One list, three conditions, one score. This table is the whole decision: every symbol the optimizer considered, the stage it failed at, and — for the survivors — the order the scheduler fills its 8 bot slots in. A symbol opens a bot only if all three pass; it is then ranked purely on funding. Carry only — spread-mean-reversion ranking was removed on 2026-08-17: it was never simulated, and on these books the executable close spread (VWAP at leg size) runs ~1% against a −0.20 target, so a basis round trip cannot pay for itself.

1. Funding paid ≥ 0pp over 14 days — the thesis, and the income we are actually buying. v16.4 retired the old 1.0pp bar: tested alone it did not improve pick quality (per-position APY 6.9% → 7.0% → 7.4% going 0 → 1.0 → 2.0pp) while slot utilisation collapsed 100% → 86% → 48%. A floor of 0 keeps only the useful part: the symbol must have paid something net positive.
2. Funding still alive today (latest ≥ 0.02%/8h), so a symbol that paid well then died cannot coast in on history. Read as an EMA (α=0.05), not a spot value — changed 2026-08-17. Gate's quoted rate is a FORECAST of the next settlement and it resets at every boundary, rebuilding from the premium index over the interval. Measured over 3,644 settlements on 287 symbols, predicting the rate actually paid: EMA r=0.984 · interval mean r=0.953 · single sample r=0.878 · last settlement r=0.850. A single sample's own accuracy swings 0.84→0.98 depending where in the interval it lands, and it disagreed with the settled rate on 14% of 916k samples across 77% of symbols — which flipped marginal names in and out every settlement and cycled their bots.
3. Expected value must be positive — realized funding × persistence, minus the 0.30pp round trip amortised over the hold. A negative score literally means "expected to lose at these thresholds". Added 2026-08-17 after two symbols ranked at −0.029 and −0.041/day (breakeven 6.9d and 27.8d, one with negative mean funding and persistence 0.20) — both qualified on sporadic spikes rather than steady carry.

Score = realized funding × persistence, net of amortised fees. Lending is not in it, and no longer gates either.

THREE EXIT RULES REMOVED on 2026-08-23 — min-hold, dead-thesis, RULE 7

Measured on research/bot_sim_v16.js, Jan–Jul 2026 (212d, 30s books, 4 slots, 340 capital, entry bar 0.30, 2×). The three rules were tested together, because they interact:

+18.20 USDT (+38.6%) on the same capital, with 87 trades instead of 132 — fewer, more profitable, higher occupancy. Min-hold was the most expensive single rule (−14.68), almost entirely basis (+27.43 → +11.85): it forbids a spread_target exit for 75–120h, so positions that reach their basis target are held through the compression and give it back. It was also what made the dead-thesis exit look necessary — with min-hold on, the dead exit is the only way out of a locked position, so removing only the dead exit is worse than removing neither. They had to go together. RULE 7 (−7.96) recycles into fresher payers but pays the round trip twice. Restore with MIN_HOLD=1, CARRY_DEAD_EXIT=1, RECYCLE_ENABLED=1.

⚠ REGIME CAVEAT — WITHDRAWN 2026-08-24. It was a DATA ARTEFACT, not a regime. The original caveat said the gain was concentrated in Jan–May and the sign flipped late (Jan01–Mar15 +10.54, Mar15–May22 +8.38, May22–Jul31 −1.84, Apr15–Jul31 −4.89), and warned the change “optimises for a regime that may have passed”. That split was measuring the data source. The simulator's runtime funding gates read FundingHistoryDaily, which changes schema mid-window — true daily (dailySum) through April, 3-day smoothed (mean) from May — so the two halves ran on different data, and the flip sits exactly on that boundary. Re-measured with --true-daily (gates rebuilt from the settlement CSVs, one method across the whole window):

Removing the three rules wins in every period once the data is honest. Mechanism: the dead-exit floor is 0.0110 and Gate's house minimum is a point mass at 0.0100. The smoothed mean sat above 0.0110 far more often than the truth did, so it MASKED how constantly the dead clock fires — with true rates the all-rules-on arm churns down to 12.24 on Apr15–Jul31 where the smoothed run flattered it to 31.74. Do not re-derive a regime conclusion from any run without --true-daily.

⚠ The published 38.91% baseline overstates the live system. research/bot_sim_v16_result.json was generated with --no-rule7 --no-minhold, i.e. without two rules that were live. Live-faithful before this change was 30.68% deployed / 23.90% total, not 38.91%. Regenerate via research/export_sim_result.js at the live config.

Two gates were REMOVED on 2026-08-17 — and why

The lending gate (spot leg must lend ≥ 0.005 pp/day) is gone. Its evidence was real but unrepeatable: EarnRateDaily spans 21 days and EarnAprHistory stops at 2026-08-04, so it could not be re-tested on the 7-month set the rest of the strategy was validated on. Against it: PTB lends 0.59% APR — far under the floor — and is the best-paying position in the book at 109%/yr; and the reading itself was broken, forcing ppDay to 0 when history was thin, so AIO reported "5.7% APR = 0.0000pp/d" and was rejected as a non-lender on stale data. Restore with EARN_GATE=1 if live trades show it earning its keep.

The reachability gate (openP95 ≥ 0.05) is gone — it was costing money. Backtested Jan–Jul on 30s order-book data, 4 slots / 300 capital: no gate 34.28% APY vs p95≥0.05 21.62%, removing only 7 of 128 trades. The mechanism: p95 measures the body of the basis distribution, but entries happen in the tail — the last 10 live entries were 0.309–0.319 against a 0.30 bar. So it discarded symbols that sit flat and spike occasionally, which is exactly the profitable carry entry. Restore with CARRY_REACH_GATE=1.

Why a rejected symbol can show HIGHER income than a ranked one

Trailing income is history; the ranking is a forecast about which of it survives. A symbol can show strong 14-day funding and still be rejected because its EV is negative once fees are amortised, or because its funding is already dead today. The Total income column looks backward; the score does not.

Hold length dominates. Friction is ~0.5–0.6pp fixed per round trip while income scales with time: median P&L is +0.165pp at 7d, +0.930pp at 14d, +1.339pp at 21d, +2.372pp at 28d (win 92%). Churn is the main way this loses money — which is also why the alive gate reads a smoothed rate rather than a spot one.

Caveats: the removals above rest on a Jan–Jul 30s backtest (4 slots, 300 capital, entry bar 0.30) and on live evidence; the lending gate could not be re-tested at all for want of data. Reachability magnitude is fragile — 12.7pp rests on a 7-trade difference — though its direction held across every setting tried. Harnesses: research/bot_sim_v16.js, carry_backtest_v15.js.