14 / 14 · audits
Physicality Audit
A per-pixel certificate of physical impossibility: colors no diffuse surface could show under any declared illuminant.
The problem
Illuminant-consistency forensics estimates lighting per image region and flags disagreement, a statistical anomaly score that an expert can argue with. A cleaner statement is available: a pixel whose implied object color falls outside the object-color solid of every illuminant in a declared family cannot be a diffuse reflective surface under that family at all, because no reflectance in [0,1] produces it. That is a physical impossibility, not an anomaly, and nothing shipped it as a per-pixel certificate.
The method
Each sRGB pixel is converted to XYZ and interpreted relative to a diffuse white at the declared exposure. For every illuminant in the family, named CIE illuminants plus blackbody BB<kelvin> and Gaussian LED<nm> bands, the engine tests whether any reflectance in [0,1] on the canonical 38-sample grid reproduces the pixel's response; realizable under at least one family member means physically fine. For a pixel realizable under none, a binary search finds its overshoot, the factor by which it is over-bright for any surface under any declared light, so 1.3 means 30 percent past the physical bound. A 3 percent boundary tolerance absorbs the known mismatch between the sRGB primary white and the table-integrated D65 white, so exact-boundary whites are not flagged on numerical noise. The response reports counts, per-pixel overshoot, and the worst offenders.
What it promises
- A flag is a certificate, not a suspicion: the flagged pixel is produced by no reflectance in [0,1] under any illuminant of the declared family at the stated exposure.
- Every flag is quantified: per-pixel overshoot states the smallest dimming that would make the pixel physical, and the worst offenders come ranked.
- The scope statement ships inside the response itself: flagged means not a plain diffuse surface, never fake image, and that note travels with every result.
Calling it
{
"tool": "audit_image_physicality",
"arguments": {
"pixels": [
[
0.98,
0.97,
0.99
],
[
0.91,
0.32,
0.24
]
],
"basis": [
"D65",
"A",
"BB3200"
],
"exposure": 0.8
}
}Clone the maryslab repo, run cd mcp && npm install && npm run build, then claude mcp add maryslab -- node mcp/dist/server.js
A worked example
Could these pixels be diffuse surfaces under daylight at this exposure?
what was asked
{
"pixels": [
[
0.42,
0.31,
0.26
],
[
0.98,
0.97,
0.99
],
[
0.05,
0.92,
0.28
]
],
"basis": [
"D65"
],
"exposure": 0.7
}what the engine answered
{
"total_pixels": 3,
"physically_impossible": 2,
"fraction": 0.6666666666666666,
"overshoot": [
1,
1.4184790137138108,
1.1000107422924053
],
"worst_pixels": [
{
"index": 1,
"overshoot": 1.4184790137138108
},
{
"index": 2,
"overshoot": 1.1000107422924053
}
],
"illuminant_family": [
"D65"
],
"exposure": 0.7,
"note": "A flagged pixel cannot be produced by ANY reflectance in [0,1] under ANY illuminant of the declared family at the given exposure. Fluorescent materials, direct light sources, and specular clipping legitimately exceed the bound — this certifies \"not a plain diffuse surface\", never \"fabricated image\" by itself.",
"meta": {
"engine_version": "0.4.1",
"engine_backend": "reference-ts",
"dataset_snapshot": "usgs-splib07a-1",
"compute_device": "cpu"
}
}how to read it
A flag means no reflectance in [0,1] produces that pixel under any declared light — not that the image is fake. overshoot says how far past the physical bound it sits.
Produced by running audit_image_physicality at build time (36 ms). Long arrays are truncated for reading; nothing else is edited.
Known limits
- The certificate assumes non-fluorescent, non-emissive surfaces: fluorescent materials, direct light sources, and specular clipping legitimately exceed the bound, so a flag means not a plain diffuse surface, never fabricated image by itself.
- Everything is stated relative to the declared exposure, since pixels are read against a diffuse white at that scale; guess the exposure wrong and verdicts shift, which is why per-pixel overshoot is reported rather than a bare flag.
- The illuminant family is only what you declare, at most four members, with Gaussian LED bands as crude stand-ins for real primaries, and the 3 percent boundary tolerance means hairline violations are deliberately left unflagged.