The hypothesis — fragility equals realness

The conceptual core of CLR, which I formalized between v1.2 and v1.3, curated from my original research notebooks and patent disclosure notes.

A Reconstruction Hypothesis

If something can be destroyed and never perfectly brought back, that suggests it was real. If something can be destroyed and then reconstructed, that suggests it was generated. Fragility equals realness. Recoverability equals simulation.

A real photograph is the output of an irreversible physical chain:

scene → optics (PSF) → sensor (Bayer, PRNU, shot noise, dark current) → ADC → image

Every stage imprints information bound to a specific piece of silicon at a specific moment — noise that is spatially incoherent, physically motivated by quantum mechanics and solid-state physics, and bound to the pixel grid. A generated image is the output of a different chain:

Gaussian noise z → denoising trajectory (T steps of a learned score function) → image

a many-to-one mapping whose "noise" is structured by the model's internals, not by pixel positions. Authenticity isn't a quality you can synthesize; it's a residue of having been caused.

The probe: a geometric round-trip

The warp probe operationalizes this. Take a small tile T, apply a small random perspective homography H at strength s, then invert it:

R(s) = | T − H⁻¹( H( T ; s ) ; s ) |

The double resampling scrambles grid-bound micro-texture proportionally to s. Real PRNU is destroyed at a characteristic rate; synthetic texture — which was never bound to a grid by physics — responds differently.

I gave the measurement a deliberate three-level structure, and getting it right mattered for the filing:

• Trials (a fixed bundle per strength): many random homographies are round-tripped and their residual maps averaged — turning directional perturbations into an approximately isotropic residual estimate at that strength. The trials are never compared to each other; they are variance reduction.
• Differential (across strengths, within a region): the fragility rate FR = (R(s_high) − R(s_low)) / (s_high − s_low), computed over a low/high strength pair chosen well inside the perceptually-invisible range. The discriminative quantity is ∂R/∂s — the rate of destruction, which cancels baseline content and corrects the empirically false "generated images are perfectly undone" framing of absolute-residual approaches.
• Cross-region (across two distant regions, at one strength): correlation and energy-ratio comparisons between the residual maps of geometrically paired tiles — top-left against bottom-right. If both patches produce residuals that cohere under the same projective transform, they inhabit one physical coordinate system; the subject was only possible when captured in the real world. The invariant being learned is the physics, not the pattern.

The slope, not the score

The hypothesis score and the local (per-tile XGBoost) score are two orthogonal probes — the outward and inward views. Neither alone suffices; their relationship is the signal:

• slope = hypothesis − local. Sign is the verdict; magnitude is confidence. Absolute scores sitting below 0.5 can still classify correctly via direction.
• R(W) = W × slope accumulates the divergence: a rising curve is physical substrate asserting itself; a falling curve is its absence. The geometry of the divergence is the verdict — threshold crossing is secondary.

The slope made spatial: the hypothesis-minus-local delta per tile. The real photograph (left) resolves to a positive slope; the generated portrait (right) lights up generation hotspots and falls negative. Where the delta is red is where your eye would pause — the hypothesis and human perception converge because both test physical plausibility.

Why this framing is durable

The smoothing inherent to generative synthesis is intrinsic to optimization-toward-a-target. More data and parameters make the generative manifold smoother, not more physical — so a detector grounded in this property doesn't decay with generator scale the way fingerprint-matchers do. FLUX, Midjourney, SD — none of them simulate PRNU. The hypothesis doesn't detect models; it detects the absence of a sensor.

Two boundary conditions I could see from the start, and want to state honestly:

• Compression erodes the substrate. Heavy JPEG, DeepFusion-style ML post-processing all weaken PRNU. Real images near slope = 0 are telling you their physical history has been partially erased. The architecture leans on relative quantities — ratios and differentials, which are mathematically invariant to uniform rescaling — precisely to survive re-encoding.
• Biometric-free by construction. No face, body, or identity model anywhere in the pipeline; salient regions are located by gradient energy alone. That's a privacy posture (GDPR/CCPA/BIPA never enter the frame) and a coverage property — the probe works on scenes, objects, and documents, not just people.

In chapter 5 I test the strongest version of this claim — architecture-agnosticity — and break it. But within the camera-photo distribution I built it on, the physics holds, and the next chapter is the empirical climb where I proved it feature by feature.