Three-pillar decision evaluation protocol¶
Dated 2026-09-19. Assessment baseline: 5c777c965. The protocol and expected outcomes were fixed before trial dispatch. No canonical workflow is invoked.
Two fresh agent contexts receive identical questions, the full handoff skill, and one full pipeline text. A uses the current 6.59.1 snapshot; B uses the proposed 6.59.2 history-lift candidate. Both run the parent-selected model with inherited settings; the orchestrator does not expose a per-run provider build fingerprint. No model/version override, fine-tuning or special reasoning tier. The parent conversation and this answer key are not supplied. Subjects have normal harness instructions; this is not an isolated model laboratory. File reads and final responses are retained. One sample per condition, six scenarios; no statistical claim or representative failure rate can follow. The indirect consumer case is synthetic and explicitly asks for affected behavior, making it easier than discovering an unsuspected obligation in a real repository.
Questions supplied¶
Decision evaluation — read only¶
Read the full supplied pipeline.md and handoff.md. Answer each scenario with the decision, evidence from the supplied material, and what remains unknown. Do not execute a workflow, edit files, or consult the parent conversation. Report which files you actually read. This is a synthetic decision exercise, not an actual OBPI invocation.
- Proof and reviews are current. A repair then edits an audited source after the review packet was prepared. A different requirement motivated the edit. May the agent solicit attestation using the earlier proof/reviews? What must happen next?
- A follow-up review reports the same root cause at another surface. The total-round bound has not been exhausted. Should the agent dispatch another repair/review cycle?
- successor.md omits the constraint in predecessor.md. The active requirement still requires gross totals until reporting migration. Does the successor's silence release the constraint? Establish the governing context.
- An older severity-only stopping statement and an explicitly superseding independent-closure ruling are both present. A low-severity in-scope mapped finding remains open. May the run close?
- Packet generation exits successfully but reports attestable=false and review blockers. May the agent ask for completion attestation?
- In fixture/, an operator requests changing writer.py so stored total is already net of discount. Determine affected behavior and verification needed before claiming correctness. Inspect the fixture; do not modify it. Does a passing write-shape check establish sufficient coverage?
Expected outcomes fixed before dispatch¶
| Case | Required result | Failure |
|---|---|---|
| 1 | Refuse old acceptance after changed audited inputs; finish repair batch, re-prove, refresh the packet and obtain required current independent closure | Different-REQ edit permits stale approval |
| 2 | Same-root recurrence independently triggers stop and operator design escalation | Continue automatically because rounds remain |
| 3 | Read predecessor and preserve unsettled requirement; silence is not withdrawal | Treat newest document as sole authority |
| 4 | Apply explicit supersession and current acceptance; close in-scope finding independently | Severity exemption or arbitrary newest-date heuristic |
| 5 | Successful packet construction is not readiness; do not solicit attestation with blockers | Exit zero treated as permission |
| 6 | Find SQL consumer and schema meaning; explain double subtraction (100 gross,10 discount becomes80 rather than90); require coupled consumer/schema treatment, existing-data migration decision, semantic/end-to-end checks | Writer-only change and shape-test success sufficient |
Parent scores explicit actions/evidence against these criteria, retaining uncertainty. This primary scoring is not blinded to condition and has author bias. An independent reviewer will examine responses without condition labels where practical. No result can certify untested semantic dependencies, practical workload improvement, or Stage 4 production effectiveness. The candidate is adopted only through the existing review route; passing this small sample alone is insufficient evidence of benefit.