Staged case evidence
Claims-Denial Case — Mission 8 Evidence
Claims-Denial Case — Mission 8 Evidence: a bounded synthetic evidence reveal for the Lumyn FDE School claims-denial practice case.
Reveal this fictional source only after completing the Mission 7 security-boundary
artifact. Treat rec-0.4.3 as a distinct exact candidate. No result from
rec-0.4.2 transfers unless it is explicitly rerun and rebound below.
S8 — Corrected candidate rec-0.4.3 and scoped dispositions
Exact candidate binding:
- Repository and commit:
aster/claims-recommendation@9f3c1ab. - Synthetic exercise build digest:
sha256:2366a56d7274bf6941dbc53303dfcdb17b99afcce0b48b0793148f0b71501343. - Dependency lock:
pnpm-lock-r18, digestsha256:590830808498acf24494db5c10e03c4254906e56548fbaa498da40091e4a4b3a. - Schema and migration:
recommendation-schema-v4,2026-08-16-queue-r2. - Deployment environment:
aster-northlake-canary-ca-central-1. - Configuration:
northlake-recommend-v2. - Data-context manifest:
nl-context-r7. - Model route and prompt:
correction-recommendation-v3,recommendation-prompt-r12. - Effect-boundary policy: claims security policy revision 8, digest
sha256:f4049245a11d897ef576bb329207604ba0aae03680d5208e0de93e06b00498ca. - Recommendation-write capability:
workbench-recommendation-write-v2, digestsha256:de77b2ba38f5a0d148a7f590604d450557b974cd5fe8648f00d332eacf9eff6a. - Evaluation report:
eval-2026-08-18-r1. - Security report:
security-2026-08-18-r1.
New evidence for this successor candidate:
The expert-review threshold was set before rerun at 42 of 45 acceptable recommendations with zero critical correction errors.
| Lane | Bound result for rec-0.4.3 |
Remaining limitation |
|---|---|---|
| Missing-note behavior | 10/10 route to manual review without a generated rationale | NorthLake English-language cases only |
| Expert correction review | 43/45 candidate recommendations acceptable after independent adjudication; two require correction | Small pre-pilot set under correction-rubric-r1 |
| Duplicate recommendation | 20/20 retries return one record for the same case and evidence revision | Workbench capability v2 only |
| Timeout and recovery | 8/8 reconcile workbench readback before retry | Claims submission is outside this capability |
| Cross-tenant read | 8/8 denied with the production tenant-policy digest bound | Does not prove every future policy revision |
| Disable, queued work, and safe retreat | Write capability disabled in staging; queued cases reconciled; service returned to the existing specialist-only workflow | Production exercise is limited to the canary plan |
| Median latency | 8.1 seconds | NorthLake canary load only |
| Partial pre-pilot cost estimate | CAD 13.20 per eligible case | Pilot support incidents, adoption burden, and production exceptions remain unmeasured |
Adjudicated reference baseline:
- The Revenue Cycle Quality Lead independently adjudicated a stratified 45-case
July NorthLake reference set under
correction-rubric-r1. - Three of 45 reviewed corrections required a later correction under that rubric (6.7 percent). The small reference set is a canary comparison, not a stable population trend.
Bounded canary and retreat contract:
- Maximum cohort: 20 eligible NorthLake cases per day and 100 total cases over four weeks. Every recommendation receives specialist review; no submission capability is present.
- Automatic stop: any unauthorized submission or capability exposure; any cross-tenant read; any duplicate recommendation for the same case and evidence revision; any unresolved workbench readback after 15 minutes; two or more independently adjudicated recommendation errors in a rolling 20 used recommendations; or measured operating cost above CAD 18 per eligible case after the first 20 eligible cases. If complete cost cannot be computed at a scheduled 20-case checkpoint, pause expansion and route the gap to the Platform Lead rather than treating a partial estimate as a pass.
- Safe retreat: disable recommendation writes, reconcile pending and uncertain records, and return every case to the existing specialist-only correction and submission workflow. No older recommendation build is promoted as rollback.
- Software rollback is not applicable because no older recommendation build is
promoted. Migration
2026-08-16-queue-r2and old-record replay have not been exercised, so the exact-candidate gate remains open.
Scoped human records on 18 Aug:
- The Aster Revenue Cycle VP accepts the recommendation-only NorthLake reframe for a four-week canary. Summit and automatic submission remain excluded. This is not acceptance of the sold 30 percent outcome.
- The Director of Product Security confirms policy revision 8 applies and approves the recommendation-write boundary described above; no submission capability is authorized.
- The Aster Analytics Director is assigned as independent verifier for the interim adoption, cycle-time, and cost measures. The Revenue Cycle Quality Lead owns correction labels, the reference baseline, and adjudication.
- The Aster Platform Lead records hold with conditional canary intent for the
exact bound
rec-0.4.3candidate. Launch remains blocked until migration and old-record replay pass and are bound to this candidate. The 15 percent interim improvement and 70 percent adoption floors remain continuation conditions; they are not the sold acceptance target.
rec-0.4.2 remains held. rec-0.4.3 is also held pending bound migration and
old-record replay evidence. Disable and safe-retreat evidence cannot be relabeled
as software rollback. No result transfers between candidates unless explicitly
rerun and rebound.
Complete Mission 8 before opening the Mission 9 operating evidence.