read-only demo — this instance accepts no writes of any kind. Adjudications, adoptions, and promotions are all refused (403). Real HITL happens only on a locally writable workbench.
Extraction correctness is untrustworthy; support is verifiable. You decide — the system shows evidence, not verdicts.
Inputs are outside the calibration set (§12): calibration numbers (4.2×, 78%) do not directly apply. Per-document mechanical checks (binding, gates, freeze, adjudication) need no calibration.
25 / 30slots reviewed
0corrections
25decisions total
Accept: 13 Confirm absent: 12 N/A: 0 Reject: 0 Correct: 0 Abstain: 0
Corrections (deliverable)
—
Residual risk: the reviewed queue covers most measured deviations (calibration 4.2×, held-out 3.04× lift), but errors outside the reviewed fraction remain — the residual is NOT zero. See docs/HELDOUT.md for the pre-registered numbers.