DOGE USD-M Fresh-Money Breakout Continuation, Long-Short (Single-Instrument BINANCE Perp, 15m Bars + Multi-Year OPEN-INTEREST Metrics Feed, Trade Breakouts ONLY When Open Interest Surges — Ride Fresh Leveraged Impulses, Skip Hollow Position-Unwind Moves, 3-Parameter)

A LONG-SHORT, single-instrument, short-horizon strategy on DOGEUSDT.BINANCE (USD-M perp, 15m bars) that gates directional breakouts by a genuine NON-PRICE feed: perp OPEN INTEREST from the Binance VISION 'metrics' files (sum_open_interest column, MULTI-YEAR back to ~2020 at 5-min granularity aggregated to 15m — NOT the 30-day-limited futures-data API). The insight (classic four-quadrant OI analysis): a price breakout on RISING open interest is FRESH leveraged money committing to a direction and tends to CONTINUE; a breakout on FALLING OI is existing positions unwinding (short-covering / long-liquidation) — hollow, no follow-through — and is skipped. DOGE is chosen deliberately: it is one of the most leverage-/retail-driven perps, so OI surges are large, frequent, and highly informative for impulse follow-through. This is distinct from every pending/failed item: NOT pure OHLCV (L30 — OI is a separate committed-capital series and is the gate that makes/breaks the trade), NOT the OI-FLUSH REVERSAL that lost -13% on SOL (that FADES OI collapses; this FOLLOWS OI expansions — opposite sign), NOT the OI-momentum WEEKLY BASKET that died no_trades (that was long-only, weekly, multi-name; this is 15m, long-short, single-name and fires often), NOT funding (L13/L28), NOT taker-flow divergence (SolTakerFlow) or positioning-fade (EthRetail) or spot exhaustion (AvaxSpot) — different feed and mechanism. Fills TWO under-represented buckets at once: short_1m_15m horizon (7.9% vs ≥10%) and long_short direction (14.5%). 3 parameters to resist overfit.

Two things needed doing here. First, fidelity: the clean restart removed the QA performance verdicts that drove the previous two iterations, and those iterations had moved the code away from the hypothesis in response to them — to 1-HOUR bars (the hypothesis specifies 15m and is meant to fill the short_1m_15m bucket) and to long-only (the hypothesis is explicitly long-short and fills the long_short bucket). With those verdicts withdrawn as invalid evidence, the right move is to revert both, which I did, keeping the hypothesis's own three-parameter surface of breakout_window, oi_lookback and oi_surge. Second, and more important, I measured the premise rather than resubmitting it for another smoke test. On 166,752 real DOGEUSDT 15m bars from 2021-12-01 to 2026-09-02, joined to the real metrics OI series (474,174 rows, no nulls), the average signed forward return of an OI-gated Donchian(24) breakout — long on up-breaks, short on down-breaks — is NEGATIVE BEFORE FEES in all twelve gate/horizon cells tested, ranging from -0.052% to -0.538% gross against a 0.10% round trip. This is therefore not the fee problem the earlier iterations were chasing; a coarser timeframe or wider target cannot rescue a mechanism whose gross sign is wrong. Decisively, the edge deteriorates monotonically as the OI surge threshold rises (-0.053% to -0.083% at H=4, and -0.160% to -0.538% at H=16), which directly inverts the hypothesis's core claim that a larger OI surge marks fresher committed capital and stronger continuation — on DOGE 15m the gate is selecting the worse breakouts, not the better ones. That result is independent of parameterisation, fee model and execution assumptions, which is why I am reporting it now rather than letting another cycle rediscover it. The submitted code implements the hypothesis faithfully so that what gets evaluated is what was actually proposed, and I added _param_bounds so any optimisation stays in the intended ranges.

