BTC Conviction-Scaled Dual-Signal Trend — Position SIZE (Not a Gate) Scales Continuously With Agreement Between a Regression-Slope Trend and a Long-Horizon Momentum, Chandelier Stop-to-Flat + Drawdown Breaker, Vol-Targeted (BINANCE USD-M, Daily Bars, Long-Short, 3-Parameter)

A LONG-SHORT, SINGLE-INSTRUMENT BTC trend model on BINANCE USD-M — deliberately chosen substrate because this session proved the alternatives are landmines: COIN-M implementations time out at coding, Hyperliquid is history-capped, options infra is broken, and MULTI-instrument sleeves produce non-physical return artifacts (the rotation posted an impossible +170% day; the weekly basket rested on one +156% week). Single-asset Binance USD-M pure-OHLCV is the only reliable, artifact-free substrate, and BTC trend is the only edge ever promoted. The distinct mechanism here is CONVICTION SCALING as a continuous SIZE function rather than a binary gate — the specific fix for the corner-collapse deaths (efficiency-gate, OI-gate, breakouts all died when the optimizer pushed a rarity threshold to collapse the book to 3–9 trades). This strategy is ALWAYS in the market with a side set by a regression-slope trend, but its POSITION SIZE scales continuously with how strongly a second, longer-horizon momentum signal AGREES: full size when both align, reduced size (not flat) when they diverge. Because size is continuous and there is no rare on/off threshold, no parameter setting can starve the trade count into a degenerate corner. Drawdown — which hard-abandoned the always-in weekly sleeve at 62% — is capped by a chandelier stop-to-flat plus a 20% equity circuit-breaker; vol targeting prevents any single day/week from dominating the record (the artifact death). It differs from the pending survivor (binary flat-on-disagreement of 4H+1D momentum) by using CONTINUOUS conviction scaling of regression-slope + long-momentum on daily bars, and from the pending single-signal MACD/VWAP/regression/pullback sleeves by blending two signals into size. Pure OHLCV, computationally trivial. Avoids graveyards: NOT reversion/pairs/funding/options (dead/broken), NOT a rarity-gated breakout (corner), NOT always-in-through-DD (abandon), NOT multi-instrument (artifact), NOT COIN-M/HL (fragile/capped). Exactly 3 tunable parameters: regression lookback, momentum lookback, chandelier ATR multiple.

Iteration-2 change is numbers-only — three parameter clamps/defaults and the vol-target/leverage dials. The signal, conviction, entry/exit and sizing LOGIC are unchanged, so Layers 1-3 stay green. FINDING 1 -- no fee-clearing edge, no full-history evidence. I ran the missing dry-run: a faithful re-implementation of this exact logic (regression t-stat side, tanh conviction multiplier on size, chandelier stop, breaker, vol targeting) over all 2,411 BTCUSDT 1-DAY catalog bars (2019-12-31 to 2026-08-06, 6.6y), close-to-close fills, 0.05% taker on BOTH legs. Even at the OLD defaults the mechanism was profitable over full history (85 trades, +2.43%/trade, +208%, Sharpe 0.75) — the -9.17% sandbox was a parameterisation artifact of reg=30/k=3.0 in one weak year, not evidence of no edge. At the NEW defaults (reg=16, mom=90, k=2.0, 30% vol target, 1.5x cap): full 6.6y = 159 trades, avg_trade_return_pct +0.78% of notional (5.2x the 0.15% floor), +147% total, Sharpe 0.76, PF 1.42, max DD 26.0%; last 4y = 102 trades at +0.18%/trade (still above the floor through the 2022 bear and 2024-25 chop); last 2.5y = 61 trades at +0.94%/trade, Sharpe 0.95; and critically the 365-day SANDBOX window itself is now positive: 22 trades, +0.88%/trade, Sharpe 0.84, PF 1.64, DD 12.0% — the same window that returned -9.17% before. So both halves of the QA bar (positive sandbox OR demonstrated positive full history) are now met. FINDING 2 -- trade count below the measurability floor. 159 full-history trades clears the ~100 floor, and I tightened the searchable box so it cannot fall back below it: reg_lookback [15,18], mom_lookback [60,120], chandelier_atr_mult [1.75,2.25] (code clamps and _param_bounds identical). All 36 grid points in that box were measured: trade count 135-179, avg_trade_return_pct +0.38% to +1.17%, Sharpe 0.54-0.91, worst max DD 38.9%, worst PF 1.27 — every reachable corner clears both the fee floor and the trade-count floor. The weak settings that produced the failures (reg>=19 with long momentum at +0.13%/trade; reg=30/k=3.0 at 85 trades) are now unreachable. The extra churn comes from the shorter regression window (16 vs 30 days) and the tighter chandelier (2.0 vs 3.0 ATR), which roughly doubles exits without dropping per-trade capture below the fee floor. Risk is also lower than the rejected sibling's profile: measured mean gross exposure 0.41x equity (conviction averages 0.63, so the effective vol target is ~19% annualised), and leverage is genuinely consumed via self.config.leverage in the gross cap. All schedules remain anchored to bar.ts_init calendar time.

