Forensic analysis report
VF-9E41C7A2
Sample · simulated result
Sample document with representative values. The detector scores are illustrative; the thresholds and detection-power figures come from calibration run full_v4.
Exhibit
- Filename
- exhibit_2216-042.jpg
- Type
- image/jpeg
- Dimensions
- 1280 × 960
- Size
- 280.9 KB
9e41c7a2d05b8ff31c26e94a7d10b5c8e2f6a03d914b7ce50a8d21f6b3e49c77Calibrated decision
0.941
Primary detector score. 0 = real, 1 = AI.
The primary score sits above the strict threshold of 0.725. Run full_v4 set that threshold to flag no more than 1% of real photographs from any source. Under Social chain conditions the detector catches 43% of AI images at this standard.
Thresholds are calibrated for the Social chain condition, never a fixed 0.5.
Primary detector · drives the verdict
Community Forensics — VF fine-tune
CVPR 2025 base · ViT-S/16 · 384px · pinned checkpoint
0.941
Supporting signals · corroboration only, never decide
Community Forensics (stock)
CVPR 2025 · ViT-S/16 · 384px
0.212
SDXL Detector
Community (HF) · Swin transformer classifier
0.716
A low supporting score does not weaken the verdict. Stock models are known to lag on 2024+ generators.
Robustness · detection power by condition
Screenshot of a WhatsApp image, which is what actually walks into a lab
Share of AI images caught at ≤5% false positives, run full_v4.
honesty gap 0.504
Headline ranking quality (AUC 0.933) minus the calibrated detection rate at your chosen standard (43%). It measures how much benchmark performance does not survive the evidentiary threshold, and we print it on every report.
Limits · read before acting on this report
- 01A flag is a calibrated statistical signal, not proof that an image was generated by AI. Check provenance and context before you act on it.
- 02The encoding fingerprint points to screenshot-style re-encoding. Sensor-level signals cannot be recovered from an image in that state.
- 03We expect the stock model to lag on 2024+ generators, and at 0.212 it does. Closing that gap is why the fine-tuned primary drives the verdict here. Supporting signals corroborate; they never decide.
- 04At the strict ≤1% false-positive standard the calibrated detection rate is 43%, so the absence of a flag would have been weak evidence that the image is authentic.
- 05Generators released after the calibration run may not be represented, and a score on a generator we have never seen can be arbitrarily wrong.
- 06This verdict applies to the exact bytes hashed above. Re-export or edit the file and you have a different exhibit.
Chain of custody
9e41c7a2d05b8ff31c26e94a7d10b5c8e2f6a03d914b7ce50a8d21f6b3e49c774c8d1a96f0b3e7255d9c02aa8f16de433b7a90c1e5f2481706bd3c5a92e8f014