results: close BabyAI population predictor rescue

1 files changed, 19 insertions, 0 deletions

diff --git a/BABYAI_SHARED_FEEDBACK.md b/BABYAI_SHARED_FEEDBACK.md
index 5749949..ea3bd99 100644
--- a/BABYAI_SHARED_FEEDBACK.md
+++ b/BABYAI_SHARED_FEEDBACK.md
@@ -236,3 +236,22 @@ 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.
+
+## P3 outcome: population-linear capacity is insufficient
+
+All three seeds fail the frozen capacity gate.  At initialization, the four
+layers have held-out `R^2` between roughly `0.972` and `0.998`, with residual
+context RMS ratios between `0.031` and `0.107`.  During clean task learning,
+the first layer loses identifiability.  At epoch 40 its held-out `R^2` is
+`0.765`, `0.774`, and `0.752` across seeds, while its residual ratios are
+`0.310`, `0.306`, and `0.327`.  The worst observed values are `R^2 = 0.7517`
+and residual ratio `0.3266`, outside both frozen limits.  Later layers remain
+more predictable, so the failure is localized but still violates the
+every-layer requirement.
+
+The population map is much stronger than the diagonal predictor, but is not a
+stable solution throughout learning.  P3 therefore closes the present BabyAI
+rescue without a local-delta implementation or another downstream endpoint.
+Any future attempt must change what the predictor observes or how early-layer
+state is represented; increasing neutral samples, changing ridge strength, or
+using task accuracy to select a predictor is outside this protocol.