01 · Build
Bind the candidate
Inspect a deterministic adapter, candidate manifest, policy revision and simulated source record.
Runnable teaching lab · local only
Claims-denial recommendation slice
Trace one bounded recommendation from an exact candidate through authorization, a simulated claims system of record, readback, telemetry, and recovery.
02 · Break
Inject the failures
Exercise missing authority, drift, stale state, duplicate requests and a lost response after commit.
03 · Prove
Read the effect back
Verify the operation in the source of record and inspect sanitized latency, cost, outcome and candidate traces.
Honest boundary
A teaching fixture, not a billing product.
The data, identities, policy and recommender are synthetic. The API records a recommendation only; it cannot approve or submit a claim. This runnable candidate (claims-denial-recommender-0.1.0) is separate from the worked-case candidate rec-0.4.3; evidence never transfers between them. Its digest demonstrates drift detection for this narrow fixture, not a complete Mission 8 release binding. It does not prove payer-rule correctness, production readiness, security certification, customer value or job readiness.
Inspect the proof
The public source and tests are part of the lesson.
Download the complete runnable lab or inspect its source files below. The original lab files are Apache-2.0 licensed and intentionally small enough to understand end to end.
Field exercises
Move from passing tests to an operating argument.
Canary
Bound the cohort, denominator, verifier, observation window, cost ceiling, specialist capacity and stop conditions.
Rollback
Preserve the previous candidate, policy, data contract and in-flight idempotency behavior.
Incident
Distinguish a lost response from an uncommitted effect using readback and the operation trace.
Transfer
Have another operator run, diagnose and recover the service without depending on the builder.