# BTC6H rejected-idea revisit protocol — September 21, 2026

Run the fragility diagnostic first, then the higher-low entry-gate replay. Both
are retrospective studies on the captured threshold-comparison data. Production
weights, parameters, promotion requirements, winner files and Pine stay unchanged.
No new training, optimization, threshold selection or deployment is authorized by
this protocol. Results do not constitute untouched validation.

## 1. Does score proximity predict execution sensitivity?

Use the completed 10,010-row threshold comparison as the sampling frame. For each
of its ten artifacts, sample up to ten configurations with `Fragile_0.05 > 0`
and ten with `Fragile_0.05 == 0`, without replacement, using RNG 20260921 and
lexically ordered artifact paths. Include the frozen incumbent and each model's
previously best-ranked configuration as separate case studies. Deduplicate exact
parameter configurations and retain sample/case-study membership. Do not select
on perturbation outcomes. Freeze the exact sample before evaluating any replay.

For every center, perturb each of the entry, exit and exit-confirmation thresholds
separately by -1, -0.05, +0.05 and +1 score units, changing nothing else: twelve
perturbations and one center. This is a diagnostic stress test, never a local
search. Do not use these rows as replacement candidates or relax a failed gate.

Compare each path with its center: changed executed entry/exit bars, position
bars, return intervals, net return, drawdown and Calmar, separately for IS and
the already-inspected OOS period. Retain all standard periods and gate outcomes.
At the center level, tabulate whether the original flag predicts any IS execution
change at magnitude 0.05 and separately at 1.0. Define a material performance
change in advance as at least 1 percentage point absolute return or drawdown
change, or 0.05 absolute calendar-Calmar change. Report actual magnitudes too.
Confusion tables use only the stratified sample, not extra best-ranked cases;
give per-model counts and denominators. They are descriptive within this sample,
not population precision estimates or independent statistical significance.

List each center's raw proximity events, threshold, confirmation-enabled state,
score smoothing windows, trading-window membership, and actual execution on that
bar and the next two bars. Raw proximity can be irrelevant to positions, disabled
conditions or smoothed crossings. Conversely, a zero raw flag need not imply
insensitivity. Preserve that distinction; do not invent a replacement gate.

## 2. Higher-low entry gates

Replay exactly four policies on the original frozen incumbent: unchanged control,
and the existing causal higher-low bullish-state entry gate at symmetric pivot
widths 2, 3 and 4. Keep weights, thresholds, stops and confirmation unchanged.
Keep the existing gate semantics: a confirmed higher low enables entries until
a confirmed lower high disables them. Do not add lower-high exits or retune.

Use the common corrected evaluator and complete promotion predicate, with distinct
policy evaluation identities and one shared experiment context. Prove an all-true
entry gate reproduces canonical actions, fees, returns and metrics. Verify prefix
causality. Retain all folds, especially those below the trade floor, and report
raw period measurements alongside eligibility. Do not zero carry-in OOS losses
merely because a window has no new entries. No new candidate advances automatically.

## Verification and stopping

Require parameter/Pine preflights. Match data, model hashes and accounting sources
to the prior comparison; abort input drift. Preserve experiment source hashes,
frozen configurations, all results, compact complete action/return traces and
proximity diagnostics. Independently reconstruct returns/drawdowns from traces
and verify reported sensitivity counts. Check all protected hashes after each
study. A failed run is retained as incomplete; use a new directory for any rerun.
Run focused tests, full fast tests, and review the implementation and findings.
Finish after the two specified studies; no adaptive follow-up search.