Scan config: compression→displacement event study (direction-free) — declared before running

Date: 2026-07-23 · Type: Edge Lab layer-3 scan (Research Program §4) — first scan executed under the program; seeds the Edge Lab pipeline (Helm #37). Trigger: Helm #39 — run 18's control arm grossed +0.0526R/403tr with random direction on these events, and Vantage-class drag (0.028–0.084R, VANTAGE-COST-UPDATE-v1.md) makes a +0.05R-class effect potentially harvestable.

Split: Research only (Binance BTCUSDT M5 from M1, 2017-08-17 → 2022-12-31). No fills, no stops, no costs — event study per §4 layer 3.

Event definition (frozen; scale-free restatement of expansion.py::detect_events)

M5 bars; range expressed in bp of close (rng_bp = (h−l)/c × 1e4) so every ratio is scale-free across BTC's 21× span. A bar is an event iff (all at bar close, rolling stats shifted one bar): - rng_bp ≥ disp_mult × baseline (baseline = 24h rolling median of rng_bp, ≥100 bars) - body ≥ 0.6 × range (one-sided bar) - prior-hour mean rng_bp ≤ baseline (compression precondition) - baseline ≥ 7 bp (the registered min-baseline; kills degenerate-baseline events) - direction = sign(c − o) ≠ 0 (recorded for the signed metric; no H4 filter — this is the direction-free study)

Condition variants (the hypothesis axis): disp_mult ∈ {3, 4, 5}.

Measurements per event, horizons H ∈ {5m, 15m, 30m, 1h, 2h, 4h, 8h}

Forward move = mid close(t+H) − close(t), in units of ATR14(M5) at the event (vol units), and in bp (for drag comparison): - M1 signed: move × event direction (continuation vs fade) - M2 absolute: |move| (direction-free volatility structure — what run 18's coin-flip arm was implicitly harvesting)

Null battery (§4 layer 4 — all four)

1. Unconditional: all-bar forward stats, same units.
2. Frequency-matched random: 200 reps of n random bars; null distribution of the mean; report where the observed mean falls (empirical p).
3. Label shuffle: direction labels shuffled across events (M1 only), 200 reps.
4. Block bootstrap: events resampled by ISO-week blocks, 1000 reps → 90% CI of the mean (clustering-honest). Never iid (§3).

Regime blocks: per calendar year (2017…2022) for sign-stability (≥3 of 4 multi-year blocks read, per §5.3).

Hypothesis accounting

3 disp variants × 2 metrics = 6 hypotheses logged against the 2,000 lifetime budget (cumulative after this scan: 16). Ledger row written to bot/runs/experiments.db (created by this scan — the §4 layer-5 ledger's first row).

Promotion context (pre-commitment, §5)

Nothing in this scan is a finding unless: effect ≥ 3× drag at a candidate stop distance (Vantage central: RT 3.1 bp ⇒ e.g. ≥9.3 bp excess at whatever geometry would harvest it), survives all four nulls, sign-stable ≥3/4 regime blocks, plateau across disp_mult, n ≥ 200, and replicates on a second instrument (XAUUSD data already on disk). Anything less gets logged and left.

Result (filled 2026-07-23, after the three runs; ledger rows 2–4; artifacts bot/reports/scan-displacement-events/)

M1 — signed continuation on BTC: the strongest scan-level effect the project has found. disp4 (n=517) and disp5 (n=282), horizons 15m–1h: mean continuation +0.29…+0.69 ATR (+10…+21 bp), label-shuffle p ≤ 0.015, block-bootstrap 90% CIs exclude zero, positive in 4/6 (disp4) and 5/6 (disp5) calendar years — including 2018 and 2022 bears; effect rises monotonically disp3→5 and decays after ~1–2h (plateau + sensible physics, not a single-setting spike). disp3 shows nothing (dilution — weak events are noise). Against Vantage-class RT cost 3.1 bp, a +19 bp/1h effect is prima facie ≥ 3× drag at any reasonable harvest geometry — the formal geometry test belongs to the promotion run.

M2 — direction-free vol elevation: confirmed and replicated (BTC all horizons rand-p ≈ 0; XAUUSD rand-p ≤ 0.015) — real but it is vol clustering, the known-physics part; context, not edge.

Replication (§5.6): NOT MET — underpowered, not refuted. XAUUSD 2024–26 yields only 36/22 events at disp4/5 (non-24/7 + 2.5y): signed means ≈ 0 with shuffle-p ≈ 0.9 — a null result with no power. GBPUSD: 0 events under the frozen 7 bp baseline floor (FX-vol class excluded by config — not comparable, not evidence). Per pre-commitment the condition is NOT promoted; no Validation look, no prereg yet.

