Skip to main content

Review assessment and case calibration

Attempt the case first. Score 0 missing/incorrect, 1 prompted/incomplete, 2 independently correct evidence with limitations, 3 justified transfer. Every category must reach 2. A complete template or agreeable meeting cannot compensate for unsupported integrity claims.

Required categories​

  • AR1 — Claim discipline: separate observations, assumptions and inferences; connect material claims to scenario, revision, method and limits. Correct the unsupported cache acceptance.
  • AR2 — Alternatives: compare three viable options under the same constraints, preserve tradeoffs and state a defensible decision or explicit deferral. Identify the cheapest discriminating experiment.
  • AR3 — Findings: write prioritized consequences, owners and falsifiable closure criteria; distinguish verified mitigation, explicit risk acceptance and preferences. Mandatory integrity gates remain noncompensatory.
  • AR4 — Review: facilitate an independent review with a retained packet, disagreements, decision, owners and unresolved work. Solo role-play is preparation; unavailable independent review remains incomplete.
  • AR5 — Executable evidence: run a relevant architectural constraint against the project, deliberately violate it, repair it and expose collection failure. Explain what it cannot establish. Reuse earlier labs where appropriate.
  • AR6 — Lifecycle: identify three real drifts, supersede a decision while preserving its history and explain which evidence the change card invalidates. Keep a dated reopening trigger and the original “no ADR needed” reasoning exercise.

Calibration — open after attempting​

E1 establishes 90/100 = 90% good attempts for the stated sample, below the 95% target. It does not establish future reliability. E2 shows 98/100 among exported successful completions, not among all eligible attempts. If the dispatch ledger's 110 attempts all belong to the agreed population, confirmed good attempts are 98 and the observed lower bound is 98/110, about 89.1%. Even if all 10 missing attempts were good, the maximum would be 108/110, about 98.2%. The missing outcomes can change the decision; neither 98% nor 89.1% is a justified final complete-window score without resolving them. Require terminal-outcome reconciliation and the agreed deadline policy.

E3 is direct counterevidence to the revoked-access requirement. A tenant-aware key alone does not prove per-principal authorization. Re-evaluate authorization at the agreed boundary and test different tenants, principals and permission changes. E4 leaves the cache deletion path untested; a direct-read check cannot close it. E5 proves only a sequential case; concurrent/restart/conflict/expiry evidence is missing. E6 therefore cannot honestly claim closure.

Suitable blocking findings concern revoked access, deletion completion and measurement coverage. Retry evidence may also block a release involving the mutation path; explain scope. A nonblocking question could concern a measured operating-cost assumption when it is not itself a hard constraint. More than one provisional decision is defensible if its unresolved requirements and limits are explicit. None permits calling the latency scenario or integrity work complete without evidence.

The change card requires old/new request/response and rollback tests plus checks affected by serialization behavior. It does not by itself prove that the business meaning of reservation changed. Preserve the earlier decision and record what actually changed rather than rewriting its history.

Delayed transfer and repair​

After at least 48 hours, review a fresh packet with a misleading average, a stale artifact and a disputed closure. Work without the calibration. Repair the weakest category, disclose assistance and retain the first attempt. Reference calculations are author calibration; independent learner/reviewer performance is separate evidence.

Generated-proposal verification​

Use the proposal-verification clinic as the concrete AR1/AR2/AR5 rehearsal within existing review hours. The real-project fitness function and independent AR4 reviewer remain required.