Declared scan: streak-effect regime conditioning (the E2 question, asked of the live candidate)

Date declared: 2026-07-25, before any computation. Split: Research only (BTCUSD and ETHUSDT — the two instruments where the effect is decisive; conditioning is only meaningful where the base effect exists). Layer 3, Edge Lab, one ledger row per instrument. Evidence basis: Findings E2 (regime non-stationarity killed displacement; run 21's positive 2021–22 blocks were the pro-look argument) and the universality sweep (the effect inverts across instruments — does it also concentrate across regimes?). Purpose: decision input for look 2 — is up-streak persistence regime-flat (candidate's case recovers) or regime-concentrated (a conditioned successor is the better Validation object)?

Conditions (depth-2 under the combination rule; constants frozen here)

Base event: up_streak == 8 (frozen). Regime variable: volatility regime = trailing 30-day realized-vol percentile rank within each instrument's own research split (computed from the feature store's 288-bar realized vol, daily-resampled, rank-transformed; declared, not tuned). Bands: low < 0.33 ≤ mid < 0.67 ≤ high.

1. up_ge8 ∧ vol_low (+1) — 4 horizons (15/30/60/240)
2. up_ge8 ∧ vol_mid (+1) — 4 horizons
3. up_ge8 ∧ vol_high (+1) — 4 horizons

12 hypotheses × 2 instruments = 24 (cumulative 199 → 223 of 2,000).

Combination-rule reading (declared): a band "concentrates" the effect only if its 4h signed mean exceeds the unconditioned up_ge8 record (BTC +0.603, ETH +0.536) by more than bootstrap-CI overlap AND the complementary bands are materially weaker. Bands partition the event set — report the n split (the three bands must sum to the unconditioned n) so a "concentration" that is really n-starvation is visible.

Pre-declared reading

Priors, stated before running: the front-loading of magnitude in 2017–18 (both instruments) suggests high-vol concentration — 2017/2020 were high-vol regimes. If confirmed, the honest conclusion is uncomfortable for the candidate: 2023–2025 (Validation) is a lower-vol regime on BTC, so an unconditioned Validation run would be betting on the band where the effect is weakest. If the effect is regime-flat, the 2021–22 positives stand as genuine regime robustness and the original candidate's look-2 case recovers toward coin-flip. Either result changes the memo; that is the point. A conditioned successor spec, if ever built, is a NEW registration and must re-clear the ex-ante cost screen on its own band's event-time ATR%.

Results (filled 2026-07-25, two ledger-logged runs — btc then eth order; 24 hypotheses,

cumulative 199→211→223/2000)

Configs: engine/scans/configs/streak-regime-{btcusd,ethusdt}.json (base event up_streak == 8, freshness == as always, +1 direction shared across all three bands). Runner: scripts/run_streak_regime_scan.py — ONE shared scan.run_scan() call per instrument (all three bands carry the same declared sign, so unlike the rhythm/upstreak/universality streak family's per-variant loop, no per-variant split is needed here); engine defaults (1000/1000/1000 nulls, min_events 200), seed 20260725. Full tables in engine/scans/results/streak-regime-{btcusd,ethusdt}.json. 12/12 dry-run hypothesis counts confirmed on both instruments before the logged run. Both loaded spans 2017-08-17→2022-12-31 (research split, 563,609 M5 bars each). 196 tests green throughout (no source changes — config + script additions only, features.py/scan.py untouched).

Vol-regime variable: derivation and the live-causality caveat

vol_pctile_30d is a runner-side derivation (scripts/run_streak_regime_scan.py: add_vol_regime), not a persisted feature-store column — chosen over a new additive feature-store column because the measure is, by the declaration's own design, non-causal (an in-split rank), unlike the store's existing additive columns (close_hour_utc, move15_atr, etc.), which are all live-computable as-is; baking a research-only rank into the shared store risked a future scan reusing it as if it were live-safe. features.py is untouched, FEATURE_VERSION is not bumped (see the script's docstring for the full rationale).

