results: close hierarchical task-scalar calibration gate

1 files changed, 22 insertions, 0 deletions

diff --git a/ORAL_A_V4.md b/ORAL_A_V4.md
index db1af1a..f2f34cf 100644
--- a/ORAL_A_V4.md
+++ b/ORAL_A_V4.md
@@ -92,3 +92,25 @@ test. It does not itself establish Harnett-specific novelty. Only after that
 innovation test passes may an independent ResNet-20/32/56 confirmation be
 specified; no test endpoint is authorized by V4-1 or V4-2.
 
+## Audited outcome (2026-07-22)
+
+V4-1 failed. The uncalibrated reference has early/all-layer alignment
+`-0.000263/0.011279`. At `eta_A=0.1`, 400 causal calibration events move these
+to only `-0.000180/0.017336`; feedback norms remain controlled. At `eta_A=1`,
+early/all alignment is `0.000348/0.007922`, but the maximum feedback/forward
+norm ratio explodes to `58.48`. The `eta_A=10` feedback state becomes
+nonfinite. Mean prediction/target cosine is at most `0.000668` across the
+finite learned candidates.
+
+Consequently the selected finite candidate fails all four substantive gates,
+and the four-record finite gate also fails. Each learned record used exactly
+800 logical batch-loss queries and estimated `1.2254e13` MACs before any task
+weight update. V4-2 and V4-3 are not launched, test remains untouched, and the
+reviewer score remains 5/10.
+
+The result falsifies this simultaneous global-task-scalar estimator at the
+frozen budget; it does not falsify the representational value of hierarchy.
+Further rate or query-budget tuning of V4 is prohibited. A next branch must
+change the information source, for example by using a locally observed child
+response as in learned FA/weight-mirroring controls, rather than estimating
+267,904 feedback parameters from the same global scalar projection.