# Experiment log

Generated by `py -3.11 -m neokernel report` from [log.jsonl](../neokernel/results/log.jsonl). Notes are historical records; later outcomes do not rewrite earlier failures.

| ID | Proposer | Item | Outcome | Local geomean tok/s | GPU seconds | Workloads | Note |
| --- | --- | --- | --- | --- | --- | --- | --- |
| 1 | human | manual | failed: latency_limit | not recorded | 61.767 | public-0, public-1, public-2 | manual benchmark; no keep decision |
| 2 | human | manual | measured, not kept | 157.866 | 104.853 | public-0, public-1, public-2 | manual benchmark; no keep decision |
| 3 | human | manual | measured, not kept | 144.597 | 101.937 | public-0, public-1, public-2 | manual benchmark; no keep decision |
| 4 | agent | baseline | kept | 216.318 | 199.618 | public-0, public-1, public-2, h-a, h-b, h-c | measured current tree before proposals |
| 5 | agent | static_kv_cache | patch rejected | not recorded | 0.000 | none | patch_failed: error: corrupt patch at line 29 |
| 6 | agent | baseline | kept | 251.336 | 171.058 | public-0, public-1, public-2, h-a, h-b, h-c | measured current tree before proposals |
| 7 | agent | static_kv_cache | failed: candidate_error | not recorded | 46.424 | public-0, public-1, public-2 | correctness check failed |
| 8 | agent | static_kv_cache | guard rejected | not recorded | 0.000 | none | engine.py:43: closing parenthesis ')' does not match opening parenthesis '[' |
| 9 | agent | static_kv_cache | failed: candidate_error | not recorded | 44.867 | public-0, public-1, public-2 | correctness check failed |
| 10 | codex | cuda_graph_decode | failed: latency_limit | not recorded | 168.953 | public-0, public-1, public-2, h-a, h-b, h-c | L4 public correctness passed; H100 rejected on TTFT latency gates. Public-0 TPOT misses 5 ms target. Prior engine restored; dependent auto run not started. |
| 11 | codex | cuda_graph_decode | kept | 323.657 | 170.733 | public-0, public-1, public-2, h-a, h-b, h-c | All six local gates pass. Kept graph foundation; strict TPOT 5/7 ms and TTFT <=1.0 targets remain unmet. One graph launch and one host synchronization verified. |
| 12 | agent | baseline | failed: latency_limit | not recorded | 194.013 | public-0, public-1, public-2, h-a, h-b, h-c | baseline failed gates |
| 13 | codex | hand_rolled_decode | rejected / incomplete | not recorded | 15.729 | none | Pre-integration L4 unit validation: four GPU tests passed; full layer switch test stopped on noncontiguous grouped SDPA view. RMSNorm/QK-RoPE/SiLU exact; attention max abs diff 0.00390625. SDPA destination-view fix passes CPU regression; GPU retry pending attended approval. Active engine unchanged. Estimated task spend USD 0.012930. |
| 14 | codex | hand_rolled_decode | validation / not kept | not recorded | 16.636 | none | Second pre-integration L4 validation passed all six tests including 96 layer/config cases. Maximum differences: RMSNorm/QK-RoPE/SiLU 0, attention 0.00390625, layer 0.03125. Integrated validated kernels and packed shared weights; CPU prefill/storage regressions and source guard pass. Full-model L4 integration/check and H100 measurements pending approval. Not a kept benchmark. Estimated task spend USD 0.048254. |
| 15 | codex | hand_rolled_decode | validation / not kept | not recorded | 251.417 | public-0, public-1, public-2 | Integrated L4 tests and unchanged public correctness check passed. 48 prefill/decode handoff checks across both attention/cache layouts have zero native margin; max logits diff 0.1875; max load+warm 20.191 s. Public near ties 1/0/6. L4 correctness only, not a performance keep. H100 measurements and freeze pending. |
| 16 | codex | hand_rolled_decode | measured, not kept | 483.481 | 95.038 | public-0, public-1, public-2 | H100 triton; selection, final freeze, and measured target review still required before keep/commit. |
