 0.1117, "code_dist_knn5": 0.0101, "code_same_as_parent": 0.436, "code_novelty": 5.5, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 27176, "forward_ms": 4.6, "error": null, "max_future_leak": 1.2695789337158203e-05, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#1350", "index": 1350, "name": "HybridCausalRouter", "internal": "delta_net_hdsr", "parent": 965, "summary": "Introduce hybrid decisive-soft routing with identity residuals, head-aware evidence, annealed epsilon, and cosine entropy scheduling.", "params_upstream": "466.93M", "upstream_fitness": 2.4518, "upstream_acc_8task": 39.04, "upstream_train_loss": 3.4233, "code_dist_deltanet": 0.1133, "code_dist_knn5": 0.0112, "code_same_as_parent": 0.475, "code_novelty": 6.7, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 30452, "forward_ms": 5.0, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#565", "index": 565, "name": "FusionGate-X", "internal": "delta_net_cagf", "parent": 471, "summary": "Introduce content-aware gated fusion with per-head statistics, path-specific bias, and learnable temperature for improved routing.", "params_upstream": "439.13M", "upstream_fitness": 2.6545, "upstream_acc_8task": 39.35, "upstream_train_loss": 3.4235, "code_dist_deltanet": 0.1074, "code_dist_knn5": 0.0051, "code_same_as_parent": null, "code_novelty": 2.3, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 27173, "forward_ms": 4.3, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#1311", "index": 1311, "name": "HybridFlowNet", "internal": "delta_net_cagf_rc_pf_hybrid", "parent": 497, "summary": "Introduce hybrid residual gating with always-on path to fix gradient starvation and restore ultra-local task capacity.", "params_upstream": "439.03M", "upstream_fitness": 2.5259, "upstream_acc_8task": 40.19, "upstream_train_loss": 3.4238, "code_dist_deltanet": 0.1106, "code_dist_knn5": 0.0056, "code_same_as_parent": null, "code_novelty": 3.0, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 26536, "forward_ms": 5.0, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#1370", "index": 1370, "name": "AdaptiveTokenRouter", "internal": "delta_net_tareia", "parent": 979, "summary": "Introduce token-adaptive \u03b5-floors scaling uncertainty per token and head to balance routing precision and exploration.", "params_upstream": "640.70M", "upstream_fitness": 2.682, "upstream_acc_8task": 38.64, "upstream_train_loss": 3.4238, "code_dist_deltanet": 0.1047, "code_dist_knn5": 0.009, "code_same_as_parent": null, "code_novelty": 3.1, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 57524, "forward_ms": 4.8, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#1045", "index": 1045, "name": "AdaptiveHybridGateNet", "internal": "delta_net_aefg_hr", "parent": 965, "summary": "Introduce adaptive annealing gates and hybrid residual scaling for decisive routing and preserved local information flow.", "params_upstream": "491.04M", "upstream_fitness": 2.3014, "upstream_acc_8task": 39.03, "upstream_train_loss": 3.4239, "code_dist_deltanet": 0.1023, "code_dist_knn5": 0.0141, "code_same_as_parent": 0.398, "code_novelty": 5.1, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 35384, "forward_ms": 4.8, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#1360", "index": 1360, "name": "AdaptiveFloorNet-HAF", "internal": "delta_net_hafs", "parent": 864, "summary": "Introduce head-adaptive annealed floors and per-head temperature for balanced routing, preserving local capacity and sharp gating.", "params_upstream": "464.22M", "upstream_fitness": 2.2673, "upstream_acc_8task": 39.42, "upstream_train_loss": 3.4241, "code_dist_deltanet": 0.1139, "code_dist_knn5": 0.0111, "code_same_as_parent": null, "code_novelty": 6.7, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 31096, "forward_ms": 4.7, "error": null, "max_future_leak": 1.341104507446289e-06, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#1393", "index": 1393, "name": "SynerFuse-LGX", "internal": "delta_net_syngf", "parent": 908, "summary": "Introduce synergetic local-global fusion with bounded gates, adaptive residuals, and entropy-aware routing for unified reasoning.", "params_upstream": "439.13M", "upstream_fitness": 2.6338, "upstream_acc_8task": 39.66, "upstream_train_loss": 3.4243, "code_dist_deltanet": 0.1107, "code_dist_knn5": 0.0062, "code_same_as_parent": 0.471, "code_novelty": 3.3, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 27176, "forward_ms": 4.9, "error": null, "max_future_leak": 0.001953125, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#1051", "index": 1051, "name": "ResiFuse-CausalGater", "internal": "delta_net_rggf", "parent": 565, "summary": "Introduce Residual-Guaranteed Gated Fusion for robust local retention, adaptive fusion, and efficient causal masking in DeltaNet.", "params_upstream": "439.13M", "upstream_fitness": 2.1384, "upstream_acc_8task": 39.16, "upstream_train_loss": 3.4243, "code_dist_deltanet": 0.1107, "code_dist_knn5": 0.0063, "code_same_as_parent": 0.798, "code_novelty": 3.3, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 27180, "forward_ms": 4.4, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#947", "index": 947, "name": "GroupTempMLP", "internal": "delta_net_gtmlp", "parent": 682, "summary": "Introduce group-wise temperature and per-head MLP gates for balanced sharpness, redundancy, and expressive conditional routing.", "params_upstream": "426.49M", "upstream_fitness": 2.4492, "upstream_acc_8task": 38.87, "upstream_train_loss": 3.4243, "code_dist_deltanet": 0.1144, "code_dist_knn5": 0.013, "code_same_as_parent": null, "code_novelty": 7.6, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 25133, "forward_ms": 4.8, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#1442", "index": 1442, "name": "DynMemGate", "internal": "delta_net_udmag", "parent": 965, "summary": "Introduce adaptive memory decay, output-aware gating, annealed routing, and corrected entropy for efficient, diverse reasoning.", "params_upstream": "492.71M", "upstream_fitness": 2.4288, "upstream_acc_8task": 39.7, "upstream_train_loss": 3.4247, "code_dist_deltanet": 0.1115, "code_dist_knn5": 0.0103, "code_same_as_parent": 0.426, "code_novelty": 5.6, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 34988, "forward_ms": 5.2, "error": null, "max_future_leak": 9.834766387939453e-07, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#750", "index": 750, "name": "AdaptiveSpanGateConv", "internal": "delta_net_adaptive_hier_gate", "parent": 560, "summary": "Introduce adaptive \u03b5-floor gating with learnable bounds, per-head temperature, and identity-initialised multi-scale convolutions.", "params_upstream": "467.86M", "upstream_fitness": 2.1674, "upstream_acc_8task": 39.93, "upstream_train_loss": 3.4247, "code_dist_deltanet": 0.1015, "code_dist_knn5": 0.0107, "code_same_as_parent": null, "code_novelty": 3.4, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 35952, "forward_ms": 4.6, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#957", "index": 957, "name": "FusionBalanceTransformer", "internal": "delta_net_psafg", "parent": 560, "summary": "Introduce normalised additive fusion with output-aware gating and identity-inclusive FIR pyramid for stable multi-path reasoning.", "params_upstream": "469.04M", "upstream_fitness": 2.6184, "upstream_acc_8task": 39.68, "upstream_train_loss": 3.4108, "code_dist_deltanet": 0.0973, "code_dist_knn5": 0.0091, "code_same_as_parent": null, "code_novelty": 2.2, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 32878, "forward_ms": 4.5, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#675", "index": 675, "name": "FusionGate-MS3E-Hybrid", "internal": "delta_net_gae_ms3e", "parent": 586, "summary": "Introduce head-grouped multi-stat adaptive gating with entropy regularization for specialized routing and efficient multi-scale fusion.", "params_upstream": "431.49M", "upstream_fitness": 2.303, "upstream_acc_8task": 39.62, "upstream_train_loss": 3.4251, "code_dist_deltanet": 0.0951, "code_dist_knn5": 0.0199, "code_same_as_parent": null, "code_novelty": 5.4, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 56686, "forward_ms": 5.5, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#511", "index": 511, "name": "FusionGateMemoryNet", "internal": "delta_net_ms_hsm_tempgate", "parent": 417, "summary": "Introduce fused multi-scale convolutions, hierarchical memory, and adaptive gating for efficient local-global reasoning and faster convergence.", "params_upstream": "466.31M", "upstream_fitness": 2.3767, "upstream_acc_8task": 38.89, "upstream_train_loss": 3.4252, "code_dist_deltanet": 0.1058, "code_dist_knn5": 0.0111, "code_same_as_parent": null, "code_novelty": 4.6, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 32046, "forward_ms": 9.3, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#1454", "index": 1454, "name": "TokenPruneRouter", "internal": "delta_net_tapr", "parent": 1370, "summary": "Introduce token-adaptive floors with progressive pruning to enable leakage-free, one-hot routing for precise late-stage tasks.", "params_upstream": "640.70M", "upstream_fitness": 2.6833, "upstream_acc_8task": 38.71, "upstream_train_loss": 3.4253, "code_dist_deltanet": 0.1056, "code_dist_knn5": 0.0103, "code_same_as_parent": 0.783, "code_novelty": 3.9, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 57524, "forward_ms": 4.9, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#1338", "index": 1338, "name": "ConvexBlendFloorNet", "internal": "delta_net_cagf_mf", "parent": 671, "summary": "Introduce immutable probability floor \u03b5 to ensure balanced path gradients while preserving convex mixture and output stability.", "params_upstream": "439.13M", "upstream_fitness": 2.497, "upstream_acc_8task": 39.62, "upstream_train_loss": 3.4265, "code_dist_deltanet": 0.1075, "code_dist_knn5": 0.007, "code_same_as_parent": null, "code_novelty": 2.8, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 27174, "forward_ms": 4.3, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#1235", "index": 1235, "name": "MinFloorRouter", "internal": "delta_net_mfg", "parent": 497, "summary": "Introduce minimum-floor gating to prevent path starvation, ensuring gradient flow while preserving sharp routing flexibility.", "params_upstream": "439.13M", "upstream_fitness": 2.4249, "upstream_acc_8task": 39.95, "upstream_train_loss": 3.4269, "code_dist_deltanet": 0.1141, "code_dist_knn5": 0.0067, "code_same_as_parent": null, "code_novelty": 4.6, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 27180, "forward_ms": 4.6, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#879", "index": 879, "name": "FusionGatedFIRNet", "internal": "delta_net_psfr", "parent": 565, "summary": "Introduce parallel sigmoid gating to eliminate path competition, enabling simultaneous local and global reasoning improvements.", "params_upstream": "466.90M", "upstream_fitness": 2.5516, "upstream_acc_8task": 40.1, "upstream_train_loss": 3.4048, "code_dist_deltanet": 0.1036, "code_dist_knn5": 0.0129, "code_same_as_parent": 0.336, "code_novelty": 4.9, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 31916, "forward_ms": 4.3, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#1786", "index": 1786, "name": "AdaptiveFusionNet-DSI", "internal": "delta_net_aft_dsi", "parent": 965, "summary": "Introduce token-adaptive routing with scheduled floors and bounded head alphas for balanced copy-context trade-offs.", "params_upstream": "466.51M", "upstream_fitness": 2.282, "upstream_acc_8task": 39.14, "upstream_train_loss": 3.427, "code_dist_deltanet": 