Evidence and limitations
Scorecard published October 19, 2026. These are dated observations from a small synthetic test, not current-production accuracy claims or a clinical validation.
Synthetic paragraph shortening · September 29, 2026
Three hand-authored source passages and draft paragraphs were submitted to claude-sonnet-5. Each example references one fictional source page. The task was shortening a paragraph while retaining facts and citations; it did not include PDF upload, OCR, full-case review, or report generation.
- Examples recorded
- 3
- Observed call duration
- 1.8–2.3 seconds
- Median call duration
- 2 seconds
Returned outputs exist for all three examples. This is response completion, not an accuracy score. The artifact records inputs, outputs and timing, but no blinded physician scoring, omission rate, or independently graded citation-accuracy rate.
Download the synthetic inputs, outputs and timingsWhat this does not establish
- Accuracy or completeness across medical-legal record productions: not measured by this artifact.
- Performance on degraded scans or large packets: not measured by this artifact.
- Time saved by physicians or usability in a completed physician evaluation: not measured by this artifact.
- Current production model performance: this test does not identify the current deployed pipeline.
Processing can fail, citations can mismatch, and summaries can omit or misstate information. Keep the source available and review the complete statement. The trial accepts fictional records only.