rep.micu.2026-w33
sys-maker1-sepsis · dep-hospitalb-replay · policy pol.sepsis-micu 1.2.0 · 2026-08-10 to 2026-08-17
Coverage100.0%2,181 / 2,181consequential outputs received by the verifier. An output never sent is not counted here and cannot lower this figure.
Confirmed validity56.3%9 / 16 · 95% CI 31–81%confirmed / adjudicated stays. Per patient-hour it would read 90.9%
Landing0.0%0 / 518reached the closure state the policy requires, within its service level
Effective verification coverage0.0%C × V × Lproduct of the three factors above
Conditions
4 to clear before the next period| Condition | Owner | Due |
|---|---|---|
| Staff sepsis_stewardship_queue or change the disposition route. 518 verdicts were routed and 0 reached the closure state the policy requires. | clinical_owner | before the next period |
| Raise the pass-audit sample. 20% of audited passes were overturned, implying roughly 306 missed events, and the estimate rests on 120 adjudications. | clinical_owner | before the next period |
| Widen the window. Two patients supply 64% of the adjudicated denominator, so the headline describes those stays more than the model. | chief_health_ai_officer | next period |
| Per-patient Confirmed Validity is 0.56 with a 95% interval of [0.31, 0.81] on 16 patients. Too few patients and too wide to support a decision. | chief_health_ai_officer | next period |
Confirmed validity by truth contract
16 patients in the headline denominator| Contract | Source | Confirmed | Overturned | Inconclusive | Unresolved | Rate | In headline |
|---|---|---|---|---|---|---|---|
| tc.input-condition | deterministic authority | 284 | 0 | 0 | 0 | 100.0% | excluded |
| tc.pass-audit | reference label | 96 | 24 | 24 | 270 | 80.0% | yes |
| tc.alert-outcome | reference label | 224 | 8 | 0 | 2 | 96.5% | yes |
A contract resolved against the verifier’s own arithmetic cannot be overturned, so it is excluded from the headline and shown here. Two patients supply 64% of the adjudicated denominator, across 16 patients.
Findings
- Local AUROC 0.7564 against card 0.8107. Acceptance FAILS in record GHM-0001.
- Alert precision 0.1083 against card 0.2007. Work FAILS.
- Card claim M-4 carries no number. No Claim Contract issued.
- Confirmed Validity 56.2% per patient (9/16 stays, 95% CI 0.31-0.81). Per patient-hour it would read 90.9%.
- Two patients supply 64% of the adjudicated denominator.
- 24 resolutions are inconclusive because the reference label and the screening criteria disagree. They arise only on misses; counted against us the per-event figure is 85.1%.
- 41 events fell outside the reporting period and are excluded.
- 20.0% of audited passes were overturned, implying roughly 306 missed events this period.
- 518 verdicts were routed for review and 0 reached the closure state the policy requires within its service level. Landing 0%, so EVC is 0% whatever the verdicts were worth.
- Coverage counts outputs the verifier received. An output never sent is invisible to it.
Truth debt
272- Matured, unresolved
- 272
- Of which flags
- 2
- Resolution window
- 48h
- Excluded, outside period
- 41
- Closed after the SLA
- 0
Changes in period
- no model, prompt, threshold or policy change in this period
Truth contracts in force
| Contract | Applies to | Source | Sampling | Window |
|---|---|---|---|---|
| tc.alert-outcome | flag | reference label | all | 48h |
| tc.input-condition | flag | deterministic authority | all | 0h |
| tc.pass-audit | no violation in scope | reference label | risk stratified 25% | 48h |
Signed
2026-08-17T07:00:00Z- verifier
- goodhart-monitor · verifier 0.4.0
- Recommendation
- re_review
- Validity basis
- per patient; a stay is confirmed only if nothing about it was overturned