0.1076, "code_dist_knn5": 0.0088, "code_same_as_parent": 0.478, "code_novelty": 3.6, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 30123, "forward_ms": 4.4, "error": null, "max_future_leak": 1.055002212524414e-05, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#629", "index": 629, "name": "ConvFusionWide31", "internal": "delta_net_ms_hsm_widefloor", "parent": 522, "summary": "Introduce \u03b5-floor fusion and extended multi-scale convolution (k=31) to enhance span tasks and gradient flow.", "params_upstream": "468.25M", "upstream_fitness": 2.5188, "upstream_acc_8task": 39.52, "upstream_train_loss": 3.4273, "code_dist_deltanet": 0.1139, "code_dist_knn5": 0.0112, "code_same_as_parent": null, "code_novelty": 6.9, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 35570, "forward_ms": 5.3, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#865", "index": 865, "name": "DynFuseFlexGate", "internal": "delta_net_dynfuse", "parent": 565, "summary": "Introduce dynamic gating with decaying local-floor, residual bypass, and entropy regularization to resolve global-local trade-offs.", "params_upstream": "439.13M", "upstream_fitness": 2.6567, "upstream_acc_8task": 39.3, "upstream_train_loss": 3.4274, "code_dist_deltanet": 0.1115, "code_dist_knn5": 0.0054, "code_same_as_parent": 0.526, "code_novelty": 3.8, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 27174, "forward_ms": 4.6, "error": null, "max_future_leak": 2.7477741241455078e-05, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#1507", "index": 1507, "name": "TokenScaleRouter", "internal": "delta_net_tarf", "parent": 965, "summary": "Introduce token-adaptive routing with multi-scale FIR and output-aware context mixing for efficient, context-sensitive reasoning.", "params_upstream": "717.33M", "upstream_fitness": 2.4876, "upstream_acc_8task": 38.92, "upstream_train_loss": 3.4274, "code_dist_deltanet": 0.1008, "code_dist_knn5": 0.0109, "code_same_as_parent": 0.483, "code_novelty": 3.5, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 71598, "forward_ms": 4.5, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#974", "index": 974, "name": "AdaptiveFusionNet", "internal": "delta_net_abrgf", "parent": 565, "summary": "Introduce adaptive bias, Dirac-initialised FIR kernels, residual bypass, and logit-dropout for balanced fusion and gradient flow.", "params_upstream": "438.96M", "upstream_fitness": 2.4764, "upstream_acc_8task": 39.66, "upstream_train_loss": 3.428, "code_dist_deltanet": 0.1082, "code_dist_knn5": 0.0076, "code_same_as_parent": 0.604, "code_novelty": 3.2, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 26732, "forward_ms": 11.7, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#1529", "index": 1529, "name": "HyenaMAFR", "internal": "delta_net_mafr", "parent": 1329, "summary": "Introduce multi-scale adaptive floors, depth-wise FIR convolutions, and per-head residuals for efficient, content-aware routing.", "params_upstream": "440.08M", "upstream_fitness": 2.6189, "upstream_acc_8task": 39.29, "upstream_train_loss": 3.4283, "code_dist_deltanet": 0.1195, "code_dist_knn5": 0.0091, "code_same_as_parent": null, "code_novelty": 6.4, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 30138, "forward_ms": 4.8, "error": null, "max_future_leak": 3.8743019104003906e-07, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#890", "index": 890, "name": "FusionGATE-HMSR", "internal": "delta_net_dmshf", "parent": 497, "summary": "Introduce hybrid hierarchical-statistical gating with adaptive regularization for dynamic multi-path routing and balanced local-global reasoning.", "params_upstream": "473.19M", "upstream_fitness": 2.3497, "upstream_acc_8task": 39.88, "upstream_train_loss": 3.4286, "code_dist_deltanet": 0.1171, "code_dist_knn5": 0.0142, "code_same_as_parent": null, "code_novelty": 8.5, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 34488, "forward_ms": 5.3, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#515", "index": 515, "name": "FusionGate-MS3", "internal": "delta_net_ms_adaptive_gstat3", "parent": 434, "summary": "Introduce adaptive gating with enriched statistics, regularization, and hybrid bias for enhanced context integration and path diversity.", "params_upstream": "464.65M", "upstream_fitness": 2.459, "upstream_acc_8task": 39.24, "upstream_train_loss": 3.4052, "code_dist_deltanet": 0.1111, "code_dist_knn5": 0.0165, "code_same_as_parent": null, "code_novelty": 7.9, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 29146, "forward_ms": 4.7, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#1693", "index": 1693, "name": "AdaptiveMixGateNet", "internal": "delta_net_ahm_gate", "parent": 965, "summary": "Introduce per-head adaptive mixing with confidence-driven suppression and stage-selective temperature annealing for dynamic specialization.", "params_upstream": "439.13M", "upstream_fitness": 2.2209, "upstream_acc_8task": 39.21, "upstream_train_loss": 3.4289, "code_dist_deltanet": 0.1156, "code_dist_knn5": 0.0099, "code_same_as_parent": 0.749, "code_novelty": 6.2, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 27181, "forward_ms": 4.7, "error": null, "max_future_leak": 6.198883056640625e-06, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#1221", "index": 1221, "name": "FusionGateX", "internal": "delta_net_oahmgr", "parent": 671, "summary": "Introduce output-aware gating, annealable epsilon floor, hybrid residual paths, Dirac+noise FIR, and learnable softplus temperature.", "params_upstream": "439.72M", "upstream_fitness": 2.5416, "upstream_acc_8task": 38.58, "upstream_train_loss": 3.4291, "code_dist_deltanet": 0.1172, "code_dist_knn5": 0.0077, "code_same_as_parent": null, "code_novelty": 5.5, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 28330, "forward_ms": 5.3, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#441", "index": 441, "name": "DualFIR-QuadFusion", "internal": "delta_net_msdfdm", "parent": 364, "summary": "Introduce dual FIR branches with adaptive fusion to balance token-level precision and mid-range context, preventing over-smoothing.", "params_upstream": "465.45M", "upstream_fitness": 2.3832, "upstream_acc_8task": 39.57, "upstream_train_loss": 3.4038, "code_dist_deltanet": 0.0998, "code_dist_knn5": 0.0103, "code_same_as_parent": null, "code_novelty": 2.9, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 31208, "forward_ms": 4.4, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#933", "index": 933, "name": "FusionGateX", "internal": "delta_net_cagf_rc_pf", "parent": 671, "summary": "Introduce \u03b5-floor normalised fusion and adaptive residual gating to balance local-global task accuracy without variance inflation.", "params_upstream": "439.33M", "upstream_fitness": 2.7053, "upstream_acc_8task": 39.28, "upstream_train_loss": 3.4298, "code_dist_deltanet": 0.1113, "code_dist_knn5": 0.0044, "code_same_as_parent": null, "code_novelty": 3.3, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 27305, "forward_ms": 4.5, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#1298", "index": 1298, "name": "AdaptiveEntropyRouterNet", "internal": "delta_net_aeoc", "parent": 649, "summary": "Introduce output-conditioned gating with multi-scale convolution, delta memory, and adaptive entropy regularization for robust multi-path routing.", "params_upstream": "491.82M", "upstream_fitness": 2.3271, "upstream_acc_8task": 39.31, "upstream_train_loss": 3.4298, "code_dist_deltanet": 0.1147, "code_dist_knn5": 0.0118, "code_same_as_parent": null, "code_novelty": 7.3, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 53106, "forward_ms": 5.1, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#1448", "index": 1448, "name": "PathGatedFusionNet", "internal": "delta_net_pathgated", "parent": 965, "summary": "Introduce path-aware head gates using fused-output statistics to enhance selective global-local reasoning with wider dynamic range.", "params_upstream": "452.03M", "upstream_fitness": 2.4963, "upstream_acc_8task": 39.72, "upstream_train_loss": 3.4301, "code_dist_deltanet": 0.1154, "code_dist_knn5": 0.0047, "code_same_as_parent": 0.622, "code_novelty": 4.3, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 29674, "forward_ms": 7.9, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#1792", "index": 1792, "name": "FusionGate-HierarchicalRouter", "internal": "delta_net_hgm_ident", "parent": 1544, "summary": "Introduce hierarchical multi-scale gating with adaptive identity routing for superior reasoning, extraction, and abstraction efficiency.", "params_upstream": "471.11M", "upstream_fitness": 2.4894, "upstream_acc_8task": 39.14, "upstream_train_loss": 3.4312, "code_dist_deltanet": 0.1133, "code_dist_knn5": 0.0106, "code_same_as_parent": null, "code_novelty": 6.2, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 33672, "forward_ms": 4.6, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#1132", "index": 1132, "name": "DualScaleStatFusionNet", "internal": "delta_net_hpaf", "parent": 864, "summary": "Introduce dual-scale FIR, per-head adaptive gating, and cross-head statistic mixing for efficient, robust sequence reasoning.", "params_upstream": "471.71M", "upstream_fitness": 2.3205, "upstream_acc_8task": 39.93, "upstream_train_loss": 3.4123, "code_dist_deltanet": 0.1208, "code_dist_knn5": 0.0149, "code_same_as_parent": null, "code_novelty": 9.0, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 35946, "forward_ms": 11.3, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#1252", "index": 1252, "name": "AdaptivePathRouter", "internal": "delta_net_adgr", "parent": 1000, "summary": "Introduce adaptive per-head bias, KL diversity regularization, and learnable epsilon floor to enhance routing flexibility and robustness.", "params_upstream": "418.93M", "upstream_fitness": 2.4715, "upstream_acc_8task": 38.59, "upstream_train_loss": 3.4332, "code_dist_deltanet": 0.0988, "code_dist_knn5": 0.01, "code_same_as_parent": null, "code_novelty": 2.6, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 23748, "forward_ms": 4.4, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#1584", "index": 1584, "name": "AdaptiveFusionNet", "internal": "delta_net_aif", "parent": 497, "summary": "Introduce adaptive token-dependent identity floor and content-position fusion router for dynamic routing and improved task precision.", "params_upstream": "466.71M", "upstream_fitness": 2.5619, "upstream_acc_8task": 39.57, "upstream_train_loss": 3.4336, "code_dist_deltanet": 0.1127, "code_dist_knn5": 0.0092, "code_same_as_parent": null, "code_novelty": 5.4, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 30381, "forward_ms": 5.1, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#1283", "index": 1283, "name": "HybridFusionFloor", "internal": "delta_net_hybfloor", "parent": 865, "summary": "Introduce hybrid floor with reserved \u03b5, per-head temperature, and gated identity residual for balanced local-global routing.", "params_upstream": "464.69M", "upstream_fitness": 2.5159, "upstream_acc_8task": 39.24, "upstream_train_loss": 3.434, "code_dist_deltanet": 0.1114, "code_dist_knn5": 0.0068, "code_same_as_parent": 0.462, "code_novelty": 3.9, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 32302, "forward_ms": 4.5, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#939", "index": 939, "name": "EntropyKLAdaptiveGateNet", "internal": "delta_net_entropy_kl_floor_gate", "parent": 471, "summary": "Introduce entropy-regularized fusion gating with learnable floors and monotonic memory decay for robust path diversity and routing.", "params_upstream": "471.62M", "upstream_fitness": 