diff options
Diffstat (limited to 'BABYAI_SHARED_FEEDBACK.md')
| -rw-r--r-- | BABYAI_SHARED_FEEDBACK.md | 16 |
1 files changed, 16 insertions, 0 deletions
diff --git a/BABYAI_SHARED_FEEDBACK.md b/BABYAI_SHARED_FEEDBACK.md index aa24b84..5749949 100644 --- a/BABYAI_SHARED_FEEDBACK.md +++ b/BABYAI_SHARED_FEEDBACK.md @@ -220,3 +220,19 @@ neutral prediction screen that reads no action, teaching, raw, SDIL, or test outcome. If it cannot explain at least 80% of held-out mission-field variance and leave at most 25% context RMS in every layer, no further PickupLoc endpoint is run. + +### P3 population-predictor capacity screen + +P3 does not train raw or SDIL and does not evaluate rollout or test outcomes. +For clean-KP networks with the frozen P2 architecture and seeds `4101--4103`, +it audits epochs `0, 1, 5, 10, 20, 40`. At each checkpoint, a full layer-local +linear map is ridge-fitted from 4,096 neutral `(h_l, C_l m)` pairs and evaluated +on 4,096 disjoint neutral pairs. The ridge coefficient is `1e-3` times the +mean somatic variance. The fit receives no action, teaching field, loss, +downstream weight, raw result, or SDIL result. + +The capacity screen passes only if every seed, checkpoint, and layer has +held-out mean per-cell `R^2 >= 0.8` and residual context RMS at most `0.25` of +the original. This ridge fit is an expressivity upper bound, not the proposed +hardware learning rule. Passing opens a separate local-delta predictor check; +failure closes the population-predictor rescue. |