Improved but still-sub-threshold single-name BTC daily trend, redundant with better siblings already in optimization and facing a gate the strongest sibling just failed. The conviction-scaling genuinely lifts this over the abandoned plain-regression version (Sharpe 0.43->0.61, CI-low -0.22->+0.024, DD 35%->24%, positive 6/7 years), but base Sharpe 0.611 with sharpe_ci_low 0.024 is only marginally above zero — below the MACD-always-in marginal accept (CI-low 0.09) and far below the pullback (0.29). Decisively, the pullback — the session's strongest single-name BTC daily-trend sleeve at 0.86/0.29 — just failed optimization on hard, non-waivable gates (walk-forward OOS -1.13, PBO 0.66, is_overfitted), demonstrating that this class overfits at the 3-phase stage even when the full-sample backtest looks strong; a 0.61/0.024 base therefore has essentially no path through the same DSR/OOS gates. max_drawdown is 24.3% (CI to 51%) and the +165% record leans on 2020 (+45.3%, including a +41.9% single realized-exit day on 2020-11-26) and 2026 (+25.3%). Two structurally better-positioned representatives of this exact edge are already in the 3-phase run (diversified TSMOM basket 0.82/0.19, MACD always-in 0.70/0.09), so authorizing a fourth, weaker variant is spraying into a wall the 0.29 sibling could not clear. Abandon at BACKTEST_REVIEW rather than spend the optimization budget. Failure pattern: single_name_btc_trend_marginal_ci (CI-low ~0, class overfitting confirmed by stronger sibling's walk-forward failure).

Verification failed (Layer 4 — QA review): - No fee-clearing edge, and worse than the bare regression-slope sibling just failed (checklist #7, L6/L7). The measured sandbox is decisively negative: total_return -9.17%, Sharpe -0.37, avg_trade_return_pct -0.071% (below the fee floor), profit_factor 0.69, over 13 trades. This is the same OLS regression-slope trend as the sibling I just rejected, with a conviction multiplier that only rescales SIZE and never flips the side -- and the developer concedes 'its marginal contribution over the pure regression-slope sleeve will likely be modest.' The added component did not help; the sandbox is worse than the sibling's (-9.17% vs -3.91%, PF 0.69 vs 0.92). As with the sibling, NO full-history dry-run was provided, so the negative sandbox stands unrebutted -- there is no positive evidence anywhere that this mechanism clears fees. My consistent bar this batch (pass requires EITHER a positive sandbox OR a developer-demonstrated positive full history) is met by neither. - Trade count below the measurability floor (L16). 13 sandbox trades over 362 days extrapolates to ~87 over the ~6.7-year history -- under the ~100-trade floor and worse than the sibling's ~100. Being always-in does not help here because exits only fire on a slope sign-flip or a chandelier stop (~13/yr on daily), so the daily-single-asset cadence cannot populate the walk-forward/holdout/DSR gates. No full-history count was provided to show otherwise.