# Registered flow-entry timing study v1 — September 8, 2026 ## Question and unchanged inputs The preceding flow study found a promising 24-hour-lag entry gate, while the primary four-hour gate failed. Test whether that result is locally stable and distinguish its performance from arbitrary entry filtering. This is follow-up exploration of already inspected history, not an independent holdout or parameter selection contest. Use exactly the frozen price, original decision and Binance BTCUSDT spot-flow inputs from SPOT_FLOW_PROTOCOL.md and SPOT_FLOW_PROVENANCE.json. All acquisition, quarantine, unit conversion, price warmup, missing-slot and historical-publication limitations remain. No new data or outcome-based data cleaning. Do not modify the preceding evaluator. ## Registered models Retain the same independent long/cash state machine: normal entry above the prior 120-bar high; shortcut above the prior 30-bar high and the current 120-close SMA; held-long exit below the prior 60-bar low; exit priority; current candle excluded from channels; full history warmup before account slicing. Only the shortcut gate changes. 1. Nine positive-flow variants with availability lag in four-hour bars {1,2,3,4,5,6,7,8,12}, corresponding to 4,8,12,16,20,24,28,32,48 hours. Six contiguous flow bars end at decision close minus lag. Require every bar and positive quote volume, with volume-weighted buy share strictly greater than 0.50. The 24-hour variant is the previously observed center; 20 and 28 hours are the primary local neighbors. All nine must be reported. 2. Inverted-direction negative control: same 24-hour feature, shortcut allowed only when buy share is strictly below 0.50. Exact 0.50 and unavailable flow block the shortcut. This is a diagnostic control, not a nominated alternative strategy. 3. Stale-flow negative control: same positive-flow rule with a fixed seven-day lag (42 four-hour bars). This may contain persistent market information, so it is an imperfect placebo; do not label it independent noise. 4. Sixteen deterministic coin-flip controls, seeds 0 through 15. At each decision, shortcut passes iff the first eight bytes of SHA256 of ASCII `flow-timing-v1||` interpreted unsigned big-endian are less than 2**63. Also require that the center's 24-hour feature is available, so missing-flow permissions match. No price/future outcome enters the hash; no seed selection. These have 50% permission probability, not the empirically observed flow permission rate, and are descriptive controls rather than a matched randomization test or statistical p-value. 5. Four references: original targets, old 120/60 breakout, unchanged price-only early-entry rule, and BTC buy-once/hold. Total 31 models: nine flow lags, two named negative controls, sixteen hash controls, four references. Missing flow disables only the shortcut, never the normal breakout or an existing position. No leverage or trained feature. ## Evaluation and interpretation Use the exact preceding 14 windows, per-dollar trading costs {6,30} bps and execution delays {1,3} four-hour bars: 31 x 14 x 4 = 1,736 account cases. Primary 6 bps and four-hour execution delay. Feature lag is additional to execution delay. Same post-fee target accounting, fixed quantity between fills, expiring scheduled fills in gaps and observed-close drawdown. No terminal liquidation, cash interest or original-short funding/borrow. No stitching returns across windows. Windows and model outcomes are correlated, not independent samples. Primary local-stability diagnostic: all three lags 20/24/28 hours must improve over the price-only shortcut in 2019-through-frozen-end, 2023-through-frozen-end and recent June-September under all four cost/execution scenarios. This is a deliberately demanding follow-up diagnostic, not a promotion gate. Failure means the prior center cannot be presented as a timing-robust improvement. Even passing does not establish a return advantage over BTC or future profitability. Report the entire neighborhood and compare each with BTC and old breakout. Report annual results and earlier 2018/2019-2022 weakness. Never choose the best lag or seed after seeing results. For every window/cost/delay, report the center's return relative to each of the 16 hash controls, their median/range, and count strictly above/equal to center. Do not call this count a p-value. Report inverted and stale controls separately. For primary-cost/delay center and its two immediate neighbors, partition incremental log wealth versus the price-only shortcut into chronological exposure-divergence episodes. Each observation's contribution is log(candidate equity growth) minus log(control equity growth), including entry/exit fees. An episode opens when the two actual post-fill long/cash exposures diverge and ends when they rejoin; include the joining interval. Both accounts start at $10,000, so all contributions plus any shared-exposure numerical residual must reconcile to log(ending candidate wealth / ending control wealth). Retain every episode, largest positive contribution, share of total positive contributions, and net advantage after subtracting the largest positive episode. That subtraction is retrospective concentration analysis, not an executable strategy. Count signal transitions and retain dated recent transitions to distinguish repeatedly delayed entry from genuinely skipped trades. ## Verification and retained evidence Verify frozen dependencies and all source/input identities. Reproduce the 23,600 canonical decisions. Four references plus 4h/24h flow variants must exactly reproduce the previous 336 matching account curves (6 x 14 x 4), including targets, equities and BTC quantities. Verify future price/flow mutations leave earlier features/signals/accounts unchanged, missing flow falls back to old breakout, uniform quote-volume scaling preserves permissions, lag endpoints are exact, hash gates are repeatable and independent of future inputs, and concentration arithmetic reconciles on constructed examples and actual runs. New account and feature traces are hashed; no overwrite of earlier experiment paths. Run meaningful targeted tests before the historical run and full discovery after code changes. Save the complete metrics, checks, diagnostic episodes, all account traces and a compact digest, then archive with source/input provenance and a SHA256 receipt. Keep failures visible. No production or shadow change. The larger performance-improvement goal stays active pending a justified improvement and new timestamped forward evidence.