lin-0195
15.8 The retained experimental sequence
The experiments were designed as a sequence in which failures determine the next structural layer. In the table, regret measures the value gap induced by the selected action, and ranking accuracy measures whether action contrasts have the correct order. A harmful admission is an update whose independently computed exact utility is negative. Exact utility is used only to score the experiment; the controller sees probe, confidence, phase, ESS, and temporal-state information. Post-shift AUC integrates return over the evaluation epochs following the hidden regime change.
|
Stage |
Retained result |
Failure isolated |
Layer added |
|---|---|---|---|
|
Value-level IL–GTD–MP |
tangent RMSE \(-24.1\% \); OOD Bellman RMSE \(-81.8\% \) |
does not imply policy improvement |
action-conditioned test |
|
Full-state policy bridge |
regret gain only \(0.073\); smooth wrong control passes |
absolute values hide rankings |
advantage quotient |
|
Advantage quotient |
regret \(3.465\to 2.081\); ranking \(+12.6\) points |
orthogonal phase failures |
coherence gate |
|
Sampled transport |
zero harmful admissions; \(95.0\% \) high-budget retention |
uncertainty causes abstention |
confidence block |
|
Off-policy transport |
zero harm but registered \(74.84\% {\lt}75\% \) threshold |
estimator depends on mismatch |
ESS meta-gate |
|
Adversarial online shift |
unconditional harm \(22.67\% \); router zero harm |
freezing delays recovery |
temporal recovery |