Skip to content

Latest commit

 

History

History
58 lines (43 loc) · 5.1 KB

File metadata and controls

58 lines (43 loc) · 5.1 KB

Brainloop A/B Wiring Forensic Audit

Generated: 2026-06-11


Verdicts Table

Benchmark / Pair Identical% Verdict Key Evidence
gguf_baseline vs gguf_deadblock 98.2% (161/164) WIRED-CORRECTLY Different model files loaded per log: baseline=qwen2.5-3b-brainloop.gguf, deadblock=cerebellum-deadblock-python.gguf. 161/164 known-expected (eval script comment says ~161). 3 differ = models are distinct.
pytorch_baseline vs pytorch_conch 45.1% differ (74/164) WIRED-CORRECTLY Published 62.2 vs 56.7 scores confirmed in eval_base.log / eval_conch.log. 74/164 completions differ — active intervention. bench_humaneval.py loads checkpoints-fusion-13k/fused_refiners.pt explicitly.
qwen7b-baseline vs qwen7b-baseline-real 100.0% (164/164) MISWIRED / DEGENERATE Both files contain 164 completions of literally " pass" (8 bytes each). Both results files show pass@1=None, elapsed=0. These are broken runs with generation failure — the "baseline" was never actually generated. File hash identical.
brainloop-rag-coding vs brainloop-sharp-rag 97.6% (160/164) MISWIRED Both runs show identical scores (32.32/36.59 pass@1_plus). 4 completions differ only in whitespace/trivial variable rename. Elapsed times differ (280s vs 294s) suggesting two real runs, but with functionally identical models or identical RAG config — the intervention (sharp vs non-sharp RAG) was effectively inert.
brainloop-best-combo vs brainloop-fix-13k 100.0% (164/164) MISWIRED Files are byte-identical (same MD5). Both results files report identical scores and elapsed_seconds=134.9. One is a file copy masquerading as a separate run.
brainloop-both-active vs brainloop-combined 70.7% differ WIRED-CORRECTLY (sufficient divergence) 48 completions differ, distinct scores, not a copy.
brainloop-code-examples vs brainloop-code-trained2 17.7% identical WIRED-CORRECTLY 82% differ — clearly distinct models.
qwen7b-true-baseline vs qwen7b-rag 13.4% identical WIRED-CORRECTLY Strongly distinct.
qwen7b-true-baseline vs qwen7b-full-13k 13.4% identical WIRED-CORRECTLY Strongly distinct.
recall_results_deadblock.json A vs B 100.0% (200/200) MISWIRED model_a=qwen2.5-3b-brainloop.gguf, model_b=cerebellum-deadblock-python.gguf. Both models returned identical completions on all 200 prompts AND identical recall scores (hits=20/200). The recall bench cannot distinguish the two models — either the server ran the same model twice, or both models produce identical recall behavior on this task.

Checkpoint Forensics

checkpoints-fusion-13k/fused_refiners.pt (PyTorch conch bench):

  • l17.gate: tanh(-0.005) = -0.005 → NEAR-INERT (scales output by 0.5%)
  • l30.gate: tanh(-0.00485) = -0.00485 → NEAR-INERT
  • l17.rag_scale: sigmoid(0.0) = 0.5; l30.rag_scale: sigmoid(0.00412) ≈ 0.501
  • l17.inj_proj.weight: L2 norm(W - I) = 0.000 → identity, did NOT train
  • l30.inj_proj.weight: L2 norm(W - I) = 4.48 → did train
  • Conclusion: refiners were near-inert at bench time. The published -5.5% "logic tax" (62.2→56.7) is confounded — the score drop likely reflects the prompt-formatting difference between the base bench and the conch bench (instruct prompt wrapping), not the refiner intervention.

checkpoints-deadblock-13k/dead_blocks.pt (GGUF deadblock bench):

  • down_proj rows[:1536] L2 norm: 0.000 (l17 and l30) → correctly zeroed
  • down_proj rows[1536:] L2 norm: 1.39 / 1.60 → trained and non-zero
  • PARITY CHECK in train log: max_abs_diff = 0.0 PASSED confirms zero-initialization
  • Conclusion: dead block architecture is correctly implemented.

checkpoints-liveblock/live_block_best.pt:

  • All weight norms >> 0 (e.g., mlp.gate_proj norm=125, down_proj norm=18.5)
  • meta: epoch=1, val_ppl=8.76, timestamp=2026-06-11 04:13:08
  • Conclusion: trained, non-trivial weights.

Provenance Gaps

All results files in bench_results/ use the same schema with 7 fields: benchmark, model, pass_at_1_plus, pass_at_1_base, total_problems, elapsed_seconds, timestamp. No file records the checkpoint path, git commit, GGUF path, or model revision that produced it.

Unauditable results (no recoverable checkpoint/config provenance):

  • brainloop-code-examples, brainloop-code-trained2, brainloop-rag-trained, brainloop-combined, brainloop-both-active, brainloop-full-corpus — no log files found, no checkpoint ref in results JSON
  • qwen3b-refiner-arc, qwen3b-baseline-arc, qwen3b-refiner-hellaswag — no log files found

Most Damning Finding

The published PyTorch A/B comparison (62.2 vs 56.7 pass@1) is confounded by near-inert gates. At bench time, fused_refiners.pt had tanh(gate) ≈ -0.005 on both injected layers (L17, L30) — scaling the refiner's output contribution by only 0.5%. L17's inj_proj.weight was an exact identity matrix (norm(W-I)=0.0), meaning it never received gradient updates. The -5.5 point drop between baseline and conch likely reflects the instruct-prompt wrapping (added <|im_start|>user\n...) rather than the refiner injection itself. The intervention was geometrically present but functionally ~zero.