2.5993, "upstream_acc_8task": 39.5, "upstream_train_loss": 3.4124, "code_dist_deltanet": 0.1115, "code_dist_knn5": 0.0167, "code_same_as_parent": null, "code_novelty": 8.3, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 34996, "forward_ms": 4.7, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#700", "index": 700, "name": "FusionGate-XR", "internal": "delta_net_amf_routing", "parent": 560, "summary": "Introduce adaptive multi-scale fusion with dynamic gating, entropy regularization, and learnable path allocation for robust sequence modeling.", "params_upstream": "469.04M", "upstream_fitness": 2.0675, "upstream_acc_8task": 39.96, "upstream_train_loss": 3.4353, "code_dist_deltanet": 0.0947, "code_dist_knn5": 0.0127, "code_same_as_parent": null, "code_novelty": 3.8, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 32894, "forward_ms": 4.7, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#702", "index": 702, "name": "BlockStateFusionNet", "internal": "delta_net_bscgf", "parent": 580, "summary": "Introduce context-aware fusion gate integrating token statistics and hidden state for adaptive multi-path memory routing.", "params_upstream": "468.47M", "upstream_fitness": 2.5632, "upstream_acc_8task": 39.11, "upstream_train_loss": 3.4085, "code_dist_deltanet": 0.1118, "code_dist_knn5": 0.0124, "code_same_as_parent": null, "code_novelty": 6.9, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 32938, "forward_ms": 4.4, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#830", "index": 830, "name": "LocalGlobalBlendNet", "internal": "delta_net_hwggm", "parent": 497, "summary": "Introduce head-wise local routing with adaptive token-level global mixing to prevent path starvation while preserving efficiency.", "params_upstream": "416.55M", "upstream_fitness": 2.3812, "upstream_acc_8task": 39.54, "upstream_train_loss": 3.4361, "code_dist_deltanet": 0.1165, "code_dist_knn5": 0.0147, "code_same_as_parent": null, "code_novelty": 8.6, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 28903, "forward_ms": 14.9, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#954", "index": 954, "name": "AdaptiveGateNet", "internal": "delta_net_acmg", "parent": 497, "summary": "Introduce adaptive gating with output-aware routing, learnable sharpness, and confidence-conditioned gradient floors for balanced performance.", "params_upstream": "447.31M", "upstream_fitness": 2.1317, "upstream_acc_8task": 39.52, "upstream_train_loss": 3.4377, "code_dist_deltanet": 0.1111, "code_dist_knn5": 0.0109, "code_same_as_parent": null, "code_novelty": 5.7, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 30053, "forward_ms": 4.4, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#804", "index": 804, "name": "PerHeadSimplexRouter", "internal": "delta_net_phsg5", "parent": 565, "summary": "Introduce per-head gated 5-way simplex routing with multi-scale FIRs, learnable annealing, and strict causality.", "params_upstream": "414.73M", "upstream_fitness": 2.2236, "upstream_acc_8task": 39.62, "upstream_train_loss": 3.4116, "code_dist_deltanet": 0.105, "code_dist_knn5": 0.0154, "code_same_as_parent": 0.486, "code_novelty": 6.0, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 23150, "forward_ms": 4.2, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#410", "index": 410, "name": "HybridScale-GateNet", "internal": "delta_net_selm", "parent": 364, "summary": "Introduce selective multi-scale memory with runtime-adaptive gating to resolve local/global tradeoffs and stabilize optimization.", "params_upstream": "469.68M", "upstream_fitness": 2.1193, "upstream_acc_8task": 39.79, "upstream_train_loss": 3.4191, "code_dist_deltanet": 0.1016, "code_dist_knn5": 0.0137, "code_same_as_parent": null, "code_novelty": 4.8, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 33767, "forward_ms": 5.0, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#714", "index": 714, "name": "ResGate_MS_FusionNet", "internal": "delta_net_ms_resgate", "parent": 560, "summary": "Introduce reserve gating with epsilon floor to ensure global path capacity while preserving sharp local branch selectivity.", "params_upstream": "468.23M", "upstream_fitness": 2.3629, "upstream_acc_8task": 38.87, "upstream_train_loss": 3.4126, "code_dist_deltanet": 0.104, "code_dist_knn5": 0.0107, "code_same_as_parent": null, "code_novelty": 4.0, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 32308, "forward_ms": 4.5, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#1595", "index": 1595, "name": "GateFusionNet", "internal": "delta_net_sigf_ptu", "parent": 1544, "summary": "Introduce rich statistical gating, dynamic identity paths, annealed \u03b5-floors, and progressive temperature untying for enhanced task performance.", "params_upstream": "472.76M", "upstream_fitness": 2.5649, "upstream_acc_8task": 39.1, "upstream_train_loss": 3.4415, "code_dist_deltanet": 0.1096, "code_dist_knn5": 0.0073, "code_same_as_parent": null, "code_novelty": 3.3, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 35688, "forward_ms": 4.9, "error": null, "max_future_leak": 5.289912223815918e-05, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#887", "index": 887, "name": "EntropyFlowGateNet", "internal": "delta_net_erfg", "parent": 649, "summary": "Introduce entropy-regularised gating and scheduled forget mechanism to prevent path collapse and improve long-memory retention.", "params_upstream": "489.71M", "upstream_fitness": 2.2415, "upstream_acc_8task": 39.73, "upstream_train_loss": 3.4166, "code_dist_deltanet": 0.1101, "code_dist_knn5": 0.0131, "code_same_as_parent": null, "code_novelty": 6.2, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 45248, "forward_ms": 4.5, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#948", "index": 948, "name": "AdaptiveFloorGate", "internal": "delta_net_acfg", "parent": 671, "summary": "Introduce adaptive gating with dynamic context floors ensuring gradient flow while optimizing copy-context trade-off per token.", "params_upstream": "464.74M", "upstream_fitness": 2.2845, "upstream_acc_8task": 39.5, "upstream_train_loss": 3.4418, "code_dist_deltanet": 0.1018, "code_dist_knn5": 0.0112, "code_same_as_parent": null, "code_novelty": 4.0, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 29002, "forward_ms": 4.2, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#1041", "index": 1041, "name": "ResConvGate", "internal": "delta_net_crdg", "parent": 565, "summary": "Introduce residual convolutional paths and stochastic dropout to prevent gate collapse and ensure balanced path utilization.", "params_upstream": "439.13M", "upstream_fitness": 2.3968, "upstream_acc_8task": 38.84, "upstream_train_loss": 3.4418, "code_dist_deltanet": 0.111, "code_dist_knn5": 0.0065, "code_same_as_parent": 0.497, "code_novelty": 3.5, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 27175, "forward_ms": 5.3, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#843", "index": 843, "name": "DualStagePathGateNet", "internal": "delta_net_hsigctx", "parent": 730, "summary": "Introduce two-stage head-wise gating to decouple identity routing from contextual paths, eliminating zero-sum trade-offs.", "params_upstream": "413.67M", "upstream_fitness": 2.4413, "upstream_acc_8task": 39.43, "upstream_train_loss": 3.4428, "code_dist_deltanet": 0.1089, "code_dist_knn5": 0.015, "code_same_as_parent": null, "code_novelty": 6.9, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 20778, "forward_ms": 13.4, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#1444", "index": 1444, "name": "FusionGate-XL", "internal": "delta_net_hmgapf", "parent": 682, "summary": "Introduce parallel additive fusion with adaptive residuals and enhanced content-aware gating for robust, balanced sequence modeling.", "params_upstream": "433.80M", "upstream_fitness": 2.3478, "upstream_acc_8task": 39.5, "upstream_train_loss": 3.443, "code_dist_deltanet": 0.111, "code_dist_knn5": 0.0161, "code_same_as_parent": null, "code_novelty": 7.7, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 27362, "forward_ms": 4.6, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#641", "index": 641, "name": "HeadWiseGateNet", "internal": "delta_net_hwg", "parent": 403, "summary": "Introduce head-wise output-aware gating to restore gradient flow, enable specialisation, and prevent head-collapse in DeltaNet.", "params_upstream": "413.67M", "upstream_fitness": 2.5573, "upstream_acc_8task": 39.79, "upstream_train_loss": 3.4131, "code_dist_deltanet": 0.1012, "code_dist_knn5": 0.0118, "code_same_as_parent": null, "code_novelty": 4.1, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 20776, "forward_ms": 4.4, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#519", "index": 519, "name": "SparseGateDelta", "internal": "delta_net_ssg_sparsemax_temp", "parent": 401, "summary": "Introduce temperature-controlled sparsemax gating for adaptive sparsity, enhancing local-precision by eliminating forced probability blending.", "params_upstream": "425.33M", "upstream_fitness": 2.3547, "upstream_acc_8task": 39.41, "upstream_train_loss": 3.4433, "code_dist_deltanet": 0.0894, "code_dist_knn5": 0.017, "code_same_as_parent": null, "code_novelty": 5.1, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 20050, "forward_ms": 34.1, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#645", "index": 645, "name": "AdaptiveHierGateNet", "internal": "delta_net_htcg", "parent": 403, "summary": "Introduce hierarchical temperature-controlled gating with adaptive entropy to balance local precision and global memory routing.", "params_upstream": "413.38M", "upstream_fitness": 1.9569, "upstream_acc_8task": 39.43, "upstream_train_loss": 3.4433, "code_dist_deltanet": 0.1109, "code_dist_knn5": 0.0117, "code_same_as_parent": null, "code_novelty": 6.0, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 20010, "forward_ms": 4.3, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#965", "index": 965, "name": "GateFlooredResNet", "internal": "delta_net_cagf_br", "parent": 671, "summary": "Introduce probability-floored gating and dynamics-aware residuals to stabilize variance and prevent convolutional path double-counting.", "params_upstream": "439.13M", "upstream_fitness": 2.7247, "upstream_acc_8task": 39.05, "upstream_train_loss": 3.4441, "code_dist_deltanet": 0.1107, "code_dist_knn5": 0.0054, "code_same_as_parent": null, "code_novelty": 3.0, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 27175, "forward_ms": 4.5, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#1349", "index": 1349, "name": "FusionFeedback-MixNormNet", "internal": "delta_net_msfr_mn", "parent": 965, "summary": "Introduce Multi-Scale Feedback-Routed MixNorm for adaptive memory, relational routing, and variance-stabilized multi-path DeltaNet processing.", "params_upstream": "476.31M", "upstream_fitness": 2.4106, "upstream_acc_8task": 39.62, "upstream_train_loss": 3.4452, "code_dist_deltanet": 0.1115, "code_dist_knn5": 0.0103, "code_same_as_parent": 0.346, "code_novelty": 5.4, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 38016, "forward_ms": 4.6, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#676", "index": 676, "name": "StatGateRouter", "internal": "delta_net_ser_minfloor", "parent": 556, "summary": "Introduce statistic-enriched routing, minimum-floor gating, identity bypass, Dirac-initialized convolutions, and optional entropy regularization.", "params_upstream": "471.51M", "upstream_fitness": 2.2237, "upstream_acc_8task": 39.33, "upstream_train_loss": 3.4465, "code_dist_deltanet": 0.1069, "code_dist_knn5": 