Pre-registration: streak_follow_btc_v1 — up-streak persistence execution test on BTCUSD

Date registered: 2026-07-24, before any execution-level test of this candidate. Lineage: hybrid — the "rhythm/streak" concept descends from the mentor's technique journals (R4 ladder rung); the specific mechanization (M5 close-streaks, ≥8 threshold, 4h harvest) is ours from the declared scans; stated per registry lineage rules. Evidence basis: R4 rhythm scan (2026-07-24-rhythm-falsification-scan.md) — BTC streak_up_ge8 +0.603 ATR/4h, CI [0.155, 1.051], 6/6 year-blocks, monotone dose — and its ETH replication (2026-07-24-upstreak-eth-replication-scan.md) — +0.536 ATR/4h, CI [0.047, 0.987], 4/6 blocks, dose and asymmetry reproduced. Amendment A1 second-instrument test satisfied at scan level. Scope: Research split ONLY (Binance BTCUSDT M1, 2017-08-17 → 2022-12-31). Validation (9/10 looks) and Vault untouched. Fresh strategy segment: run 1 of 3.

Ex-ante cost arithmetic (TEMPLATE requirement — computed BEFORE running)

• Measured event-time ATR14(M5): BTC up-ge8 median 0.273% of price (n=498, from the replication scan's descriptive add-on; ETH 0.326%).
• The 3× screen is stop-invariant in R (gross_R/drag_R = gross_bp/cost_bp): 4h scan gross 0.603 ATR × 27.3bp ≈ 16.5bp vs Vantage-class RT 3.1bp → 5.3× (ETH: 17.5bp → 5.6×). Dukascopy pessimistic bound 8.3bp → 2.0× (reported, not the decision case).
• 1h harvest: BTC 9.3bp ≈ 3.0× (at the line), ETH 7.4bp = 2.4× (fails) → the 1h harvest is NOT registered; 4h only.
• Break-even RT cost at the 4h harvest: ≈ 5.5bp (Vantage-class 3.1 sits 44% below).
• Kill rule: passes (≥3× at the realistic variant). Registration proceeds.

Frozen spec (zero free parameters)

• Event: up_streak == 8 on M5 closes (exactly 8, the scans' freshness rule — feature-store semantics, open-labeled grid), long only (the down-side never earned a registration). Signal at the 8th bar's close.
• Entry: open of the next M1 bar.
• Stop: 2.0 × ATR14(M5) at the event, below entry. (Declared for geometry, not fitted: a 1×ATR stop sits inside 4h noise — unconditional 4h |move| ≈ 3.7 ATR — and would convert the time-exit harvest into a stop-lottery; 2× is the smallest round multiple that keeps the median 4h adverse excursion plausibly survivable. The 3× screen is stop-invariant, so this choice affects truncation realism, not the admission test.)
• Exit: time exit at the open of the first M1 bar ≥ 240 minutes after entry, unless stopped. No TP, no trail — the faithful harvest of the scanned close-to-close move.
• One position at a time; events while in-position skipped.
• Arms: A1 primary (as above) · C1 control: same events/geometry, coin-flip direction (seed 20260724).
• Costs: charged as RT % of entry: Vantage-class 0.031% (decision case) + Dukascopy bound 0.0832% (reported). BTCUSD_VANTAGE conventions.

Registered hypotheses (4; cumulative 133 → 137 of 2,000)

• H1 — survives execution: A1 gross > 0 AND > C1 gross. (Stop truncation will cut the +0.60 close-to-close; disp-follow's 1×ATR/1h haircut was ~40% — with 2×ATR/4h the truncation is milder but real.)
• H2 — survives realistic costs: A1 net (Vantage) > 0.
• H3 — year stability carries to execution: A1 gross positive in ≥4 of 6 year blocks.
• H4 — drag law: measured mean drag within ±30% of cost/stop prediction (0.031%/(2×ATR%) per trade; charged-not-price-paid caveat as always).

Pre-declared "worth a Validation conversation" bar (NOT a Validation trigger)

All four: A1 net (Vantage) ≥ +0.05R · n ≥ 200 · beats C1 net · ≥4/6 year blocks gross-positive. Meeting it triggers a decision memo to Richard (his look, his call — lesson of run 20 fully priced in: research-split success is a hypothesis about the past).

