Results

Historical campaign review: 2026-09-22; Limite endpoint review: 2026-10-07. Values below are historical results, not live run status.

Main findings

Experiment	Result	Interpretation
Original 100M/300M sweeps	GDN had strong BPB; Engram offered a favorable quality/throughput tradeoff	Scaling recipes gave larger variants more tokens; this is not a pure equal-token effect
10M backend comparison	Architecture-delta correlation 0.971361; 13/15 directions agree	Broad signal transfers, but the backends differ numerically
Qwen W320, 3.003B tokens	CE: global 3.663525; local 3.641465; simplicial 3.649627	Ordinary local attention won this pilot
Frozen-pretrained Qwen	CE 1.936017 at initialization to 1.862897 at step 8000	Learning occurred; no trained ordinary-adapter control isolates the mechanism
DeepSeek final matched full fine-tuning	CE: normal 0.700900; simplicial 0.701895	Small normal advantage; normal was also faster
Original DeepSeek scratch pair	CE: normal 4.122233; simplicial 4.160268 at 240.875M tokens	Early-stopped comparison, not a converged 10B result

Limite endpoint comparison

At 10B total supervised targets, AIME26 mean pass@1 is 15.83% normal and 20.83% 2-simplicial (30 problems, four draws each, native full context). Subsequent historical RL gives 12.50% and 20.83%. These results do not isolate architecture: parameter counts differ, only one training pair is available, and historical SFT omitted most document endings. See the complete SFT comparison.

Interpretation limits

• Completed comparisons do not establish a simplicial-specific capability advantage.
• The DeepSeek full comparison has one training pair. Its evaluation bootstrap does not measure variation across training seeds.
• Small capability subsets are regression screens, not proof of general equivalence.
• Corrected random-feature runs are separate experiments; the invalid sampler's curves remain excluded.
• Compare only compatible metrics, datasets, budgets and evaluation protocols.

Evidence map

The repository's docs/RESULTS.md indexes the frozen campaigns. docs/TRAINING_CONCLUSIONS_20260922.md audits Qwen/DeepSeek lineages and source artifacts. Compact evidence is versioned under docs/recorded-results/; detailed runs may require team storage.

Next: Qwen Experiments · DeepSeek Experiments