0.0093, "code_same_as_parent": 0.59, "code_novelty": 3.8, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 33964, "forward_ms": 13.7, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#295", "index": 295, "name": "DynamicMemGateNet", "internal": "delta_net_dyn_gate_mix", "parent": 137, "summary": "Introduce dynamic per-token, per-head gating to blend Delta-rule and EMA memory for adaptive context/task precision.", "params_upstream": "412.15M", "upstream_fitness": 2.2621, "upstream_acc_8task": 38.59, "upstream_train_loss": 3.4469, "code_dist_deltanet": 0.0921, "code_dist_knn5": 0.0162, "code_same_as_parent": null, "code_novelty": 5.1, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 17570, "forward_ms": 1790.3, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#1175", "index": 1175, "name": "AnnealedPathFusionNet", "internal": "delta_net_annealed_eklf", "parent": 908, "summary": "Introduce annealed regularisation, temperature floor, and structural hard floor to prevent path collapse and improve routing.", "params_upstream": "442.55M", "upstream_fitness": 2.1498, "upstream_acc_8task": 39.57, "upstream_train_loss": 3.4472, "code_dist_deltanet": 0.1105, "code_dist_knn5": 0.0134, "code_same_as_parent": 0.463, "code_novelty": 6.3, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 28408, "forward_ms": 5.0, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#988", "index": 988, "name": "LexiFuse-Percept", "internal": "delta_net_pfr", "parent": 865, "summary": "Introduce persistent local-floor schedule and per-head residual modulation for balanced local-global routing in DeltaNet.", "params_upstream": "439.13M", "upstream_fitness": 2.6696, "upstream_acc_8task": 39.8, "upstream_train_loss": 3.4327, "code_dist_deltanet": 0.1113, "code_dist_knn5": 0.0046, "code_same_as_parent": 0.766, "code_novelty": 3.3, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 27175, "forward_ms": 4.6, "error": null, "max_future_leak": 2.0503997802734375e-05, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#1031", "index": 1031, "name": "FusionGate-X", "internal": "delta_net_hefth", "parent": 580, "summary": "Introduce \u03b5-floor, length-aware scaling, and talking-heads mixer to address gate-collapse, local-swamp, and head isolation.", "params_upstream": "471.70M", "upstream_fitness": 2.2218, "upstream_acc_8task": 39.28, "upstream_train_loss": 3.4512, "code_dist_deltanet": 0.1088, "code_dist_knn5": 0.0105, "code_same_as_parent": null, "code_novelty": 4.8, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 35182, "forward_ms": 5.1, "error": null, "max_future_leak": 0.0078125, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#1733", "index": 1733, "name": "FusionGate-XL", "internal": "delta_net_ddfsanr", "parent": 1544, "summary": "Introduce dynamic contextual residual gating with enriched stats, adaptive normalization, and progressive specialization for balanced reasoning-extraction fusion.", "params_upstream": "451.84M", "upstream_fitness": 2.1085, "upstream_acc_8task": 39.63, "upstream_train_loss": 3.4536, "code_dist_deltanet": 0.1157, "code_dist_knn5": 0.0078, "code_same_as_parent": null, "code_novelty": 5.4, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 29587, "forward_ms": 5.1, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#330", "index": 330, "name": "AdaptiveFractalGateNet", "internal": "delta_net_dyn_decay_fractal_gate", "parent": 238, "summary": "Introduce adaptive decay gating and gated fractal mixing to balance global reasoning and local precision under sub-quadratic complexity.", "params_upstream": "412.20M", "upstream_fitness": 2.1079, "upstream_acc_8task": 39.88, "upstream_train_loss": 3.4351, "code_dist_deltanet": 0.096, "code_dist_knn5": 0.016, "code_same_as_parent": null, "code_novelty": 5.1, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 18083, "forward_ms": 970.4, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#1408", "index": 1408, "name": "AdaptiveEntropyGateNet", "internal": "delta_net_efagm", "parent": 864, "summary": "Introduce entropy-floored adaptive-feedback gated memory for dynamic path routing, diversity preservation, and robust task performance.", "params_upstream": "496.59M", "upstream_fitness": 2.2337, "upstream_acc_8task": 38.72, "upstream_train_loss": 3.4558, "code_dist_deltanet": 0.1191, "code_dist_knn5": 0.0139, "code_same_as_parent": null, "code_novelty": 8.6, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 37044, "forward_ms": 5.4, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#1155", "index": 1155, "name": "TriScaleFusionNet", "internal": "delta_net_triscale", "parent": 865, "summary": "Introduce tri-scale FIR memory with persistent floor and 5-way fusion gate for enhanced locality and mid-range context.", "params_upstream": "439.62M", "upstream_fitness": 2.5772, "upstream_acc_8task": 39.41, "upstream_train_loss": 3.4318, "code_dist_deltanet": 0.1134, "code_dist_knn5": 0.0073, "code_same_as_parent": 0.726, "code_novelty": 4.7, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 28456, "forward_ms": 4.9, "error": null, "max_future_leak": 9.834766387939453e-07, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#259", "index": 259, "name": "BAMG_MemoryGate", "internal": "delta_net_bias_init_mix_gate", "parent": 151, "summary": "Introduce bias-initialised adaptive mixing gate to prevent premature gate collapse, enhancing long-range memory utilization during training.", "params_upstream": "411.95M", "upstream_fitness": 2.0173, "upstream_acc_8task": 39.35, "upstream_train_loss": 3.4587, "code_dist_deltanet": 0.0736, "code_dist_knn5": 0.0115, "code_same_as_parent": 0.628, "code_novelty": 3.1, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 17444, "forward_ms": 3.8, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#380", "index": 380, "name": "DualScaleGateNet", "internal": "delta_net_ndg", "parent": 302, "summary": "Introduce normalised dual-scale gating for output mixing and state updates, ensuring balanced global-local context integration.", "params_upstream": "412.34M", "upstream_fitness": 2.113, "upstream_acc_8task": 38.59, "upstream_train_loss": 3.4589, "code_dist_deltanet": 0.1027, "code_dist_knn5": 0.0158, "code_same_as_parent": null, "code_novelty": 5.8, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 17700, "forward_ms": 7.1, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#1081", "index": 1081, "name": "EntropyFusionNormX", "internal": "delta_net_entropy_cagf_rc_norm", "parent": 671, "summary": "Introduce entropy-regularized gating, RMSNorm fusion, and dynamic residual scaling for robust DeltaNet multi-path performance.", "params_upstream": "451.83M", "upstream_fitness": 2.3826, "upstream_acc_8task": 40.06, "upstream_train_loss": 3.4592, "code_dist_deltanet": 0.1106, "code_dist_knn5": 0.0072, "code_same_as_parent": null, "code_novelty": 3.5, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 29351, "forward_ms": 4.7, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#151", "index": 151, "name": "AdaptiveMixTransformer", "internal": "delta_net_adaptive_mix_gate", "parent": 1, "summary": "Introduce Adaptive Mixing Gate for dynamic recurrent-attention blending, enhancing length extrapolation and task robustness with minimal overhead.", "params_upstream": "411.95M", "upstream_fitness": 2.078, "upstream_acc_8task": 38.55, "upstream_train_loss": 3.4593, "code_dist_deltanet": 0.0646, "code_dist_knn5": 0.0158, "code_same_as_parent": null, "code_novelty": 4.6, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 17440, "forward_ms": 3.9, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#437", "index": 437, "name": "FusionConv-AMG", "internal": "delta_net_dual_path_fusion", "parent": 268, "summary": "Introduce additive dilated-conv fusion with decoupled gate bias and AMG for balanced local-global reasoning and stable optimisation.", "params_upstream": "437.21M", "upstream_fitness": 2.2909, "upstream_acc_8task": 38.39, "upstream_train_loss": 3.4593, "code_dist_deltanet": 0.0835, "code_dist_knn5": 0.0145, "code_same_as_parent": null, "code_novelty": 4.2, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 21792, "forward_ms": 4.1, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#709", "index": 709, "name": "HierGate-MEM", "internal": "delta_net_htgmsm", "parent": 403, "summary": "Introduce hierarchical two-stage gating with dual-scale EMA for efficient, adaptive local-global memory routing and precision.", "params_upstream": "425.33M", "upstream_fitness": 2.5176, "upstream_acc_8task": 39.64, "upstream_train_loss": 3.4398, "code_dist_deltanet": 0.1184, "code_dist_knn5": 0.0189, "code_same_as_parent": null, "code_novelty": 9.6, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 20052, "forward_ms": 8.1, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#562", "index": 562, "name": "SparseTempGateNet", "internal": "delta_net_sparsemax_temperature", "parent": 401, "summary": "Introduce sparsemax gating with learnable per-head temperature to enable sharper, adaptive multi-scale memory routing.", "params_upstream": "425.33M", "upstream_fitness": 2.2204, "upstream_acc_8task": 39.1, "upstream_train_loss": null, "code_dist_deltanet": 0.0912, "code_dist_knn5": 0.0145, "code_same_as_parent": null, "code_novelty": 4.4, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 20050, "forward_ms": 10.7, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#1490", "index": 1490, "name": "AdaptiveContextFusionNet", "internal": "delta_net_ahic", "parent": 1409, "summary": "Introduce adaptive hybrid identity-context routing with dynamic copy-floor, bounded scaling, and annealed entropy regularization for robust reasoning.", "params_upstream": "466.71M", "upstream_fitness": 2.1154, "upstream_acc_8task": 38.56, "upstream_train_loss": 3.4624, "code_dist_deltanet": 0.1124, "code_dist_knn5": 0.0086, "code_same_as_parent": null, "code_novelty": 5.0, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 30253, "forward_ms": 4.8, "error": null, "max_future_leak": 5.364418029785156e-07, "grads_finite": true, "dtype": "fp32"}}, {"kind": "asi105", "id": "ASI#322", "index": 322, "name": "AdaptiveTokenGate", "internal": "delta_net_head_gate_ema", "parent": 137, "summary": "Introduce fine-grained per-head/token mixing gate for adaptive Delta-EMA fusion, enhancing precision and generalization efficiently.", "params_upstream": "412.15M", "upstream_fitness": 2.1269, "upstream_acc_8task": 39.94, "upstream_train_loss": 3.44, "code_dist_deltanet": 0.0974, "code_dist_knn5": 0.0162, "code_same_as_parent": null, "code_novelty": 5.3, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 17579, "forward_ms": 3076.4, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#1621", "index": 1621, "name": "HybridGateTransformer", "internal": "delta_net_hhgass", "parent": 1393, "summary": "Introduce hybrid hierarchical gating with adaptive scheduling for dynamic routing, sharp selectivity, and robust identity-processor balance.", "params_upstream": "464.71M", "upstream_fitness": 2.3731, "upstream_acc_8task": 39.49, "upstream_train_loss": 3.4656, "code_dist_deltanet": 0.1145, "code_dist_knn5": 0.0103, "code_same_as_parent": 0.57, "code_novelty": 6.1, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 32362, "forward_ms": 5.7, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "bf16-autocast (as upstream trained it)"}}, {"kind": "asi105", "id": "ASI#720", "index": 720, "name": "HybridSparseGateMemoryNet", "internal": "delta_net_hsgm", "parent": 522, "summary": "Introduce temperature-controlled sparsemax gating and orthogonal-noise