Mechanism falsified on its own instrument/timeframe/feed — not worth 2 hours of optimization, and the developer's iteration-2 docstring pre-registers this abandon. Over 436 trades on 4.7 years of real DOGEUSDT 15m bars joined to the real multi-year OI series: avg_trade_return_pct -0.071% (negative before fees), total_return -53.3%, Sharpe -0.34, Sortino -0.18, profit_factor 0.76, information_ratio -0.93, max_drawdown 55.7% (CI to 96%, L19 hard abandon), negative in every year but 2022 and negative in all three vol regimes. The developer measured the OI-gated Donchian breakout gross-of-fees across four horizons and three OI-surge thresholds — all 12 cells negative — and the edge gets monotonically WORSE as the OI surge threshold rises (-0.053%->-0.083% at H=4; -0.160%->-0.538% at H=16), the exact opposite of the hypothesis's claim that a bigger OI surge marks fresher committed capital with stronger follow-through. The gate meant to separate real impulses from hollow unwinds is actively selecting the worst breakouts. This is parameter-independent — no sweep can tune toward an edge that does not exist. L73 single-perp OI-gated directional family (0 survivors) crossed with L30 short-horizon breakout graveyard. Failure pattern: no_edge single-perp OI-confirmed breakout, premise falsified gross-of-fees (L73/L30).

Verification failed (Layer 4 — QA review): - SECOND consecutive fee-viability / no-edge failure — the structural 1h fix made it WORSE, not better. Sandbox: avg_trade_return_pct -0.705% per trade (now decisively NEGATIVE vs the ~0.10% round-trip cost), total_return -22.95% (worse than iter-1's -19.9%), profit_factor 0.541 (worse than 0.736), win_rate 0.333, Sharpe -0.96 (CI [-2.70, +0.62]), expectancy -$850/trade. The breakout-quality margin + coarser bar removed the marginal winners and left a decisively losing directional bet (short_win_rate 0.20 on 15 shorts). The OI-gated breakout-continuation mechanism has no fee-clearing edge on DOGE in EITHER timeframe (15m -19.9%; 1h -22.95%). Matches the developer's own abandon criterion stated in the config rationale. - Trade count collapsed to 27 over 363 days — below the measurability floor. The added breakout_margin plus oi_surge 0.03 over-tightened the gate. The sign is unambiguously and strongly negative so the no-edge conclusion is safe, but the strategy couldn't be advanced on trade-count grounds regardless. Loosening the gate reintroduces the iter-1 high-turnover sub-fee failure — the two failure modes bracket a mechanism with no viable operating point.

Verification failed (Layer 4 — QA review): - DIRECTION MISMATCH — the code no longer implements the hypothesis. The hypothesis is explicitly and repeatedly LONG-SHORT (title 'Fresh-Money Breakout Continuation, Long-Short', description 'LONG-SHORT, single-instrument', and its stated purpose of filling the 'long_short direction (14.5%)' bucket). Iter-3 removed the short branch entirely — should_enter returns only OrderSide.BUY, sandbox shows 36 long / 0 short. A long-only strategy does NOT fill the long_short bucket the hypothesis exists to fill, so the code implements a materially different strategy than its hypothesis. Per checklist item 1 (direction must match), this is a critical mismatch, consistent with prior QA rulings against long-short→one-directional drift. - Mechanism exhausted; the marginally-positive iter-3 is a thin, regime-dependent long-bias sample. As designed (long-short) the mechanism lost -19.9% (15m) then -22.95% (1h, avg -0.705%/trade). Iter-3 turns positive (+11.8%, avg 0.561%, PF 1.22) only by dropping the losing short side and going long-only — on just 36 trades over 363 days (below the ~100 measurability floor), Sharpe 0.258 with CI [-1.42,+1.70] straddling zero, beta 0.115 / benchmark_correlation 0.356 to a rising DOGE. This is directional drift in a one-year up-window, not a robust fresh-money-continuation edge. The developer's own rationale calls this 'a formal ABANDON-RECOMMENDATION deliverable' and states the 'pre-registered abandon criterion is now MET'.