# Registered sustained-flow experiment v1 — September 8, 2026

The one-day 24h-lag feature failed its local timing test. A week-old control did better over some long windows, which may reflect persistent conditions or retrospective luck. Test the separate hypothesis that volume-weighted aggressive buying over multiple days adds information to early price entries.

Use the exact frozen inputs from FLOW_TIMING_PROVENANCE.json. All 14 prior windows, warmup, missing-data, source/instrument, historical-availability and revision limitations remain. No refresh or retrospective cleaning. History is already explored.

The price state machine remains exactly entry_path from evaluate_flow_timing.py: close above prior 120-bar high enters unconditionally; close above prior 30-bar high and current 120-close SMA allows an optional shortcut; exit below prior 60-bar low. No flow exit and no leverage. Signals warm on all prior prices before account slicing.

Six candidates combine aggregation horizons {3,7,14} days (18,42,84 contiguous four-hour bars) with availability lags {1,6} bars (4h,24h), additional to execution delay. At decision close D, the latest constituent flow close is D minus the availability lag. Sum taker-buy quote volume divided by sum total quote volume over exactly that horizon; require strictly above 0.50 to permit the shortcut. Every row must exist with positive quote volume. Missing, quarantined or zero-volume constituent bars invalidate the whole feature and disable only the shortcut. This measures aggregate pressure; it does not require every individual day to have a buying majority.

Two negative controls reverse the seven-day feature to strictly below 0.50 at each availability lag. Seven exact prior references remain: original targets, old 120/60 breakout, price-only shortcut, BTC buy-once/hold, one-day flow at 4h/24h lag, and the week-old one-day positive-flow control. Total 15 models, 14 windows, costs {6,30} bps per dollar traded and execution delays {1,3} four-hour bars: 840 cases. All results retained; no choosing the best horizon/lag after seeing outcomes.

Report returns, observed-close drawdowns, costs, exposure and fills at every setting; show paired differences against price-only shortcut and BTC. Primary 6 bps/4h execution. Report the number of positive paired differences across all six candidates in the three core windows (2019 onward, 2023 onward, recent) and four cost/execution settings, 72 comparisons. A universal local-consistency claim requires all 72 to be positive; this is descriptive and not a promotion gate. Report earlier eras and annual failures. Inverted controls are diagnostics, not nominated replacements. Save all signals/features and 18 primary concentration records for the six variants across the three core windows using the previously tested divergence accounting.

Verify 23,600 canonical decisions, frozen dependencies and 392 unchanged reference curves (7 x 14 x 4), including targets/equity/quantity. Test exact variable-horizon boundaries, complete-window validation, normalization, future-price/flow invariance, missing-flow breakout fallback and long/cash account prefix invariance. All code is frozen before evaluation. Full suite, trace hashes, digest, provenance archive and workspace backup required. Original-short funding/borrow, cash interest and intrabar liquidation remain omitted. No production or cloud strategy changes; future performance is unproven.