| 17 | codex | hand_rolled_decode | measured, not kept | 438.243 | 81.059 | public-0, public-1, public-2 | H100 sdpa_grouped; selection, final freeze, and measured target review still required before keep/commit. |
| 18 | codex | hand_rolled_decode | failed: latency_limit | not recorded | 153.937 | public-0, public-1, public-2, h-a, h-b, h-c | H100 full_bench; selection, final freeze, and measured target review still required before keep/commit. |
| 19 | codex | hand_rolled_decode | measured, not kept | 3476.524 | 49.024 | h-c | H100 bench_h-c; selection, final freeze, and measured target review still required before keep/commit. |
| 20 | codex | hand_rolled_decode | failed: latency_limit | not recorded | 233.997 | public-0, public-1, public-2, h-a, h-b, h-c | Final five-sample all-workload freeze failed h-c TTFT: ratio 1.0513944201 exceeds unchanged local 1.05 gate. All correctness checks passed. Public-0/2 TPOT 4.951553/6.482385 ms; profiles 439/475 kernels. No freeze artifact accepted, no commit, no push. Passing five-sample isolated h-c rerun retained separately. |
| 21 | codex | hand_rolled_decode | measured, not kept | 191.790 | 30.854 | public-0 | H100 bench_public-0; selection, final freeze, and measured target review still required before keep/commit. |
| 22 | codex | hand_rolled_decode | measured, not kept | 175.868 | 39.150 | public-0 | H100 bench_public-0; selection, final freeze, and measured target review still required before keep/commit. |
| 23 | codex | hand_rolled_decode | failed: incorrect_output | not recorded | 81.996 | public-0, public-1, public-2 | Rejected fused prefill norms: fresh L4 public-2 replay failed at [0,14,40], margin 2.125 > 2.0. Native norms restored; proceeding with native-operation prefill CUDA graph. |
| 24 | codex | hand_rolled_decode | measured, not kept | 190.579 | 28.499 | public-0 | H100 bench_public-0; selection, final freeze, and measured target review still required before keep/commit. |
| 25 | codex | hand_rolled_decode | kept | 722.578 | 180.448 | public-0, public-1, public-2, h-a, h-b, h-c | Five-sample all-workload freeze passed at unchanged 1.05 latency gate; user requested pause before further tests. Website runs left to user. |
| 26 | agent | lm_head_argmax | failed: candidate_error | not recorded | 56.824 | public-0, public-1, public-2 | correctness check failed |
| 27 | agent | attention_impl_kv_layout | interrupted / recovered | not recorded | 0.000 | none | interrupted; recovered without merge |
| 28 | agent | lm_head_argmax | failed: candidate_error | not recorded | 62.369 | public-0, public-1, public-2 | correctness check failed |
| 29 | agent | attention_impl_kv_layout | rejected / incomplete | 693.083 | 278.588 | public-0, public-1, public-2, h-a, h-b, h-c | gates failed or improvement did not exceed 1 percent |
| 30 | agent | lm_head_argmax | failed: candidate_error | not recorded | 61.588 | public-0, public-1, public-2 | correctness check failed |
| 31 | agent | lm_head_argmax | failed: candidate_error | not recorded | 44.970 | public-0, public-1, public-2 | correctness check failed |
| 32 | agent | lm_head_argmax | failed: candidate_error | not recorded | 44.792 | public-0, public-1, public-2 | correctness check failed |
| 33 | agent | prefill_packed_weights | failed: candidate_error | not recorded | 31.017 | public-0, public-1, public-2 | correctness check failed |
| 34 | agent | prefill_packed_weights | failed: candidate_error | not recorded | 31.942 | public-0, public-1, public-2 | correctness check failed |
| 35 | agent | prefill_cuda_graph | rejected / incomplete | 733.436 | 238.678 | public-0, public-1, public-2, h-a, h-b, h-c | gates failed or improvement did not exceed 1 percent |
| 36 | agent | attention_impl_kv_layout | interrupted / recovered | not recorded | 0.000 | none | interrupted; recovered without merge |