HMNPROOF
Live model status

The engine gets sharper with every scan.

HMNPROOF runs a self-training signal weighter on top of its detector stack. Every scan and every "this seems wrong" flag updates per-signal weights, so verdicts that were borderline yesterday become confident today.

Scans processed
41
Signals tracked
28
Model version
v1.0.1
Accuracy estimate
51.1%
Feedback signals
16
Last trained
Aug 19, 2026, 02:44:13 PM

Top learned signals

Updated in real time

Each bar shows the learned weight of a signal — negative pushes a verdict toward authentic, positive pushes toward AI / manipulated.

File uploadsource
n=138 · weight -0.29
Reverse-image links generatedneutral
n=136 · weight -0.30
Byte forensics: high entropy (≥7.5)neutral
n=130 · weight -0.32
Image media typemedia
n=129 · weight -0.31
EXIF: missing or strippedrisk
n=108 · weight -0.08
No C2PA manifestneutral
n=91 · weight -0.42
Face(s) present in mediamedia
n=80 · weight -0.20
AI vision: low likelihoodauthentic
n=71 · weight -0.73
Face forensics: clean face regionauthentic
n=71 · weight -0.23
Camera-native filename / RAWauthentic
n=52 · weight -0.68
AI vision: high likelihoodrisk
n=37 · weight 0.47
Embedded provenance declares AI generationrisk
n=33 · weight -0.05

How it learns

  1. 1. Observe. Every scan extracts signals from every layer of the stack — source (host, keywords, scheme), metadata (EXIF present, C2PA manifest), byte forensics (entropy band, JPEG DQT count, double-compression), AI vision verdict band, face-region deepfake band, opt-in identity match outcome, reverse-image availability, and NCII flags.
  2. 2. Track. The engine records how often each signal accompanied a high, moderate, or low verdict.
  3. 3. Correct. Every "this seems wrong" flag tells the engine when it was off — false positives push the signal toward authentic, false negatives push it toward AI.
  4. 4. Re-weight. Weights are smoothed by sample size, so a brand-new signal can't dominate until it has earned the data behind it.
  5. 5. Apply. Next scan, those weights nudge the score — capped so the base detectors always have the final say on extreme verdicts.