FIR branches to resolve path-collapse and gradient starvation issues.", "params_upstream": "425.58M", "upstream_fitness": 2.5803, "upstream_acc_8task": 39.37, "upstream_train_loss": 3.4469, "code_dist_deltanet": 0.1082, "code_dist_knn5": 0.0139, "code_same_as_parent": null, "code_novelty": 6.3, "source": "https://github.com/GAIR-NLP/ASI-Arch/blob/3113c51b04a1d4a7520a3b0238c6a08d3194d58f/data/106.json", "status": "upstream-only", "real_test": {"ok": true, "causal_ok": true, "params_d64": 21834, "forward_ms": 4.5, "error": null, "max_future_leak": 0.0, "grads_finite": true, "dtype": "fp32"}}, {"kind": "grid200", "id": "ASI200-001", "name": "ar_dense fp w32 L2 H1", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": false, "width": 32, "depth": 2, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-002", "name": "ar_dense fp w32 L2 H2", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": false, "width": 32, "depth": 2, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-003", "name": "ar_dense fp w32 L2 H4", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": false, "width": 32, "depth": 2, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-004", "name": "ar_dense fp w32 L2 H8", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": false, "width": 32, "depth": 2, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-005", "name": "ar_dense fp w32 L2 H16", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": false, "width": 32, "depth": 2, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-006", "name": "ar_dense fp w32 L4 H1", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": false, "width": 32, "depth": 4, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-007", "name": "ar_dense fp w32 L4 H2", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": false, "width": 32, "depth": 4, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-008", "name": "ar_dense fp w32 L4 H4", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": false, "width": 32, "depth": 4, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-009", "name": "ar_dense fp w32 L4 H8", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": false, "width": 32, "depth": 4, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-010", "name": "ar_dense fp w32 L4 H16", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": false, "width": 32, "depth": 4, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-011", "name": "ar_dense fp w48 L2 H1", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": false, "width": 48, "depth": 2, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-012", "name": "ar_dense fp w48 L2 H2", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": false, "width": 48, "depth": 2, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-013", "name": "ar_dense fp w48 L2 H4", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": false, "width": 48, "depth": 2, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-014", "name": "ar_dense fp w48 L2 H8", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": false, "width": 48, "depth": 2, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-015", "name": "ar_dense fp w48 L2 H16", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": false, "width": 48, "depth": 2, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-016", "name": "ar_dense fp w48 L4 H1", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": false, "width": 48, "depth": 4, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-017", "name": "ar_dense fp w48 L4 H2", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": false, "width": 48, "depth": 4, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-018", "name": "ar_dense fp w48 L4 H4", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": false, "width": 48, "depth": 4, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-019", "name": "ar_dense fp w48 L4 H8", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": false, "width": 48, "depth": 4, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-020", "name": "ar_dense fp w48 L4 H16", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": false, "width": 48, "depth": 4, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-021", "name": "ar_dense binary w32 L2 H1", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": true, "width": 32, "depth": 2, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-022", "name": "ar_dense binary w32 L2 H2", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": true, "width": 32, "depth": 2, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-023", "name": "ar_dense binary w32 L2 H4", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": true, "width": 32, "depth": 2, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-024", "name": "ar_dense binary w32 L2 H8", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": true, "width": 32, "depth": 2, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-025", "name": "ar_dense binary w32 L2 H16", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": true, "width": 32, "depth": 2, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-026", "name": "ar_dense binary w32 L4 H1", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": true, "width": 32, "depth": 4, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-027", "name": "ar_dense binary w32 L4 H2", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": true, "width": 32, "depth": 4, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-028", "name": "ar_dense binary w32 L4 H4", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": true, "width": 32, "depth": 4, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-029", "name": "ar_dense binary w32 L4 H8", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": true, "width": 32, "depth": 4, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-030", "name": "ar_dense binary w32 L4 H16", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": true, "width": 32, "depth": 4, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-031", "name": "ar_dense binary w48 L2 H1", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": true, "width": 48, "depth": 2, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-032", "name": "ar_dense binary w48 L2 H2", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": true, "width": 48, "depth": 2, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-033", "name": "ar_dense binary w48 L2 H4", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": true, "width": 48, "depth": 2, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-034", "name": "ar_dense binary w48 L2 H8", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": true, "width": 48, "depth": 2, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-035", "name": "ar_dense binary w48 L2 H16", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": true, "width": 48, "depth": 2, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-036", "name": "ar_dense binary w48 L4 H1", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": true, "width": 48, "depth": 4, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-037", "name": "ar_dense binary w48 L4 H2", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": true, "width": 48, "depth": 4, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-038", "name": "ar_dense binary w48 L4 H4", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": true, "width": 48, "depth": 4, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-039", "name": "ar_dense binary w48 L4 H8", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": true, "width": 48, "depth": 4, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-040", "name": "ar_dense binary w48 L4 H16", "family": "ar_dense", "spec": {"family": "ar_dense", "binary_transformer_projections": true, "width": 48, "depth": 4, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-041", "name": "ar_local16 fp w32 L2 H1", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": false, "width": 32, "depth": 2, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-042", "name": "ar_local16 fp w32 L2 H2", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": false, "width": 32, "depth": 2, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-043", "name": "ar_local16 fp w32 L2 H4", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": false, "width": 32, "depth": 2, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-044", "name": "ar_local16 fp w32 L2 H8", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": false, "width": 32, "depth": 2, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-045", "name": "ar_local16 fp w32 L2 H16", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": false, "width": 32, "depth": 2, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-046", "name": "ar_local16 fp w32 L4 H1", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": false, "width": 32, "depth": 4, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-047", "name": "ar_local16 fp w32 L4 H2", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": false, "width": 32, "depth": 4, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-048", "name": "ar_local16 fp w32 L4 H4", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": false, "width": 32, "depth": 4, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-049", "name": "ar_local16 fp w32 L4 H8", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": false, "width": 32, "depth": 4, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-050", "name": "ar_local16 fp w32 L4 H16", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": false, "width": 32, "depth": 4, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-051", "name": "ar_local16 fp w48 L2 H1", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": false, "width": 48, "depth": 2, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-052", "name": "ar_local16 fp w48 L2 H2", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": false, "width": 48, "depth": 2, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-053", "name": "ar_local16 fp w48 L2 H4", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": false, "width": 48, "depth": 2, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-054", "name": "ar_local16 fp w48 L2 H8", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": false, "width": 48, "depth": 2, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-055", "name": "ar_local16 fp w48 L2 H16", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": false, "width": 48, "depth": 2, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-056", "name": "ar_local16 fp w48 L4 H1", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": false, "width": 48, "depth": 4, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-057", "name": "ar_local16 fp w48 L4 H2", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": false, "width": 48, "depth": 4, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-058", "name": "ar_local16 fp w48 L4 H4", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": false, "width": 48, "depth": 4, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-059", "name": "ar_local16 fp w48 L4 H8", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": false, "width": 48, "depth": 4, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-060", "name": "ar_local16 fp w48 L4 H16", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": false, "width": 48, "depth": 4, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-061", "name": "ar_local16 binary w32 L2 