Eve library

Why one cutoff cannot work (2026-09-10)

Guide

Revision 1

On this page

Evidence for 182.C.19.08: calibrate identity and artifact thresholds by visibility, shot size, motion blur and stylization rather than one universal cutoff, per embedding model.

Fitting the operating point needs the labelled corpus 182.C.33 gathers, and that is still true — this item stays unchecked. What does not need it, and is measured here, is how far each condition moves the score. That is the part that makes a corpus usable: whatever cutoff the labels give has to be adjusted by these amounts, and under some conditions no cutoff exists at all.

Method#

One real face from the committed frame, embedded with SFace, then measured against itself under each condition with the model and framing fixed. Two different people from the same frame give the scale.

Two different people score 0.2817.

Margin below is the self-cosine minus that fixed baseline; "retained" is that margin as a fraction of the clean 0.718.

Condition self-cosine margin retained
clean 1.000 0.718 100%
lighting, −1 stop 0.960 0.679 95%
lighting, +1 stop 0.830 0.549 76%
shot size, face 14% of frame 0.927 0.645 90%
shot size, face 8% 0.852 0.570 79%
shot size, face 5.6% 0.718 0.436 61%
shot size, face 4.4% 0.451 0.169 24%
motion blur, 5 px 0.906 0.624 87%
motion blur, 11 px 0.679 0.397 55%
motion blur, 21 px 0.313 0.031 4%
motion blur, 31 px 0.251 −0.031 −4%
stylization, grayscale 0.916 0.635 88%
stylization, bilateral 0.756 0.474 66%
stylization, posterize 0.624 0.342 48%
stylization, edge-preserving 0.267 −0.014 −2%

The finding#

Two conditions put a face below the different-person baseline against itself. At 31 pixels of motion blur, and under an edge-preserving stylization, a person's own face resembles itself less than two strangers resemble each other.

No cutoff separates those cases, because there is nothing left to separate: a threshold low enough to admit the blurred self-match admits every stranger too. That is not "the threshold is not yet fitted" — it is "the decision is impossible", and assertDecidable refuses rather than returning one.

Conditions do not compensate, either. A face at 31 px of blur is undecidable however well lit it is, so the worst condition present governs.

Lighting is the mild one here — 76% retained at a full stop — which is worth saying because 182.C.19.03 measured a lighting ramp costing 0.25 of cosine. The difference is that the ramp compounded with framing change across a sequence; a stop of exposure on a fixed framing does not.

Adjusting a fitted point#

adjustForConditions pulls a fitted operating point toward the baseline by the measured retention. It refuses a base that was never fitted on labels — scaling an unfitted number by a measured factor produces a number that looks calibrated and is not, which is exactly what the inherited 0.35 gate was when SFace put two strangers at 0.422.

The case against one universal cutoff#

Five independent measurements, each with its own evidence directory, carried in code so a reader asked to accept per-condition calibration can see why:

  1. Two strangers at 0.422 under SFace against a 0.35 inherited gate — evidence/2026-09-10-cpu-identity-embedding/
  2. A lighting ramp costing 0.25 of cosine on a held identity — evidence/2026-09-10-identity-tracking/
  3. Blur leaving the detector 0.879 confident while the embedding fell to a self-cosine of 0.033 — evidence/2026-09-10-visual-integrity/
  4. One stop of exposure moving skin-tone ITA by 58° — evidence/2026-09-10-appearance-attributes/
  5. A face below the different-person baseline against itself under blur and stylization — this directory.

Per-item status this evidence supports#

Item Was Is
182.C.19.08 not-started blocked-corpus: per-condition margins measured and shipped; the operating point still needs 182.C.33