Structural Predictive Gate

One question across six experiments: does the Image Structure Reader (and the ported variance-of-disagreement feature) add predictive information beyond a capable conventional pixel + regional + frequency baseline? Every test freezes its protocol, validates the estimator with planted controls, and reports a verdict or explicitly withholds one.

Results

Control battery

Resolution-floor curves

Test	n	conventional	+ structure	gain	95% CI (group boot)	build valid	disposition
1. Clustered vs dispersed synthetic, matched MAE	216	0.977	0.983	+0.005	[-0.009, +0.019]	yes	STRUCTURE DOES NOT ADD INFORMATION beyond the conventional baseline
2. Coherent vs phase-scrambled synthetic, matched spectrum	432	0.999	0.998	-0.001	[-0.004, +0.000]	yes	STRUCTURE DOES NOT ADD INFORMATION beyond the conventional baseline
3. Real vs AI, global ISR real images, 16-D ISR block	45	0.870	0.843	-0.028	[-0.148, +0.065]	no	BUILD INVALID — disposition withheld
4. Omega dispersion real vs AI, luminance-centroid reads	45	0.870	0.856	-0.014	[-0.154, +0.056]	yes	OMEGA-DISPERSION DOES NOT ADD INFORMATION beyond the conventional baseline
5. Tonal Omega (fair shot) real vs AI, tonal-band reads	45	0.870	0.960	+0.089	[+0.000, +0.133]	no	BUILD INVALID — disposition withheld
6. Steered vs default POWERED: n=859, 106 groups, real intervention	859	0.663	0.667	+0.003	[-0.016, +0.023]	yes	STRUCTURE DOES NOT ADD INFORMATION beyond the conventional baseline

Test	planted + (must recover ≥ +0.03)	planted − (must stay \|·\| < 0.02)	duplicate baseline (< 0.015)	noise max	read independence (want < 0.90)
1. Clustered vs dispersed	+0.520	-0.012	—	+0.012	—
2. Coherent vs phase-scrambled	+0.455	-0.014	—	-0.000	—
3. Real vs AI, global ISR	+0.509	-0.081	—	+0.108	—
4. Omega dispersion	+0.490	+0.004	-0.093	-0.034	0.969
5. Tonal Omega (fair shot)	+0.510	-0.049	-0.093	-0.034	0.937
6. Steered vs default	+0.490	+0.000	-0.000	-0.015	0.958

Per the ported RESOLUTION_FLOOR_PROTOCOL, the whole scale sweep is reported rather than the best rung. A genuine scale-organized signal should hold one sign and stabilize as the grid coarsens. Both Omega variants instead flip sign across rungs (tonal Omega: +0.63 at 2×2, −0.51 at 6×6), which is the signature of an unstable association, not a resolution floor.

The fair shot, and why it was not certified

Test 5 produced the only substantial positive of the whole series: conventional 0.870 → 0.960, gain +0.090 linear (+0.028 nonlinear), Brier −0.030, and the gain far exceeds the noise controls. Taken alone that reads like a discovery.

The ported gates refused to certify it, for three concrete reasons:

Independence gate FAILED (0.937, needs < 0.90). The tonal reads are still near-collinear, so Omega is not yet measuring genuine disagreement among independent reads. This is the same "fake plurality" failure that demoted Materials to toolmaker floor.
Planted-negative leaked (−0.049, needs |·| < 0.02). On a label driven purely by a conventional feature, the estimator still invented a gain of that size. At n=45 it can manufacture ±0.05, so a +0.090 is not separable from that.
Resolution curve flips sign (+0.63 at 2×2 to −0.51 at 6×6) and never matches the pre-registered direction, so there is no stable floor underneath the effect.

Had this been built without the controls ported from pathology, biomechanics and seismology, it would have been reported as "structure adds nine points of AUROC." The instrument caught its own most attractive result.

Resolved by Test 6. The controls did not just flag the result, they predicted what would happen to it. Rerun on a corpus 19x larger (n=859, 106 groups), the estimator stopped manufacturing gains (planted-negative went from -0.049 to +0.000) and the effect went with it: +0.090 collapsed to +0.003, CI [-0.016, +0.023]. The fair shot was estimator noise, and the control battery called it in advance.

Test 1 — clustered vs dispersed (matched MAE)

Test 2 — coherent vs phase-scrambled (matched spectrum)

Both classes share an identical amplitude spectrum, so every energy and frequency feature is matched by construction and only spatial coherence differs. The conventional baseline still saturated at 0.999, leaving no headroom: with the baseline at ceiling, incremental value is mathematically capped at zero. A design finding, not an ISR result.

Tests 3–5 — real Cezanne vs AI generation

Test 6 — steered vs default generation (powered, decisive)

Where this leaves it

Six experiments. No certified evidence that the structural block adds predictive information beyond conventional features. Test 6 is the decisive one, and the only test with no alibi available: the conventional baseline sits at 0.663, far from ceiling, so there was ample headroom; n=859 across 106 groups is properly powered; the estimator is provably honest on this corpus (planted-negative +0.000, duplicate -0.000, planted-positive +0.490); and the label is a real generator intervention rather than a proxy. Structure added +0.003, CI [-0.016, +0.023].

Every earlier null had a possible alibi: baseline saturation in Tests 1-2, underpowering in Tests 3-5. Test 6 removes all of them. The claim is therefore not merely unproven but falsified under conditions built to let it win.

The one honest boundary. The independence gate still fails (0.958), so what is falsified is tonal-Omega as implemented on single 2D images, not variance-of-disagreement in principle. The domains where it worked (pathology tissue, seismic arrays) drew independence from genuinely different physical observables; one grayscale image may simply not contain five things that can meaningfully disagree. That is a measurement problem, not a modelling one.

Structural Predictive Gate — labbook