H1", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": true, "width": 32, "depth": 2, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-062", "name": "ar_local16 binary w32 L2 H2", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": true, "width": 32, "depth": 2, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-063", "name": "ar_local16 binary w32 L2 H4", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": true, "width": 32, "depth": 2, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-064", "name": "ar_local16 binary w32 L2 H8", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": true, "width": 32, "depth": 2, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-065", "name": "ar_local16 binary w32 L2 H16", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": true, "width": 32, "depth": 2, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-066", "name": "ar_local16 binary w32 L4 H1", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": true, "width": 32, "depth": 4, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-067", "name": "ar_local16 binary w32 L4 H2", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": true, "width": 32, "depth": 4, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-068", "name": "ar_local16 binary w32 L4 H4", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": true, "width": 32, "depth": 4, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-069", "name": "ar_local16 binary w32 L4 H8", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": true, "width": 32, "depth": 4, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-070", "name": "ar_local16 binary w32 L4 H16", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": true, "width": 32, "depth": 4, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-071", "name": "ar_local16 binary w48 L2 H1", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": true, "width": 48, "depth": 2, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-072", "name": "ar_local16 binary w48 L2 H2", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": true, "width": 48, "depth": 2, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-073", "name": "ar_local16 binary w48 L2 H4", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": true, "width": 48, "depth": 2, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-074", "name": "ar_local16 binary w48 L2 H8", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": true, "width": 48, "depth": 2, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-075", "name": "ar_local16 binary w48 L2 H16", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": true, "width": 48, "depth": 2, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-076", "name": "ar_local16 binary w48 L4 H1", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": true, "width": 48, "depth": 4, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-077", "name": "ar_local16 binary w48 L4 H2", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": true, "width": 48, "depth": 4, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-078", "name": "ar_local16 binary w48 L4 H4", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": true, "width": 48, "depth": 4, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-079", "name": "ar_local16 binary w48 L4 H8", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": true, "width": 48, "depth": 4, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-080", "name": "ar_local16 binary w48 L4 H16", "family": "ar_local16", "spec": {"family": "ar_local16", "binary_transformer_projections": true, "width": 48, "depth": 4, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 1}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-081", "name": "diffusionblocks_ar fp w32 L2 H1", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": false, "width": 32, "depth": 2, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 2}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-082", "name": "diffusionblocks_ar fp w32 L2 H2", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": false, "width": 32, "depth": 2, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 2}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-083", "name": "diffusionblocks_ar fp w32 L2 H4", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": false, "width": 32, "depth": 2, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 2}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-084", "name": "diffusionblocks_ar fp w32 L2 H8", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": false, "width": 32, "depth": 2, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 2}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-085", "name": "diffusionblocks_ar fp w32 L2 H16", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": false, "width": 32, "depth": 2, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 2}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-086", "name": "diffusionblocks_ar fp w32 L4 H1", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": false, "width": 32, "depth": 4, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 4}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-087", "name": "diffusionblocks_ar fp w32 L4 H2", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": false, "width": 32, "depth": 4, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 4}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-088", "name": "diffusionblocks_ar fp w32 L4 H4", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": false, "width": 32, "depth": 4, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 4}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-089", "name": "diffusionblocks_ar fp w32 L4 H8", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": false, "width": 32, "depth": 4, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 4}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-090", "name": "diffusionblocks_ar fp w32 L4 H16", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": false, "width": 32, "depth": 4, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 4}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-091", "name": "diffusionblocks_ar fp w48 L2 H1", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": false, "width": 48, "depth": 2, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 2}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-092", "name": "diffusionblocks_ar fp w48 L2 H2", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": false, "width": 48, "depth": 2, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 2}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-093", "name": "diffusionblocks_ar fp w48 L2 H4", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": false, "width": 48, "depth": 2, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 2}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-094", "name": "diffusionblocks_ar fp w48 L2 H8", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": false, "width": 48, "depth": 2, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 2}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-095", "name": "diffusionblocks_ar fp w48 L2 H16", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": false, "width": 48, "depth": 2, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 2}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-096", "name": "diffusionblocks_ar fp w48 L4 H1", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": false, "width": 48, "depth": 4, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 4}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-097", "name": "diffusionblocks_ar fp w48 L4 H2", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": false, "width": 48, "depth": 4, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 4}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-098", "name": "diffusionblocks_ar fp w48 L4 H4", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": false, "width": 48, "depth": 4, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 4}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-099", "name": "diffusionblocks_ar fp w48 L4 H8", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": false, "width": 48, "depth": 4, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 4}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-100", "name": "diffusionblocks_ar fp w48 L4 H16", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": false, "width": 48, "depth": 4, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 4}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-101", "name": "diffusionblocks_ar binary w32 L2 H1", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": true, "width": 32, "depth": 2, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 2}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-102", "name": "diffusionblocks_ar binary w32 L2 H2", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": true, "width": 32, "depth": 2, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 2}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-103", "name": "diffusionblocks_ar binary w32 L2 H4", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": true, "width": 32, "depth": 2, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 2}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-104", "name": "diffusionblocks_ar binary w32 L2 H8", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": true, "width": 32, "depth": 2, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 2}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-105", "name": "diffusionblocks_ar binary w32 L2 H16", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": true, "width": 32, "depth": 2, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 2}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-106", "name": "diffusionblocks_ar binary w32 L4 H1", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": true, "width": 32, "depth": 4, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 4}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-107", "name": "diffusionblocks_ar binary w32 L4 H2", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": true, "width": 32, "depth": 4, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 4}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-108", "name": "diffusionblocks_ar binary w32 L4 H4", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": true, "width": 32, "depth": 4, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 4}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-109", "name": "diffusionblocks_ar binary w32 L4 H8", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": true, "width": 32, "depth": 4, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 4}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-110", "name": "diffusionblocks_ar binary w32 L4 H16", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": true, "width": 32, "depth": 4, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 4}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-111", "name": "diffusionblocks_ar binary w48 L2 H1", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": true, "width": 48, "depth": 2, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 2}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-112", "name": "diffusionblocks_ar binary w48 L2 H2", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": true, "width": 48, "depth": 2, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 