BTC Order-Flow-Confirmed Range Breakout, Long-Short (Single-Instrument BINANCE USD-M Perp, DEEP aggTrades, Trade a Range Break ONLY When a Same-Direction TAKER-FLOW SURGE Confirms Real Demand Is Behind It — Flow-PRIMARY Continuation, the Non-Price Filter That Separates Genuine Breaks From the False Ones OHLCV Breakout Can't — Fast Hold, ATR Stop, 3-Parameter)

A LONG-SHORT, single-instrument MICROSTRUCTURE CONTINUATION strategy on BTCUSDT.BINANCE (USD-M perp) — deliberately the OPPOSITE direction from my three pending order-flow FADE ideas (CVD absorption-divergence, large-print sweep exhaustion, perp-index overshoot), to diversify the one live space this session left open. It fits the viable-space lessons: L30 (pure-OHLCV trend/breakout = 0/213 → the PRIMARY signal here is TAKER ORDER FLOW from the DEEP aggTrades feed, a genuine non-price dependency L30 permits; the range break is only the context), L29 (no options), L28 (no funding carry — immune to the funding-edge decay). The core insight addresses exactly WHY pure-OHLCV breakout dies: most range breaks are false (no real demand), so an unfiltered breakout has no edge. This trades a break of a short rolling range ONLY when it coincides with a SURGE in same-direction aggressive taker volume (computed from aggTrades) — i.e. real, aggressive buyers/sellers are lifting/hitting through the level, not a thin drift. Flow-confirmed breaks are the minority that continue; the flow surge is the filter OHLCV cannot see. Enter in the break direction on flow confirmation, ride the short continuation, exit on flow exhaustion or an ATR stop. BTC is chosen for the deepest, cleanest aggTrades tape and tightest execution. Single instrument, LONG-SHORT (fills the 85%-long imbalance), short intraday hold. 3 core parameters (range lookback, taker-flow surge threshold, ATR stop/trail).

Iteration-3 fix targeting the CRITICAL Layer-4 rejection: the code implemented open-interest expansion, but the hypothesis's stated core edge is TAKER ORDER FLOW ('same-direction taker-flow surge'). Smallest correct change: swap the confirming signal from OI back to a DIRECTIONAL aggressive-taker-flow surge (metrics taker_long_short_vol_ratio -> taker imbalance z, gated same-direction: >=surge_z for up-breaks, <=-surge_z for down-breaks), so the code now implements the mechanism it is filed under; the entire breakout/ATR-trail/continuation structure that already passed Layers 1-3 is unchanged. I also addressed the reviewer's secondary points: (4) the carry-forward z-baseline bias is removed by z-scoring the current taker imbalance against a baseline of DISTINCT metric samples (appended only when the underlying metric timestamp advances), so repeated/carried values no longer deflate the baseline std; (3) trade frequency is lifted by loosening the break margin (0.2->0.1 ATR) and the surge gate (1.2->1.0) toward the >100-trade measurability floor; (2) per-trade gross is given more room to clear the fee line by letting winners run (max_hold 16->24 bars) under the wide ATR trail while losers are cut at the trailing stop and the taker-reversal exit only fires on a strong opposite flow (exhaust_z 0.8->1.0) so it doesn't clip winners early. Data reality (disclosed honestly in the docstring): the raw aggTrades tape is un-wired/un-testable in this factory, so the deepest available taker-flow signal is the 5-min metrics taker buy/sell volume ratio; this is the aggregate stand-in for aggressive taker flow and is the correct feed for the stated edge. Flow is PRIMARY (the range break is only context), so this is not pure OHLCV (L30-acceptable). Risk controls unchanged: single 0.5x-equity position, no leverage, ATR trailing stop -> no liquidation/drawdown blowups (L15); min_notional=100 respects BTCUSDT USD-M minimums; long-short fills the long-imbalance bucket.

Verification failed (Layer 4 — QA review): - The hypothesis's core edge is explicitly TAKER ORDER FLOW: the title and description name a 'SAME-DIRECTION TAKER-FLOW SURGE' computed from the DEEP aggTrades tape as the PRIMARY confirming signal (aggressive buyers/sellers lifting/hitting through the level). The code does NOT read taker flow at all — it substitutes OPEN INTEREST expansion (_oi_surge_z from the 5-min metrics open_interest column). OI expansion (position build-up) and aggressive taker volume (who crosses the spread) are DISTINCT microstructure signals; a break can occur on rising OI with balanced or opposing taker flow. Per the QA mandate, the hypothesis's stated core edge is absent from the code = critical: the strategy is a different mechanism than the one filed under this hypothesis. The developer transparently documents the swap (aggTrades un-wired, taker-ratio proxy edge-dead), which is honest, but that does not make the substituted code implement the hypothesis. It should be re-filed under an OI-confirmed-breakout hypothesis. - Fee viability fails on the measured sandbox result. avg_trade_return_pct = 0.097% is below the 0.15% futures floor and barely covers the ~0.10% round-trip taker cost; the dollar-weighted result is already a net loss (avg_trade_pnl_usd -$20.88, profit_factor 0.84, total_return -1.90%, Sharpe -0.92, CI [-2.39, +0.78]). The break-margin / fresh-cross / higher-surge changes did not lift per-trade gross above the fee line. Same sub-fee, no-net-edge signature that has killed this family repeatedly. - 91 trades over 364 days is below the ~100-trade measurability floor; the fresh-cross + 0.2-ATR margin + higher surge gate further suppress trigger frequency, so the result cannot cleanly distinguish edge from noise. - OI is carried forward when no metrics row falls in a 15m bar, so carried-forward bars contribute an exact-0 OI log-change into _oi_surge_z's baseline, deflating std(dlog) and biasing the z upward on the bars where OI actually updates (metrics cadence vs bar cadence mismatch).

Verification failed (Layer 4 — QA review): - The edge is decisively negative on a now-measurable sample, and it is NOT a code defect. 367 trades (metrics_reliable=true): avg_trade_return_pct -0.0074% (below the 0.15% futures floor and net-negative), avg_trade_pnl_usd -$68, total_return -25.04%, profit_factor 0.56, win_rate 0.36, Sharpe -1.64 with CI [-3.16, -0.002] lying ENTIRELY below zero, max_drawdown 25.3%. Signal polarity is correct (up-break + buyers-dominating z -> LONG, down-break + sellers-dominating z -> SHORT; exits on opposite-flow flip), so this is not a polarity/sizing bug — it is a genuine falsification of the continuation thesis. A 0.36 win rate means taker-flow-confirmed breaks FADE more often than they continue. With 367 trades this is decisive, not noise. - The confirming signal is again the 5-min metrics taker buy/sell VOLUME ratio proxy, because the DEEP aggTrades tape the hypothesis names is un-wired/un-testable here. That proxy is already closed-loop established as edge-dead across this exact family in BOTH fade and continuation orientations (per the developer's own rationale). This is the 6th variant, and the well-sampled run confirms the continuation orientation is dead too (PF 0.56, Sharpe CI fully negative). The correct Layer-4 action is to close the loop rather than spend another backtest/optimization cycle rediscovering a documented-dead signal. - High-turnover fee/impact death compounds the negative edge: turnover 228, 367 trades in 364 days, total_impact_usd $6,941 on ~$100k. A negative gross signal plus this cost stack produces the -25% observed.