# Execution sensitivity and frozen observation contract

Implement the September 24 recommendations without changing any installed model,
winner parameters or Pine presets. This is a validation change, not a deployment.

## Execution sensitivity v1

Raw score proximity remains a diagnostic; it no longer acts as a numerical veto.
Before eligibility for promotion, replay each otherwise eligible configuration at
entry, exit and exit-confirmation thresholds shifted independently by -1, -0.05,
+0.05 and +1 score units. Apply the same twelve perturbations to the incumbent.
Keep every other parameter, data input, fee and execution rule unchanged.

The guard fails if any perturbation loses more than 1% of terminal wealth relative
to its center, or worsens close-equity drawdown by more than one percentage point,
in IS, either IS subperiod, or any historical WFO period (including low-trade ones).
Improvements do not count as harm. Record all action differences and performance
deltas, not just the pass/fail result. These are fixed engineering tolerances for
small parameter changes, not empirically established probabilities of robustness.
Do not tune them to admit the hybrid or any other inspected candidate.

Missing, failed, stale or configuration-mismatched sensitivity evidence cannot pass.
Candidates that already fail drawdown/trade, subperiod or WFO requirements can skip
the expensive perturbations, with an explicit unassessed status; that status cannot
qualify. Research validation may request the full test for ineligible candidates.
The existing drawdown/trade, positive subperiod, WFO, incumbent metric and conditional
OOS protection requirements remain. No search or promotion is part of this work.

## Frozen forward observation

Freeze the original hybrid9090 near miss, the original incumbent with the three-bar
higher-low entry gate, and the unchanged incumbent control. Copy model bytes and
parameters into a versioned local bundle with hashes. Preserve the original hybrid
threshold -730, not the shifted diagnostic setting. Freeze the execution/sensitivity
code identity and an observed UTC cutoff. Only closed bars strictly after that
cutoff may count as forward evidence. Never backdate the freeze.

Provide a repeatable observation command taking a new chart export. It must verify
the bundle, reject source/parameter drift, validate required features and timestamps,
exclude the unconfirmed last candle, and report gaps/history revisions explicitly.
Prior captured input rows may not silently change. Start all policies flat at the
forward boundary, warm indicators on the same preceding history, retain subsequent
positions continuously, and mark open holdings to market. Keep all policies in every
report; never rank or retune from these observations. Record input hashes, exact
bounds, trades/exposure, return/drawdown, and append-only observation receipts.

No available post-freeze data means `awaiting_data`, never a zero-return success.
Until at least ten completed trades and 180 calendar days per candidate, report
`insufficient_coverage`; reaching coverage permits review only, never automatic
promotion. Paired TradingView chart/trade export parity is a separate prerequisite
to any eventual live change. New historical exports alone cannot establish
publication-time feature availability or live parity.

## Completion evidence

Tests must cover raw-proximity counterexamples, harmful versus harmless action
changes, each unchanged promotion hurdle, stale/missing evidence, disabled gate
equivalence and prefix causality, frozen-artifact tampering, future/unclosed bars,
changed historical rows, repeated observations and empty forward windows. Run
parameter/Pine preflights, focused and full fast tests, a real fixed-candidate
validation, freeze the actual bundle, and exercise the observer against currently
available data. Preserve the no-data status honestly and document collection steps.