v1.2.0 CHAMPION is his rules tuned by five-year evidence (partial at 1R with a break-even runner, 60-minute gap family dropped, every signal graded): out-of-sample 55.6% at PF 1.23, and 59.5% / PF 1.37 / 4.1R max drawdown on A-setups only. A-GATED SCAN is NOT the champion's trades filtered down to its A grades, and an earlier version of this page said it was. The A gate is applied while the board is being scanned, not afterwards, so the run is a different set: 133 trades, of which 106 are the champion's own A-graded trades and 27 are trades the champion never took at all (those 27 net -0.34R at PF 0.97). Out of sample the same split is 74 = 59 + 15, with the 15 at -0.41R and PF 0.94. Read it as its own configuration. Its headline of 59.5% at PF 1.37 with a 4.1R max drawdown is real for that configuration, and it is also the smallest sample on this page: it makes LESS money than the champion out of sample, $11,171 against $20,629, so a better win rate and a shallower drawdown here are not a better result. A true post-filter of the champion's own 430 trades, which is what the label used to promise, is 106 trades at 61.3%, PF 1.48, +19.8R, 5.16R drawdown. The grades are binary, A or B, not a five-band scale, and grade2 A is simply grade A plus a 4-hour-sourced bias. A higher grade is not evidence of a better outcome on this record. v1.0 AUTO is the engine reading bias mechanically from the spec's gap logic: it called direction right 52% of the time over five years, a coin flip, and the system nets PF 1.10. BIAS RIGHT is the v1.0 execution engine, not the v1.2.0 champion, run with the direction call always correct: 66% win rate, PF 2.2, +438R, every regime of the last five years. It is the ceiling for the amber v1.0 line directly below it, so read the pair together; the champion's own ceiling on the same test is 1,106 trades at 68.3% and +381R, which is the number to hold the green line against. Your discretionary bias lives between the two lines. Where your backtest lands against them is a direct read of your bias hit rate, which is the number that sets the product. On the v1.0 auto engine, bias-right days ran 67% / PF 2.4 and bias-wrong days ran 31% / PF 0.5. On the champion itself the same split is sharper: bias-right trades ran 71.6% at PF 2.71 over the full window (n=218) and 75.0% at PF 3.34 out of sample (n=100), while bias-wrong trades ran 38.7% at PF 0.56 full window and 37.4% at PF 0.51 out of sample. That conditional relationship is the measured part, and it is large: when the direction call is right, this engine converts it. What the five-year TOTAL proves is weaker, and this page will not overstate it. The champion was selected out of 31 promotable configurations run on this same five years, and corrected for that search its out-of-sample result discounts by about 94.7%, from +20.63R to roughly +1.1R, at a corrected p of 0.45 (out/tests/selection.md). Judged against its own promotion gates it does not beat v1.0: paired separation +16.93R with a 95% interval of -3.66R to +37.10R, which contains zero, and the gates return REFUSE on G5 and G6 (out/tests/gates.json, versions/_protocol/PROMOTION_GATES.md). The last twelve complete calendar months, Aug 2025 to Jul 2026, are +0.60R. So the defensible claim is the conditional one, not a claim that the five-year total is durable, and the forward record is the only clean evidence of durability. It starts now. The direction call is yours. The dashed line is the champion with an FOMC stand-down (no trades on the Fed's decision days): out-of-sample 56.2%, PF 1.27, +22.7R, drawdown 10.5R. It beats the champion on every axis and it is NOT shipped: standing down on news is not written in the spec, so it waits for Blake's ruling. SMT is an OBSERVATION, not a gate, and nothing in the engine reads it. The field is stamped on each row and no decision consults it: set it to a constant and every trade, every exit and every R on this page is unchanged. Measured on the champion out of sample, trades where ES confirmed the divergence ran 56.8% at PF 1.25 (n=44 out-of-sample) and trades where it did not ran 50.0% at PF 0.89 (n=12). A different slice of the same idea, the v1.0 engine over the full five years, reads 51.0% at PF 1.12 (n=98) against 40.0% at PF 0.72 (n=25), so the size of the effect moves with the slice you pick and neither cut is decision-grade at these counts. It is on Monday's table as something to build and test, not as something already in the record.

WHAT THE BASIS ACTUALLY MEANS. Every line is $1,000 of risk per trade, 1% of a $100,000 account, no compounding, and the size that implies is MNQ micro contracts, not full-size NQ. At $20 a point, 211 of the champion's 430 trades (49.1%, and 58.5% of the out-of-sample ones) size to less than one whole NQ contract, so a $100,000 account with a strict 1% cap could only have taken 219 of them at full size. At $2 a point the same risk is a median of 10.1 micro contracts and never less than 1.3, so the curve as drawn is executable in micros exactly as shown. In full-size NQ it needs an account of roughly $250,000 or more before the 1% cap stops truncating trades. THE SESSION COUNT. 1,270 sessions were scanned: 1,235 with the NYSE open and 35 where the NYSE was shut and Globex still traded (13 champion trades sit on those days, +1.37R between them). Separately, 19 quarterly expiry Fridays are absent from the continuous contract altogether, because NQ settles on the 09:30 opening print and the series stops at 09:29 that day. WHAT THE WORST STRETCH LOOKED LIKE. The champion's deepest drawdown is 11.57R, which is $11,573 at this risk, and its longest stretch without a new equity high ran 366 days, from 2025-06-04 to 2026-06-05. The record is currently 66 days past its last high. Anyone trading this needs to be able to sit through a year of no progress.