Honest prior

Genuinely uncertain. For: two-instrument scan replication, 6/6 BTC year signs, 5.3× ex-ante margin. Against: the 4h magnitude is front-loaded in 2017–18 thin-liquidity years on both instruments; a 2×ATR stop still truncates some winners' paths; n≈450 after in-position skips is thinner than disp-follow's. Expected failure mode: H1 holds weakly but H2 lands in the −0.05..+0.05R dead zone. And the run-20 lesson stands: even a clean pass here is a research-split fact, not an edge claim.

Iteration budget

Run 1 of 3 on this segment. Any change after seeing results = new registration.

Result (filled 2026-07-24, after the single registered run)

ALL THREE execution hypotheses (H1–H3) HELD; the pre-declared bar is MET. H4 (drag law) did NOT hold within tolerance — measured drag ran ~46% above the scalar ex-ante prediction (see below; not a bar criterion). Registry run 21 (streak_follow_btc_v1, lineage hybrid); artifacts reports/streak-follow-btc-v1/, dashboard reports/review-streak_follow_btc_v1.html.

Event count vs the rhythm scan: 498 up_streak==8 events (exact match to the scan's n=498). 11 dropped to in-position skips → 487 traded, pre-cutoff. 0 dropped by the split-end entry-cutoff (last event closes 2022-12-28, ATR-based 240m window never approaches the split boundary) → 487 final traded.

• H1 ✓ survives execution: A1 gross +0.1785R > 0 and > C1 gross +0.0950R — the 2xATR/240m harvest truncates the scan's raw +0.603 ATR close-to-close move (as expected) but keeps a real, control-beating edge; the coin-flip control's small positive gross reproduces the same post-event-volatility artifact seen in disp-follow's C1 and costs would erase it regardless.
• H2 ✓ survives realistic costs: A1 net(Vantage) +0.1074R, n=487.
• H3 ✓ year stability carries to execution: gross positive in 4/6 year blocks (2017 +0.273 · 2018 +0.389 · 2019 −0.017 · 2020 −0.056 · 2021 +0.293 · 2022 +0.225) — both thin-liquidity-era misses (2019, 2020) are near-zero, not catastrophic; both post-2020 years and both early years positive.
• H4 — drag law, NOT within tolerance. Measured mean drag (gross − net-Vantage) = 0.0711R; the ex-ante scalar prediction 0.031%/(2×mean-ATR%) = 0.0486R → ratio 1.46×, outside the pre-declared ±30% band. Root cause: atr_pct (2×ATR14/entry) is right-skewed (mean 0.64% vs median 0.55%, max 4.7%), and drag is a convex function of it (cost/atr_pct) — Jensen's inequality means the true mean drag exceeds the mean-ATR-based point estimate. A median-ATR-based version of the same scalar formula (0.031%/(2×0.55%) = 0.0569R) gives ratio 1.25×, inside tolerance — flagged here as an honest sensitivity, not a re-run: the mean-ATR convention was fixed before looking at results (matching disp-follow's own H4 precedent language, "predicted ≈ cost/mean-ATR"), and it is the one reported as the test. The drag direction and rough magnitude are still right; the point estimate undershoots on this event set's fat right tail.
• Bar: net(Vantage) +0.1074R ≥ +0.05R ✓ · n=487 ≥ 200 ✓ · beats C1 net (+0.1074 > +0.0239) ✓ · 4/6 year blocks gross-positive ✓ → BAR MET.
• Distribution note (honest, not a bar criterion): per-trade net R is heavy-tailed (std 2.36R, max +16.99R, min −1.45R, cumulative max drawdown −31.6R against a +52.3R total) — a direct consequence of normalizing R by event-time ATR14 across a sample whose raw ATR spans roughly 150× (2.76 to 420 price units). A handful of low-ATR events that keep running produce large R-multiples; this inflates both the upside and the drawdown when expressed in R terms. Worth carrying into any Monte Carlo pass (Helm #32) before this candidate goes further — the R-normalization itself, not just win rate, drives tail risk here.

Per pre-commitment, nothing further runs without Richard: this is an execution test on the Research split; it does not by itself license a Validation look or any change to the frozen spec. Next steps (R7 S/R engine, a possible Validation conversation, or a fresh registration for any parameter change) are his call.