diff options
Diffstat (limited to 'SHARED_FEEDBACK.md')
| -rw-r--r-- | SHARED_FEEDBACK.md | 18 |
1 files changed, 18 insertions, 0 deletions
diff --git a/SHARED_FEEDBACK.md b/SHARED_FEEDBACK.md index a7d34c2..8808dca 100644 --- a/SHARED_FEEDBACK.md +++ b/SHARED_FEEDBACK.md @@ -114,3 +114,21 @@ innovation, context lesions, and an explicit accounting of the extra clean-KP instruction wire. Injected-bias or clean-KP scaling results cannot substitute for those checks. +## S0 outcome + +The sole frozen run from clean source revision `b345784` fails the gate. The +architecture and identification checks worked: oracle reached `99.316%` +validation accuracy, removing context reduced it by `22.998` points, the +innovation predictor explained `95.06%` of per-cell context variance on +average, its largest residual-context RMS ratio was `0.01553`, and it observed +zero task instructions. But the endogenous shared field did not obstruct +learning. Raw shared KP reached `99.023%`, innovation reached `99.219%`, and +norm-matched raw reached `99.170%`. The raw--oracle gap was only `0.293` +points and the innovation--raw gain only `0.195` points, far below the frozen +five-point requirements. + +This rules out the proposed sufficient condition: task-required ordinary +apical activity can be large, necessary for inference, and locally predictable +without producing a meaningful raw-learning failure. This realization is +closed without a context-scale, task, optimizer, width, depth, or predictor +screen, and it does not open a scaling experiment. |