Formula: (1) rvol_288 (the store's existing 288-M5-bar = 24h realized-vol column, bp) daily-resampled via .resample("1D").mean(); (2) a 30-day trailing measure via .rolling(30, min_periods=1).mean() on that daily series (min_periods=1 rather than requiring a full 30-day window — see the n-split check below, which confirms this choice costs zero events on either instrument); (3) .shift(1) to lag the daily series by one day before broadcasting, so every bar of day D sees the trailing measure through the END of day D−1, not day D's own incomplete realized vol; (4) .rank(pct=True) taken once, on the daily series, over the whole research split — this is the acknowledged non-causal step; (5) broadcast to M5 by calendar day.

Caveat, stated explicitly per the declaration's own instruction: this in-split percentile rank is fine for a research conditioning study — it answers "does this band, defined with full-split hindsight, show a different effect?" — but it is not how a live system could compute a vol regime. A live version needs a strictly causal rolling rank (e.g. a rank against only the trailing N days/years of history available at each timestamp, not the whole split), which would shift band boundaries over time and is a materially different (and harder) engineering problem than what this scan measures. Nothing here should be read as a live-deployable regime filter.

N-split check

Bands partition the up_ge8 event set exactly on both instruments (zero events lost to the vol-regime warmup window, confirming the min_periods=1 choice was costless):

Per-band × horizon tables

BTCUSD (unconditioned up_ge8 record: +0.603 ATR @ 4h, n=498, rhythm-btcusd.json)

ETHUSDT (unconditioned up_ge8 record: +0.536 ATR @ 4h, n=625, upstreak-eth.json)

(U) = flagged underpowered (n < min_events 200). 5 of 6 bands are underpowered on this partition (only ETH-mid and ETH-high clear 200) — the 3-way split roughly thirds an already-thin ~500-600-event base, which is the central fact the mechanical read below has to contend with.

Mechanical verdict per instrument (declared rule, applied literally)

Rule: a band "concentrates" the effect only if its 4h signed mean exceeds the unconditioned record by more than bootstrap-CI overlap (operationalized as: the record value must fall outside, below, the band's own 90% CI) AND the complementary bands are materially weaker.

• BTC — no band clears the CI-overlap leg. Record +0.603 falls inside all three bands' 4h CIs: low [−0.473, 1.347], mid [−0.053, 1.380], high [0.022, 1.361]. High has the largest point estimate (+0.741) and is the closest candidate, but its own bootstrap distribution comfortably contains the unconditioned record, so "exceeds beyond CI overlap" fails. The complementary-bands-weaker leg is also equivocal for high: mid (+0.663) is nearly as large as high (+0.741), not "materially weaker." Mechanical verdict: NOT CONFIRMED. Directionally, though, BTC's three point estimates rank monotonically with vol (low +0.382 < mid +0.663 < high +0.741) — suggestive, not CI-clean, weak support for the high-vol-concentration prior.
• ETH — no band clears the CI-overlap leg either. Record +0.536 falls inside low's CI [0.126, 2.170] and high's CI [−0.022, 1.366]; mid's CI [−0.806, 0.524] sits just below the record (upper bound 0.524 vs record 0.536) but mid's point estimate is negative (−0.163) — the record exceeding a weak band from above is not the declared criterion (which requires a band to exceed the record). Low has the standout point estimate (+1.120, ~2× the record) and the complementary bands are genuinely weaker (mid −0.163, high +0.693) — the second leg is satisfied — but low's own 90% CI still contains the record, so the first leg fails. Mechanical verdict: NOT CONFIRMED, though closer to qualifying than BTC's best candidate. Directionally, ETH's ranking is non-monotone and low-vol-favoring (low +1.120 > high +0.693 > mid −0.163), the opposite shape from BTC's own monotone-with-vol pattern.
• Cross-instrument reading: neither instrument produces a CI-clean concentration under the declared test — at ~150-215 events per band, the 90% bootstrap CIs at 4h are wide enough that no band is statistically distinguishable from the unconditioned pooled record. This is itself the honest finding: the scan is underpowered to resolve the E2 question either way at this partition's event counts, not evidence for regime-flat either. What can be said: the two instruments' point-estimate shapes disagree with each other (BTC monotone-increasing-with-vol; ETH non-monotone, low-vol-favoring), which argues against a single shared "vol regime" story generalizing cleanly across BTC and ETH, echoing the project's now-familiar instrument-specificity pattern (universality sweep, gold-inversion finding) rather than resolving it.

