nx_frameab.nx

nx_frameab.nx -- THE FIRST REAL A/B ON THE ESTATE: two frametrace windows, three axes, one corrected verdict. WHY THIS ORGAN AND WHY NOW. nx_abstat_lib shipped the two-arm contrast arithmetic, and the acceptance panel supplies exactly ONE voting lens (a single percept scalar over a whole frame), so K=1 and the family-wise correction was a no-op with nothing to correct across. This organ is the wiring that fixes that for the PERFORMANCE subject, and it needs no new capture, no lens and no calibration corpus, because the arms ALREADY EXIST: knowledge/store/frametrace- is append-only and its last column carries EVERY INDIVIDUAL FRAME TIME of the window -- 256 replicated samples per row against a floor of 30. THE CONFOUND GUARD IS THE POINT, NOT A COURTESY. The live plane holds beach rows at budget 94, 84 and 6 ms. Budget is derived from the display refresh, so ROWS AT DIFFERENT BUDGETS CAME FROM DIFFERENT MACHINES, and contrasting them would measure the hardware while reporting a verdict about the build. This organ REFUSES such a pair by name (CONFOUNDED-BUDGET) instead of returning a confident number. * A comparison whose arms differ in a dimension you did not intend to vary is not an experiment, and it fails in the flattering direction: a faster machine reads as a better build. THREE AXES, ALL PER-FRAME, ALL FROM THE SAME WINDOW, SO ALL WITH REAL WITHIN-ARM VARIANCE: 0 frame time ms -- LOWER is better 1 over-budget indicator -- LOWER is better (1 when the frame missed its budget) 2 jank indicator -- LOWER is better (1 when the frame exceeded 2x its own window median, the shipped-industry definition nx_perf_lib already uses) K=3 puts the Bonferroni critical t at 2447 rather than 2007: three correlated views of one window would otherwise give three chances to find an improvement. Bonferroni is conservative under ANY correlation, which is exactly why it is the right correction for axes drawn from the same frames. THE MINIMUM DETECTABLE EFFECT IS DERIVED, NOT PICKED. Frame samples are INTEGER MILLISECONDS, so an effect below 1 ms is beneath the instrument's own resolution and must not be called an improvement however significant it is. That is the MDE for axis 0 and its provenance is printed. The indicator axes are already rates, so their MDE is 0 and BELOW-MDE can never fire on them -- stated here rather than left to be found. KNOWN LIMIT, NAMED RATHER THAN HIDDEN: a rate axis whose arms are BOTH constant (every frame over budget in one window, none in the other) has an undefined t statistic, so this organ reports INSUFFICIENT with the reason ZERO-VARIANCE even though the underlying shift is real and large. That is the safe direction -- abstain, never acquit -- but the right instrument for a rate is a two-proportion test, which this organ does not have. Measured live on its own fixture: a 100-percent-to-0 over-budget shift read INSUFF. COMPOSES, NEVER RE-IMPLEMENTS: fpc_row_frames for the CSV (the incumbent parser, with its own bad-input handling), pf_sample/pf_n/pf_budget/pf_pct for the ring, ab_push/ab_axis_verdict/ab_family_verdict for the statistics. This organ owns only the SELECTION of two rows and the refusal.