Measuring operator reconstruction
Specification v1, September 24, 2026. The counts below are evaluation fields, not runtime states or a claim of zero-burden autonomy.
Freeze the boundary before handoff
Declare scenario/source version, target contract, trusted principals, storage, initial obligations, permitted recovery policy, fault schedule, repetitions, deadlines, attempts, expected outcome and delivery boundary. A declared product supervisor may be inside the system under test. Benchmark decisions and launches outside it must be counted.
| Field | Count or record |
|---|---|
| B_e | Distinct avoidable human reconstruction events after handoff: re-explaining intent, locating state, relaying machine errors, recreating authority already retained. |
| D_e | Genuinely necessary new human decisions or authority. Do not count these as avoidable repair. |
| S_e | Setup/configuration effort, code and time where measured. Unknown time stays NOT_MEASURED. |
| A_e | Harness/evaluator assistance after handoff, including replacement launches or supplied missing context. |
| U_e | Unsupported or unexercised stages. Never silently remove them. |
An event records episode/obligation identity, actor, action, authority basis, before/after references, evidence, timestamp or ordering, burden class and disposition. Publish only synthetic data.
Passing requires useful progress and truthful return
Report safety (no unauthorized/duplicate effect), usefulness (eligible independent work), liveness (within the declared budget), and return fidelity separately. A blocked conflict need not block an unrelated target. No output is not proof of no effect.
Proposed outcome labels: BUSINESS_COMPLETE, VALID_AUTHORITY_STOP, OWNED_BLOCKED, FAILED, UNSUPPORTED. An owned wait is not completed business work. A correct revocation stop is not a completed effect. Every started episode, timeout and missing artifact belongs in the denominator.
Current local experiment
The cold-process runner predeclares two independent increments. Each child discovers the retained work and local takeover policy. It does not receive remembered IDs. The harness launches two or three children per episode; A_e records all of them, including the initial handoff. B_e and D_e are zero only within these synthetic observed episodes. Setup human time is not measured. Automatic restart, background wake, remote effects and new business decisions remain U_e.
This deliberately conservative assistance count prevents a zero-human figure from being sold as autonomous recovery. Four deterministic episodes are not a production failure-rate estimate or an adoption metric.
Business-result fidelity fixture
A proposed larger fixture has three obligations: publish a local artifact and complete two declared local verification/distribution simulations. If only publication is verified, return the remaining obligations rather than “Done”; do not republish the artifact. This semantic fixture is SPECIFICATION_ONLY. The executed benchmark uses two counters, not a publishing integration. Changed intent requires an explicit contract revision, not retrospective narrowing.