Date: 2026-07-25. Status: synthesis of record — every claim below carries its source (registry run, ledger row, or scan doc). Research-split evidence unless marked otherwise. This document is the project's internal citation base: future declarations should cite these findings instead of re-deriving them, and contradicting evidence supersedes on sight.
A1. The cost-drag law: drag_R = round_trip_cost / stop_distance. Validated 4× on 2
instruments (GBPUSD, XAUUSD), consistent on BTC within Jensen skew (run 21's H4: the
point estimate needs the distribution of ATR, not its mean — right-skewed event ATR
makes realized drag run ~25–46% above the scalar estimate). Consequence: cost admission
is checked ex ante (TEMPLATE kill rule), and tight-stop styles die on high-cost venues
regardless of signal quality. Sources: COST-DRAG-FINDING-v1, runs 8/9/16/18/21.
A2. Event-conditioned coin-flip controls show positive gross. Every control arm on volatility events (runs 17/19/21 C1) grosses positive (+0.05..+0.25R) because events mark elevated-volatility moments with upward drift in crypto bull segments — and costs erase it. Consequence: a strategy must beat its event-matched control, never zero; "gross positive" alone is the event-vol artifact. This artifact bit expansion v1 (+0.98R gross → mostly artifact) before controls were standard.
A3. Vol elevation after events is universal physics; direction is not. Unsigned vol/range elevation replicates everywhere tested (BTC, ETH, gold, session blocks, streak events, H4 opens: rand-p ≈ 0 routinely). Signed direction is instrument- and regime-specific. Consequence: any scan reporting only unsigned effects has found nothing tradeable; the null battery's frequency-matched arm exists to say so.
A4. Time-of-day structure is real but is a conditioner, not an edge. Europe block (08–15 UTC) is the one genuinely vol-elevated session on BTC (+4–6% abs move, rand-p .000 all horizons; ledger 11). Session filters improved net/tails every time they appeared (runs 17/18 H3) — but the 09–16 window does NOT amplify event continuation (combination rule failed; in-window was weaker at 2 of 4 horizons). Consequence: session terms belong in risk/exposure control, not in signal claims.
B1. Compression→displacement continuation is crypto-class structure. After a large displacement bar (range ≥4–5× baseline, body ≥60%, prior compression), price continues in the bar's direction 15m–1h: BTC +0.29..+0.67 ATR, ETH near-identical (+0.72 at disp5), monotone in displacement size, both bears sign-stable (ledger 2, 6). Refuted on 10y gold — sign inverts (ledger 5). Not universal; crypto-class. Updated 2026-07-25 (universality sweep): 4/5 crypto instruments replicate (BTC, ETH, BNB, XRP; SOL under-powered-not-refuted at n=65). XRP is the cleanest reading in the project (+0.55/+0.83/+0.89 ATR at 15m/30m/1h, all CIs clear, monotone) — a harvest-viability question (XRP cost model) is a declared-lead, not a finding.
B2. …and it survived execution on BTC research data (+0.28R net Vantage, run 19) but FAILED the Validation holdout decisively (run 20): 2023–2025 direction information absent/inverted; the coin-flip control beat both signed arms; event frequency and ATR% decayed (drag 0.13→0.21R). Status: dead as a spec; the event class remains a research object. The definitive lesson: research-split selection flatters — see §E.
B3. WHERE the move happens matters more than size. The same-size move anchored at an H4 boundary shows zero-to-negative continuation, anti-monotone in size (ledger 9) — opposite of the anywhere-in-time result (B1). Boundary-anchored moves are a different population (scheduled-time repositioning, not organic order flow). Consequence: event definitions must not conflate calendar-anchored and flow-anchored volatility.
B4. The taught pullback-wait destroys this specific edge. Run 18's pullback-confirmation entry forfeits exactly the 15m–1h continuation window that carries the information (displacement close carries it; waiting spends it). Mentor technique and measured market structure disagree on entry timing for this family.
C1. Extended up-streak persistence is a two-instrument, asymmetric effect. After 8+ consecutive up-closes on M5: BTC +0.34 ATR/1h, +0.60/4h (6/6 years); ETH +0.23/1h, +0.54/4h (5/6, 4/6) — dose-responsive (ge8 ≫ ge5 both instruments), and asymmetric: down-streaks show no clean effect anywhere (BTC's breaks dose-response; ETH's never clears CI). Symmetric "rhythm" is dead; this is upside-persistence, plausibly short-squeeze/momentum-ignition microstructure. Ledger 10, 12. Updated 2026-07-25 (universality sweep): the map is patchy — 2/5 decisive (BTC, ETH), SOL supporting-only (power-capped), BNB narrow miss (4h CI crosses zero), and XRP INVERTS (−0.41 ATR/4h, CI-clean negative). The effect is NOT crypto-class the way displacement is; it looks specific to the large leveraged-flow venues — and the XRP inversion proves this family can flip sign across instruments, the same fragility displacement showed across time. Per the sweep's pre-declared score-keeping, this materially weakens the Validation look-2 case (recorded in the decision memo).
C2. It survived BTC execution: +0.107R net Vantage, n=487, beats control, 2021–22 positive (run 21, BAR MET). Caveats of record: magnitude front-loaded 2017–18 both instruments; ETH 2021/2022 negative at 4h; heavy-tailed per-trade distribution (few big winners carry the total — MC: p95 drawdown 68R, median path 40% under water). Pending: Richard's Validation look 2 decision; ETH execution test gated on the ETH cost model.
2026-07-25-streak-follow-eth-REFUSED.md).
Consequence: the streak candidate is BTC-only at this cost structure; ETH's role is
signal-replication evidence, not a second harvest venue. Signal universality and
harvest viability are separate claims — the cost law (A1) decides the second.E1. Research-split success ran 2-for-2 at the bar; the one Validation test failed. Selection-on-the-same-data flatters even a zero-free-parameter spec, because the hypothesis was selected on that data. Expected posture: research numbers are upper bounds; the holdout is the only arbiter (TEST-CATALOG "overfitting" section).
E2. Regime non-stationarity is the dominant risk for this asset class. Displacement died specifically post-2022 (post-ETF vol compression: event rate −30%, drag +58%, direction inverted). Any 2017-heavy effect must discount its early-year magnitudes; recent-year sign-stability is worth more than aggregate size. (This is why run 21's positive 2021/22 blocks matter more than its +0.107R mean.)
E3. The four-null battery + declared constants catch what matters — with one known degeneracy: label-shuffle is a no-op for one-directional conditions (recorded ledger 8/10/12); frequency-matched random + block bootstrap carry those cases. Block (never iid) resampling is load-bearing: clustered trade sequences show ≥1.5× wider drawdown tails than iid pretends.
E4. Execution haircuts are predictable in direction, not size: stop truncation cut the scanned close-to-close magnitude ~40% (run 19, 1×ATR/1h) and ~70% (run 21, 2×ATR/4h gross 0.18R vs scan 0.60 ATR). Scans license execution tests; they do not predict execution results.
E5. Fail-fast arithmetic works. The ex-ante cost screen has now refused one harvest (1h streak) before any spend — the cheapest kill in the project's history. Every refused registration is a positive result of the method.
Supersession rule: any future run/scan contradicting a finding above updates this document in the same commit as its own artifacts, citing both.