The unlock is data, not thresholds: extend XAUUSD M1 history to ~2014 (Dukascopy serves deep gold history; ~10y ⇒ expected ~145 disp4 events — real power). Then rerun THIS config unchanged on the extended gold set. If gold confirms, the §5 checklist is complete and the promotion prereg (event-follow entry at close, ATR-scaled stop, 30m–1h horizon, Vantage + Dukascopy cost variants) gets written. If gold refutes with power, the effect is BTC-specific and gets logged as such.

Interpretation note for the mentor branch: run 18's directed spec failed while this scan's direction signal is strong — the difference is the pullback-confirmation entry: waiting for a 25% retrace + reclaim forfeits the 15m–1h continuation window. The mentor's journals teach the pullback; the data says the displacement close itself carries the information. That divergence is now measured, on the record, and belongs in the eventual formalization discussion (hybrid lineage, R5).

Hypothesis accounting: 18 logged across the three runs (BTC 6, XAUUSD 6, GBPUSD 6) — cumulative project total 28 of 2,000.

Replication rerun (2026-07-23 midday, deep gold): §5.6 REFUTED — the effect is BTC-specific

Data unlock delivered: HistData.com XAUUSD M1 2014-01 → 2023-12 (120 monthly files, bot/data/histdata/xauusd-1m/, ET-no-DST → UTC conversion documented; single-quote series — irrelevant to the scale-free event ratios). Config byte-identical. 699,193 M5 bars.

Read: this is not underpowered ambiguity — a BTC-sized effect (+0.4…+0.7 ATR) would be visible at n=61; instead the sign is consistently opposite. On gold, displacement bars mean-revert (disp3 marginally significant at p=0.065). The direction-free vol elevation (M2) still replicates at disp3/4 — vol clustering is universal; the directional continuation is not. It is a BTC-specific (plausibly crypto-specific) phenomenon.

Pre-committed consequence (§5.6): NOT PROMOTED. No promotion prereg, no Validation look (budget stays 10/10). The finding is logged as: displacement continuation exists on BTC across 5 regime-years with a full null battery behind it, and inverts on gold across a decade. Whether an instrument-specific effect should ever be promotable is a question about §5 itself — that belongs to the quarterly review (D4, 2026-10-22) as a possible amendment debated cold, NOT to a post-result exception here. Amending the bar after seeing a finding fail it is the exact failure mode the program §0 names.

Possible follow-ups (each a new hypothesis, none licensed here): replication on a second crypto instrument (ETHUSDT, same Binance source) would test "crypto-specific" vs "BTCUSD-artifact" without touching §5's letter — worth an Edge Lab config; the gold mean-reversion hint (disp3, p=0.065, wrong-side) is weak and stays unpursued.

Hypothesis accounting: +6 (deep-gold rerun) = 34 of 2,000.

ETHUSDT replication config (declared 2026-07-23, before running)

Purpose: split "crypto-market structure" from "BTCUSD artifact" after the gold refutation. Instrument: Binance spot ETHUSDT 1m (same source as BTC), window matched to the BTC research split 2017-08-17 → 2022-12-31 (ETHUSDT listed 2017-08). Config byte-identical: same event definition, thresholds (incl. the 7 bp baseline floor — ETH vol class is crypto-normal, so unlike GBPUSD this is a fair test), horizons, and full four-null battery. Hypotheses: +6 (3 disp variants × 2 metrics). Read declared in advance: replication within the crypto class supports escalating the §5.6 interpretation question to Richard/D4 with evidence; refutation kills the displacement-continuation finding entirely (BTC-only effects with n=1 instrument class stay logged, unpursued).

Result (run same day): REPLICATES, with near-identical magnitudes. 563,609 M5 bars (identical window). Signed 1h: disp3 +0.221 ATR (p=0.035, n=1374) · disp4 +0.393 ATR / +15.0 bp (p=0.055, boot90 [+0.09,+0.71], n=496) · disp5 +0.720 ATR / +19.4 bp (p=0.020, boot90 [+0.26,+1.20], n=238). Per-year: disp4 4/6 positive, disp5 5/6 (only 2020 negative). BTC disp5 was +0.670/+19.2 bp — the magnitudes match to within noise. Verdict: the displacement-continuation effect is crypto-class structure, not a BTCUSD artifact; the universal-structure claim stays dead (gold inverts). The §5.6 interpretation question (within-class replication for a crypto-only candidate) goes to Richard/D4 with this evidence. Hypothesis accounting: +6 = 40 of 2,000.