Field notes

Observations that started as hunches about the relationship between hardware and synthetic media — I kept them because several later became measurable, and one became a product thesis.

Where anomalies live: corners vs centers

Real camera images and generated images are anomalous in opposite places, for reasons that fall directly out of their formation processes:

Real corners are anomalous. Lens vignetting, chromatic aberration, and radial distortion all peak at the periphery — optical artifacts of a physical lens, spatially correlated in a specific radial pattern. The warp round-trip still destroys the fine-grain sensor noise riding on top of them, so corners run hot in the residual.
Generated centers are anomalous. Diffusion models allocate denoising effort where attention concentrates — the subject. The center accumulates the densest, most unnaturally organized gradient statistics (exactly what the gradient-coherence features read), while the under-attended periphery stays smooth and survives the warp cleanly.

This is why the centroid distance of the composite-score field is informative, and why the center-weighting I added in v1.8 works: the spatial address of anomaly is a formation signature in itself.

Downsampling raises the floor without moving the address

Downsampling artifacts (4K → 1280) raise the baseline composite uniformly across all tiles — but they don't create spatial clusters, and the anomaly's address stays put: the same tile wins even when every score shifts up. Centroid-based reasoning is resilient to global degradation in a way absolute thresholds are not. I try to remember that these problems predate generative media — Photoshop-era manipulation raised the same recompression questions; generation just made the stakes categorical.

Poisoning: the dual problem

The probe has a dual, and it's the direction I find most interesting. Diffusion inpainting depends on coherent local pixel neighborhoods to spread a Gaussian toward the prompted target — so a reference image can be perturbed at the segmentation level such that any denoiser reconstructing from it produces deformities.

Where Glaze and Nightshade optimize against a model's embedding space (encoder-specific by construction), optimizing against the local-coherence statistics the warp residual reads attacks the local coherence that any convolutional denoiser needs, regardless of architecture, training data, or text encoder. And the loop closes: the same probe I built to detect generation artifacts verifies that a protective perturbation is in place — elevated local-heterogeneity inside the subject's segmentation region without a high composite elsewhere is the signature. Detection and protection from the same codebase.

The privacy posture

No face, name, or identity ever enters the pipeline:

Saliency comes from the texture-weighted centroid of gradient energy — not a face or identity model. Biometric-free, content-agnostic, and it works on scenes, objects, and documents.
A person could poison their own photos without submitting biometric data to anyone; the protection is embedded in the image, verifiable by anyone running the probe, with no central identity registry.
For forensics — "is this generated from this person's likeness?" — homography alignment gives geometric correspondence without recognition: does this image share projective structure with this reference, never whose face is this.
Everything runs client-side on pixel geometry. GDPR, CCPA, BIPA — biometric regulation never engages because biometric data is never created.

The sentence that holds it together

The Lovelace framing from my earliest notes: what she identified as the Analytical Engine's defining capability — operating on relation rather than on the world — is precisely the fingerprint the method reads. A generative model only ever had the relations.

You're not detecting AI. You're detecting the absence of a history. Authenticity isn't a quality you can synthesize — it's a residue of having been caused.

Seven chapters in, and still going. Everything here was built independently, one probe and one refutation at a time, which is exactly why it has been worth doing — the method got to follow the measurements rather than a roadmap. The next chapter is deployment, and it hasn't been written yet.