2}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-113", "name": "diffusionblocks_ar binary w48 L2 H4", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": true, "width": 48, "depth": 2, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 2}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-114", "name": "diffusionblocks_ar binary w48 L2 H8", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": true, "width": 48, "depth": 2, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 2}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-115", "name": "diffusionblocks_ar binary w48 L2 H16", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": true, "width": 48, "depth": 2, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 2}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-116", "name": "diffusionblocks_ar binary w48 L4 H1", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": true, "width": 48, "depth": 4, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 4}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-117", "name": "diffusionblocks_ar binary w48 L4 H2", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": true, "width": 48, "depth": 4, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 4}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-118", "name": "diffusionblocks_ar binary w48 L4 H4", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": true, "width": 48, "depth": 4, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 4}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-119", "name": "diffusionblocks_ar binary w48 L4 H8", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": true, "width": 48, "depth": 4, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 4}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-120", "name": "diffusionblocks_ar binary w48 L4 H16", "family": "diffusionblocks_ar", "spec": {"family": "diffusionblocks_ar", "binary_transformer_projections": true, "width": 48, "depth": 4, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 1, "training_blocks": 4}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-121", "name": "diffusionblocks_sat8 fp w32 L2 H1", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": false, "width": 32, "depth": 2, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 2}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-122", "name": "diffusionblocks_sat8 fp w32 L2 H2", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": false, "width": 32, "depth": 2, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 2}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-123", "name": "diffusionblocks_sat8 fp w32 L2 H4", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": false, "width": 32, "depth": 2, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 2}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-124", "name": "diffusionblocks_sat8 fp w32 L2 H8", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": false, "width": 32, "depth": 2, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 2}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-125", "name": "diffusionblocks_sat8 fp w32 L2 H16", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": false, "width": 32, "depth": 2, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 2}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-126", "name": "diffusionblocks_sat8 fp w32 L4 H1", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": false, "width": 32, "depth": 4, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 4}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-127", "name": "diffusionblocks_sat8 fp w32 L4 H2", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": false, "width": 32, "depth": 4, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 4}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-128", "name": "diffusionblocks_sat8 fp w32 L4 H4", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": false, "width": 32, "depth": 4, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 4}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-129", "name": "diffusionblocks_sat8 fp w32 L4 H8", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": false, "width": 32, "depth": 4, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 4}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-130", "name": "diffusionblocks_sat8 fp w32 L4 H16", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": false, "width": 32, "depth": 4, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 4}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-131", "name": "diffusionblocks_sat8 fp w48 L2 H1", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": false, "width": 48, "depth": 2, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 2}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-132", "name": "diffusionblocks_sat8 fp w48 L2 H2", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": false, "width": 48, "depth": 2, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 2}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-133", "name": "diffusionblocks_sat8 fp w48 L2 H4", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": false, "width": 48, "depth": 2, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 2}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-134", "name": "diffusionblocks_sat8 fp w48 L2 H8", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": false, "width": 48, "depth": 2, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 2}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-135", "name": "diffusionblocks_sat8 fp w48 L2 H16", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": false, "width": 48, "depth": 2, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 2}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-136", "name": "diffusionblocks_sat8 fp w48 L4 H1", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": false, "width": 48, "depth": 4, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 4}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-137", "name": "diffusionblocks_sat8 fp w48 L4 H2", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": false, "width": 48, "depth": 4, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 4}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-138", "name": "diffusionblocks_sat8 fp w48 L4 H4", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": false, "width": 48, "depth": 4, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 4}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-139", "name": "diffusionblocks_sat8 fp w48 L4 H8", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": false, "width": 48, "depth": 4, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 4}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-140", "name": "diffusionblocks_sat8 fp w48 L4 H16", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": false, "width": 48, "depth": 4, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 4}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-141", "name": "diffusionblocks_sat8 binary w32 L2 H1", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": true, "width": 32, "depth": 2, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 2}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-142", "name": "diffusionblocks_sat8 binary w32 L2 H2", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": true, "width": 32, "depth": 2, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 2}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-143", "name": "diffusionblocks_sat8 binary w32 L2 H4", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": true, "width": 32, "depth": 2, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 2}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-144", "name": "diffusionblocks_sat8 binary w32 L2 H8", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": true, "width": 32, "depth": 2, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 2}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-145", "name": "diffusionblocks_sat8 binary w32 L2 H16", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": true, "width": 32, "depth": 2, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 2}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-146", "name": "diffusionblocks_sat8 binary w32 L4 H1", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": true, "width": 32, "depth": 4, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 4}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-147", "name": "diffusionblocks_sat8 binary w32 L4 H2", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": true, "width": 32, "depth": 4, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 4}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-148", "name": "diffusionblocks_sat8 binary w32 L4 H4", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": true, "width": 32, "depth": 4, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 4}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-149", "name": "diffusionblocks_sat8 binary w32 L4 H8", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": true, "width": 32, "depth": 4, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 4}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-150", "name": "diffusionblocks_sat8 binary w32 L4 H16", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": true, "width": 32, "depth": 4, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 4}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-151", "name": "diffusionblocks_sat8 binary w48 L2 H1", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": true, "width": 48, "depth": 2, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 2}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-152", "name": "diffusionblocks_sat8 binary w48 L2 H2", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": true, "width": 48, "depth": 2, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 2}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-153", "name": "diffusionblocks_sat8 binary w48 L2 H4", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": true, "width": 48, "depth": 2, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 2}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-154", "name": "diffusionblocks_sat8 binary w48 L2 H8", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": true, "width": 48, "depth": 2, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 2}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-155", "name": "diffusionblocks_sat8 binary w48 L2 H16", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": true, "width": 48, "depth": 2, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 2}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-156", "name": "diffusionblocks_sat8 binary w48 L4 H1", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": true, "width": 48, "depth": 4, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 4}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-157", "name": "diffusionblocks_sat8 binary w48 L4 H2", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": true, "width": 48, "depth": 4, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 4}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-158", "name": "diffusionblocks_sat8 binary w48 L4 H4", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": true, "width": 48, "depth": 4, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 4}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-159", "name": "diffusionblocks_sat8 binary w48 L4 