One ranked table. Every candidate is scored on what your binary, sparse, streaming system needs. 50 = your current block; higher means better.
Loading…
| # | Architecture | Fit | Quality | Streaming | KV/token | Long ctx | Stability | Novelty | Evidence |
|---|---|---|---|---|---|---|---|---|---|
| Loading the index… | |||||||||
The upstream ASI-Arch objective scored linear-attention kernels on benchmarks. Your cluster is a different problem, so the objective was rebuilt around it.
| Weight | Component | What it measures | The live problem it targets |
|---|---|---|---|
| 55% | Quality at equal box time | Held-out cross-entropy (AR 60, SAT 20, NAT 20) compared at the token count each candidate reaches in the same wall-clock as your block, using its real streaming bytes and measured step cost | 278 s/step dominated by latent streaming; 12.8M tokens seen at 600B; deep undertraining, so learning per step and per byte is what counts |
| 15% | Scale | Latent bytes at real 600B scale (from the candidate's actual module catalog) | Sub-bit storage on the path to 1T; disk and streaming bills |
| 10% | Stability | Local-denoise loss spike rate and level | The run halted on the local-denoise loss guard (26.2 vs limit 25) |
| 10% | Long context | Held-out NLL gain from 1,024 tokens of context vs 256 (trained at 256). Avoiding harm earns up to the gap to your block; only a real gain from the extra context earns more | RoPE fails past the trained length; NoPE hurts NAT |
| 10% | Serving | KV state bytes per generated token at 600B scale | 24.9 MB/token is the binding term on the 1660 |
Earlier versions of this page ranked the 105 upstream programs on the upstream goal and kept a separate 200-entry grid that was never scored. Those pages are kept here: old 105 index · old 200-grid gallery.