Inspect evidence, not just verdicts.
Compare the displayed fields, ground-truth tamper region, structured prediction, parser state, and recorded latency.
No tamper detected
none confidence 100.0%
Precomputed label-derived Demo Stub output for interface walkthrough only.
0.0 ms · parser valid first pass
Trace one document through the evaluation pipeline.
Use the selected synthetic sample to replay how an image becomes a structured, scored prediction. This walkthrough uses the existing precomputed gallery output and clearly separates inference from ground-truth scoring.
- 1Load the document image
Read only the selected card pixels; no label or manifest enters inference.
- 2Prepare visual evidence
Normalize the image and preserve text regions, degradation cues, and possible tamper boundaries.
- 3Run the configured adapter
Transcribe fields and estimate the tamper verdict, type, confidence, and localization.
- 4Validate structured output
Check the JSON contract and record malformed or missing values as failures.
- 5Compare with held-back labels
Only after prediction, score extraction, verdict, type, confidence, and bounding-box overlap.
- 6Surface the research finding
Show whether the error came from perception, reasoning, formatting, calibration, or localization.