diff options
| author | YurenHao0426 <Blackhao0426@gmail.com> | 2026-07-22 14:30:04 -0500 |
|---|---|---|
| committer | YurenHao0426 <Blackhao0426@gmail.com> | 2026-07-22 14:33:51 -0500 |
| commit | 39d1839ead0709798c2eb07104c3764ccf1ed3ae (patch) | |
| tree | d8813f14d590d3b9ec4368bcac4ee78584957a9d /KP_BASELINE.md | |
| parent | 41048d9d9c399ded6f5c473b578f59e6ac375ae9 (diff) | |
results: pass reciprocal KP short gate
Diffstat (limited to 'KP_BASELINE.md')
| -rw-r--r-- | KP_BASELINE.md | 21 |
1 files changed, 21 insertions, 0 deletions
diff --git a/KP_BASELINE.md b/KP_BASELINE.md index 25696f7..78bb756 100644 --- a/KP_BASELINE.md +++ b/KP_BASELINE.md @@ -76,3 +76,24 @@ the reviewer score. A KP-2 pass only opens a separately frozen raw-versus-norm-matched-raw-versus-innovation mixed-traffic experiment. The somato-dendritic innovation must be load-bearing there to affect the SDIL paper assessment. + +## Audited KP-1 outcome + +The sole frozen record from source revision `8f420f2` passes every gate: + +| metric | KP-1 | frozen requirement/reference | +|:--|--:|--:| +| epoch-20 validation accuracy | **82.66%** | at least 70%; BP epoch 20 is 81.02% | +| early-third teaching alignment | 0.885583 | at least 0.50 | +| final feedback/forward cosine | 0.901034 | at least 0.80 | +| epoch-11--20 mean feedback cosine | 0.854594 | at least 0.70 | +| MAC ratio to matched BP | 1.3261x | at most 1.40x | +| task-loss queries for feedback | 0 | exactly 0 | + +All epoch losses and tracking diagnostics are finite. The run uses +`1.451874816e14` MACs, 2.163 GB peak allocated memory, and 558.1 s on a GTX +1080. Its final mean parameter relative error is still 0.445, so KP does not +require exact weight equality to obtain a strong teaching direction. Unlike +the failed intermittent-mirror trajectory, these alignment gates cover the +constant-LR training interval rather than only a decayed-LR endpoint. KP-2 is +opened without changing any optimizer or feedback setting. |