H8", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": true, "width": 48, "depth": 4, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 4}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-160", "name": "diffusionblocks_sat8 binary w48 L4 H16", "family": "diffusionblocks_sat8", "spec": {"family": "diffusionblocks_sat8", "binary_transformer_projections": true, "width": 48, "depth": 4, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 8, "training_blocks": 4}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-161", "name": "diffusionblocks_nat32 fp w32 L2 H1", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": false, "width": 32, "depth": 2, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 2}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-162", "name": "diffusionblocks_nat32 fp w32 L2 H2", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": false, "width": 32, "depth": 2, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 2}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-163", "name": "diffusionblocks_nat32 fp w32 L2 H4", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": false, "width": 32, "depth": 2, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 2}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-164", "name": "diffusionblocks_nat32 fp w32 L2 H8", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": false, "width": 32, "depth": 2, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 2}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-165", "name": "diffusionblocks_nat32 fp w32 L2 H16", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": false, "width": 32, "depth": 2, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 2}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-166", "name": "diffusionblocks_nat32 fp w32 L4 H1", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": false, "width": 32, "depth": 4, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 4}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-167", "name": "diffusionblocks_nat32 fp w32 L4 H2", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": false, "width": 32, "depth": 4, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 4}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-168", "name": "diffusionblocks_nat32 fp w32 L4 H4", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": false, "width": 32, "depth": 4, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 4}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-169", "name": "diffusionblocks_nat32 fp w32 L4 H8", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": false, "width": 32, "depth": 4, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 4}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-170", "name": "diffusionblocks_nat32 fp w32 L4 H16", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": false, "width": 32, "depth": 4, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 4}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-171", "name": "diffusionblocks_nat32 fp w48 L2 H1", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": false, "width": 48, "depth": 2, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 2}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-172", "name": "diffusionblocks_nat32 fp w48 L2 H2", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": false, "width": 48, "depth": 2, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 2}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-173", "name": "diffusionblocks_nat32 fp w48 L2 H4", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": false, "width": 48, "depth": 2, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 2}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-174", "name": "diffusionblocks_nat32 fp w48 L2 H8", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": false, "width": 48, "depth": 2, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 2}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-175", "name": "diffusionblocks_nat32 fp w48 L2 H16", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": false, "width": 48, "depth": 2, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 2}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-176", "name": "diffusionblocks_nat32 fp w48 L4 H1", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": false, "width": 48, "depth": 4, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 4}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-177", "name": "diffusionblocks_nat32 fp w48 L4 H2", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": false, "width": 48, "depth": 4, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 4}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-178", "name": "diffusionblocks_nat32 fp w48 L4 H4", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": false, "width": 48, "depth": 4, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 4}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-179", "name": "diffusionblocks_nat32 fp w48 L4 H8", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": false, "width": 48, "depth": 4, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 4}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-180", "name": "diffusionblocks_nat32 fp w48 L4 H16", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": false, "width": 48, "depth": 4, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 4}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-181", "name": "diffusionblocks_nat32 binary w32 L2 H1", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": true, "width": 32, "depth": 2, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 2}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-182", "name": "diffusionblocks_nat32 binary w32 L2 H2", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": true, "width": 32, "depth": 2, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 2}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-183", "name": "diffusionblocks_nat32 binary w32 L2 H4", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": true, "width": 32, "depth": 2, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 2}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-184", "name": "diffusionblocks_nat32 binary w32 L2 H8", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": true, "width": 32, "depth": 2, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 2}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-185", "name": "diffusionblocks_nat32 binary w32 L2 H16", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": true, "width": 32, "depth": 2, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 2}, "params": 28896, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-186", "name": "diffusionblocks_nat32 binary w32 L4 H1", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": true, "width": 32, "depth": 4, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 4}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-187", "name": "diffusionblocks_nat32 binary w32 L4 H2", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": true, "width": 32, "depth": 4, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 4}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-188", "name": "diffusionblocks_nat32 binary w32 L4 H4", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": true, "width": 32, "depth": 4, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 4}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-189", "name": "diffusionblocks_nat32 binary w32 L4 H8", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": true, "width": 32, "depth": 4, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 4}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-190", "name": "diffusionblocks_nat32 binary w32 L4 H16", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": true, "width": 32, "depth": 4, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 4}, "params": 49568, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-191", "name": "diffusionblocks_nat32 binary w48 L2 H1", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": true, "width": 48, "depth": 2, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 2}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-192", "name": "diffusionblocks_nat32 binary w48 L2 H2", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": true, "width": 48, "depth": 2, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 2}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-193", "name": "diffusionblocks_nat32 binary w48 L2 H4", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": true, "width": 48, "depth": 2, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 2}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-194", "name": "diffusionblocks_nat32 binary w48 L2 H8", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": true, "width": 48, "depth": 2, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 2}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-195", "name": "diffusionblocks_nat32 binary w48 L2 H16", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": true, "width": 48, "depth": 2, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 2}, "params": 58704, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-196", "name": "diffusionblocks_nat32 binary w48 L4 H1", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": true, "width": 48, "depth": 4, "attention_heads": 1, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 4}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-197", "name": "diffusionblocks_nat32 binary w48 L4 H2", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": true, "width": 48, "depth": 4, "attention_heads": 2, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 4}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-198", "name": "diffusionblocks_nat32 binary w48 L4 H4", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": true, "width": 48, "depth": 4, "attention_heads": 4, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 4}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-199", "name": "diffusionblocks_nat32 binary w48 L4 H8", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": true, "width": 48, "depth": 4, "attention_heads": 8, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 4}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}, {"kind": "grid200", "id": "ASI200-200", "name": "diffusionblocks_nat32 binary w48 L4 H16", "family": "diffusionblocks_nat32", "spec": {"family": "diffusionblocks_nat32", "binary_transformer_projections": true, "width": 48, "depth": 4, "attention_heads": 16, "ffn_expansion": 2, "generation_block_size": 32, "training_blocks": 4}, "params": 105072, "status": "not-run-under-objective", "novelty": 0, "novelty_reason": "A grid over width, depth, heads and binary flag across five fixed templates; no new mechanism."}], "winners": {"best": "markov_corpus_prior", "fastest": "nope"}, "status": {"started": 1794399604.7248912, "jobs_done": 38, "jobs_failed": 140, "pauses": 104, "paused_seconds": 30531.55717229843, "current": null, "state": "queue empty", "last_update": 1794569598.2358065, "host": "8dfbada1207f", "job_started": 1794522359.671632}, "upstream_tests": {"total": 105, "ok": 105, "causal_ok": 105, "failed": 0, "wall_s": 47.1, "torch": "2.9.1+cu128", "library": "flash-linear-attention 0.2.2 (real GPU kernels)", "pinned_commit": "3113c51b04a1d4a7520a3b0238c6a08d3194d58f"}, "counts": {"measured": 38, "queued_native": 0, "upstream": 105, "grid": 200}, "ref_tokens": 1228800}