Band-by-year event-count concentration

(n_events_by_year, additive field computed by the runner post-scan.run_scan, not part of scan.py's own output)

BTCUSD

ETHUSDT

Reading, against the pre-declared prior ("front-loading of magnitude in 2017–18... suggests high-vol concentration — 2017/2020 were high-vol regimes"):

• 2017 confirmed as purely high-vol on this measure, both instruments — 100% of 2017 up_ge8 events fall in the high band (BTC 55, ETH 41; zero in low or mid on either instrument). 2018 also skews high-band-heavy on both (BTC 39 of 95 2018 events in high; ETH 75 of 151, the single largest band-year cell in the whole table). The 2017–18 half of the prior holds up cleanly.
• 2020 does NOT confirm as high-vol — it inverts the prior, on both instruments. 2020 is the low/mid-band-heaviest year, not high: BTC 2020 events are 59 low / 26 mid / 8 high; ETH 2020 events are 60 low / 48 mid / 9 high. Under this specific 30-day trailing realized-vol measure, up-streak-8 events in 2020 concentrated during low-vol windows, not the high-vol regime the prior assumed (likely reflecting that most of 2020's up_ge8 events — sustained multi-bar continuation runs — occurred during the low-vol grinding recovery/rally (post-March-crash through Q4) rather than the single sharp high-vol crash week itself, which a 30-day trailing average dilutes). This is a genuine, reportable divergence from the stated prior, not a confirmation of it.
• 2021 and 2022 sit mostly in mid/high (both instruments), 2019 sits mostly in low/mid — consistent with 2021 (bull-to-bear cycle) carrying more sustained realized vol than 2019 or 2022's later grind.

Ambiguity resolutions (explicit)

• Ledger cumulative count matches the declaration's own assumption exactly (ledger.cumulative_count() returned 199 immediately before this scan; two runs of 12 hypotheses each landed at 211 then 223, verified via ledger.append's return value after each run) — no discrepancy to flag this time, unlike the universality sweep's 137-vs-133 mismatch.
• min_periods on the 30-day rolling window: a stricter min_periods (e.g. requiring a full 30-day history before assigning any regime) would have NaN'd out the first ~30 days of the research split, and the n-split check is the mechanism that would have caught any resulting event loss. min_periods=1 was chosen and verified costless (zero up_ge8 events fall in the ~2-day true warmup window — rvol_288 itself needs a full 24h to be non-NaN — on either instrument); this is recorded as a deliberate, checked choice, not an unexamined default.
• Daily-resample method (mean vs last): the declaration says "daily-resampled" without specifying mean/last/other; .resample("1D").mean() (averaging rvol_288's ~288 intraday observations per day) was used rather than an end-of-day snapshot (.last()), on the reasoning that a mean is less exposed to a single noisy point-in-time reading and is the more standard resample convention when collapsing a rolling series to a lower frequency. Documented here since the declaration does not pin this down and a .last() reading would be a legitimate alternative.
• Rank taken on the daily series, then broadcast (rather than ranking the already-broadcast M5-level series) — avoids implicitly weighting a calendar day's contribution to the rank by however many M5 bars that day happens to have (data gaps are rare in this corpus but not absent); a design choice in the runner's favor of the daily series being the natural unit of "a day's vol regime," consistent with "trailing 30-day" being a per-day, not per-bar, concept.

Consequence

Per the declaration's own framing ("Either result changes the memo; that is the point"): this result is neither of the two clean outcomes the declaration anticipated. It is not "regime-flat, original candidate's case recovers" (BTC's high band and ETH's low band both show elevated point estimates that would, if CI-clean, argue for concentration) — but it is also not "high-vol concentration confirmed" (no band on either instrument clears the declared CI-overlap bar, and ETH's shape actively contradicts the high-vol-favors reading). The honest read for Richard's look-2 decision: this specific 3-way vol-band partition is underpowered to settle the E2 question at ~150-215 events per cell: BTC suggestively supports high-vol concentration (monotone ranking, though no CI clears), ETH suggestively contradicts it (low-vol band is the standout, mid is actually negative), and the 2020 band-year table falsifies the "2020 was high-vol" half of the pre-declared prior on both instruments (2020 events concentrate low/mid, not high). A conditioned successor spec, if pursued, is a new registration (per the declaration) and would need either a coarser 2-way split (more events per band) or a larger combined instrument pool before the vol axis itself can be resolved — the present result does not license either "regime-flat" or "regime-concentrated" as the basis for a validation-object decision.