results: close stagewise no-KP feedback screen

1 files changed, 29 insertions, 0 deletions

diff --git a/ORAL_A_V6.md b/ORAL_A_V6.md
index 7663f27..6e597f4 100644
--- a/ORAL_A_V6.md
+++ b/ORAL_A_V6.md
@@ -94,3 +94,32 @@ recovery after the endpoint.  Failure closes V6 and the present no-KP causal
 vectorizer route.  Passing opens only a separately frozen short task gate;
 capture alone cannot raise the reviewer score.
 
+## Audited outcome (2026-08-06)
+
+V6-1 fails on clean source revision `98bfda7`.  The run executes the exact
+frozen 19 stages, 380 edge events, 760 batch-loss queries, and 48,640
+per-example observations.  All forward weights, readout weights, BatchNorm
+state/affines, and forward optimizer tensors remain bitwise fixed; the local
+fits use zero forward-weight reads and zero reverse-mode learning operations.
+
+The readout stage succeeds.  Its final feedback/forward parameter cosine is
+`0.81319`, and final-hidden teaching alignment is `0.87219`, compared with
+`0.00377` and `0.07787` under fixed HFA.  This confirms the post-failure V5
+diagnosis: per-example causal observations have enough signal to identify the
+small dense feedback map when fitted appropriately.
+
+The convolutional diagonal approximation fails.  Feedback norm ratios already
+reach `10.17--49.08` on finite late edges.  The predicted parent-field MSE
+grows from `0.0258` at edge 18 to `578` at edge 16, `3.12e7` at edge 15, and
+`6.84e30` at edge 9.  Edge 9 and all earlier stages become nonfinite.  The
+child contexts are strongly correlated, so treating their covariance as
+diagonal creates a multiplicative error cascade rather than a stable local
+inverse.
+
+Consequently every alignment and norm gate fails, no task-accuracy run is
+opened, and validation/test endpoints remain untouched.  Per the frozen rule,
+there is no ridge, clipping, or full-covariance recovery.  V6 closes the
+present task-loss-only no-KP causal vectorizer route.  Recovering convolutional
+feedback would require a materially different information source or an
+expensive full local system-identification method, which would need to beat
+the already measured KP, response-mirror, and BP cost/accuracy baselines.