Scan declaration: failed-break reversion (edge #2 hunt, family 1)

Date declared: 2026-07-27 (written BEFORE implementation; run follows same night). Layer: 3 (no execution, no costs). Split: BTCUSDT research 2017-08-17→2022-12-31. Lineage: our own — successor to the level-reaction scan's logged-not-promoted observation (ledger 8: break-continuation mildly INVERTED, ≈−0.13 ATR/4h), the strongest unpursued lead in the ledger. Osler's stop-cluster literature motivates the mechanism (breaks that fail strand breakout entrants; their unwind is the reversion). Charge: 10 hypotheses at registration.

Question

After a level break that FAILS (price re-enters below/above the broken level within 30 minutes), does price revert in the failure direction with harvestable magnitude?

Definitions (frozen)

• Level book, break events: exactly the level-reaction scan's (full R7 book, break_up_event/break_down_event in data/features/btcusd-levels-m5.parquet; broken level price L = the pre-break res_level_price/sup_level_price).
• Failed break (up): break_up at bar t; failure = first M5 close < L within bars (t, t+6]. Event time = that re-entry close; direction = short. Mirrored for failed break (down) → long. One event per break; debounce: a break bar inside a prior event's horizon window is skipped (no overlapping positions in the same family — same convention as prior event scans).
• Re-entry speed dose: FAST = failure within ≤3 bars, SLOW = 4–6 bars.

Hypothesis cells (10)

• Signed reversion return (failure-direction-aligned, close-to-close) at {1h, 4h, 24h} × {failed-up, failed-down} = 6 cells
• Dose: {FAST, SLOW} × {failed-up, failed-down} at 4h = 4 cells (mechanism predicts FAST > SLOW: a sharper trap, a harder unwind)

Metrics in ATR14(M5) and bp. Four-null battery; label-shuffle degeneracy for these one-directional conditions noted in advance (carried by frequency-matched random + placebo + bootstrap, as in ledger 8/10/12). Per-year sign blocks all cells.

Pre-declared interpretation

• Live lead iff some direction shows CI-clear reversion at ≥2 horizons with ≥4/6 years sign-stable AND the dose ordering (FAST ≥ SLOW) holds in sign.
• A follow-on prereg is licensed only after the TEMPLATE ex-ante cost screen on the measured event ATR% (≥3× Vantage drag at a stop-invariant harvest) — same gate that refused three registrations before this one.
• Null/weak ⇒ the ledger-8 inversion gets recorded as artifact-or-tiny and the failed-break family retires; the edge #2 hunt moves to the next declared family (volume-conditioned displacement, then BTC→alt lead-lag — Helm #53/#54).

What this scan cannot say

Nothing about Validation-era regimes, nothing about execution truncation (E4: scans license tests, they do not predict results), nothing about non-BTC instruments.

RESULT (run 2026-07-27, same night) — NULL, FAMILY RETIRED

Ledger row 26 (+10 hypotheses → 261/2000); artifacts bot/reports/failed-break-scan/results-btcusd.json; detector src/tradingbot/edgelab/failed_break_events.py (causality-tested); commit 04abf68; suite 274 green.

All 10 cells null: every bootstrap 90% CI straddles zero (n=1,777 failed-up / 1,784 failed-down — well-powered); no random-p or placebo-p below 0.32; best cell by magnitude is failed-down@4h at −6.8 bp — which is weak continuation, the wrong sign for the reversion hypothesis, and fails the 9 bp ex-ante cost gate regardless. The FAST≥SLOW dose ordering technically holds but gates nothing without a CI-clear signal. Label-shuffle degenerate as pre-declared (one-directional families; on record in the JSON).

Reading: the ledger-8 "inverted break-continuation" lead was noise-level weak continuation, not a harvestable trap-unwind. Osler stop-cluster mechanics do not manifest at R7-level breaks on BTC M5 at measurable size. Failed-break family retired; edge #2 hunt proceeds to volume-conditioned displacement (Helm #53), then BTC→alt lead-lag (#54). Implementation note of record: debounce resolved to the max horizon (288 bars, per-family) — documented in the module.