Michael Oakes
AI operations case study · November 2026

Project Conductor: defining what “done” means

A historical study of AI coordination: an incomplete run record, a later accepted reporting-control run, and task outcomes that remained unverified.

Coordination designSource reviewHistorical evidence

The problem

When several assistants work on a project, a status update can travel farther than the evidence behind it. A scheduler starts a process, a process exits successfully, and the next report describes the work as complete. Conductor’s central question was how to keep those events separate from proof that the intended work happened. [C1]

Role and approach

Michael retained the project’s scope, adoption and publication decisions. The project charter assigns AI assistants distinct intake, building and review roles. This is a human-directed, AI-assisted effort; the reviewed records do not establish that Michael personally authored every implementation detail. [C1] [C4]

The documented design centers on a registry: owner, state, artifacts, blockers and next action. Bounded work packets connect a request to an expected output. The packet-state rules distinguish routing a packet from executing its contents and require an explicit option before moving it. These are documented controls, not proof that a complete orchestration service operated in production. [C1]

Three different kinds of evidence

  1. LaunchDid the process start and exit?
  2. ObservationWhat did that exact run inspect?
  3. EffectDid the intended work happen?

The included effect-check test source makes the distinction concrete. Undeclared output evidence should produce a warning; declared but missing or stale evidence should produce an error. A fresh artifact satisfies the freshness check. That does not establish the artifact’s correctness or the completion of its underlying task. [C2]

The useful failure: October 2

The October 10 review of the October 2 scheduled run found one validated telemetry record reporting success. However, the acceptance criteria also required the report and logs belonging to that run. By the evaluation date, the rolling report and standard-output log reflected a later run. The retained evidence could not reconstruct the October 2 report or its per-task observations. The composite execution gate therefore failed. [C3]

Coverage also remained incomplete. A grouped warning and omitted entries did not establish eight explicit, supported source-access classifications. The evaluator recorded coverage as “not demonstrated”; every coordinated task effect remained unverified. That finding identifies an evidence-retention and observation gap. It does not prove that the tasks failed to run. [C3]

What the source review supports

Claim boundaries
EvidenceSupported conclusionLimit
Charter, schemas and packet rulesCoordination design and inspectable artifacts exist.No production-orchestration claim.
Effect-check test sourceAssertions distinguish missing, fresh and stale evidence.The packet omits the monitor implementation; tests were not rerun.
Retained evaluation outputThe historical acceptance review recorded FAIL and incomplete coverage.No fresh runtime inspection or task-effect verification.
October installation and matched scheduled-run evidenceEleven-file installation recorded; one reporting-control run accepted.All eight task effects remained unverifiable; no current-operation claim.

A later, narrower success: October 30

The October 2 evidence gap remains part of the record. A separately preserved October 23 installation receipt documents eleven verified file changes: the monitor, its configuration and schema, plus eight task receipts. The original scheduling state was restored; installation itself did not establish task execution. This later evidence is separate from the earlier committed packet. [C5]

For October 30, a retained scheduled-run record matches the output log, and the report’s generation time falls inside that run’s start and end times. The report records zero errors and eight warnings. A separate acceptance check rechecked the eleven file identities and accepted the event for receipt controls only. This supports one ordinary scheduled execution of the reporting mechanism—not merely a synthetic test. [C5]

All eight underlying task outcomes remained unverifiable. Their receipts lacked task-owned action evidence and authoritative destination readback. The monitor successfully preserved those unknowns as warnings; it did not establish successful completion of the coordinated work. The frozen report, ledger and logs make this later event inspectable without using a subsequently overwritten report to reconstruct it. No current runtime or provider status is claimed. [C5]

November 3: the local baseline was committed

The November 3 closeout incorporates the accepted October 30 receipt-control layer into local commit dc77ad29117c4f022e15524f96df7dbb76bc86c0 on main. Its 21 changed paths comprise 17 receipt-control files and four status documents. The complete committed tree matched the isolated candidate that underwent local validation; direct local Git checks for this case study confirmed the tree identity and all 21 committed file hashes. This closes the defined narrow local baseline. [C6]

Locally committed, not pushed or published as a source release in that closeout. The closeout explicitly records no remote access, push or publication. Publishing this case-study narrative is separate from publishing the source repository. Full-charter acceptance remains a separate decision, and all eight workflow outcomes remain unverifiable in the closeout. Local validation and commit identity do not supply downstream action evidence. [C6]

Where the control boundary sits

The packet-state rules permit a pending packet to become routed, held or rejected only when the routing verdict and an explicit move request agree. With no move request, the documented default is to generate reports and indexes. The rules prohibit executing packet contents as part of that transition. This separates a coordination decision from permission to carry out the work; it is a documented policy, not evidence that every boundary was enforced in operation. [C1]

What would establish a task result?

A stronger claim needs evidence for the specific task and action: the approved scope, the source revision, a contemporaneous execution record, the resulting artifact or changed destination, and an authoritative readback tied to that result. An acceptance record must say which completion criteria passed and which remain open. The October 30 acceptance closes the reporting-control question only. Repeated reliability or savings claims would additionally require observations across a defined period and a defensible comparison baseline. [C5]

What I take from the work

A useful acceptance rule must survive contact with incomplete evidence. Retain the report for the exact run, record coverage item by item, and distinguish a good launch from a demonstrated result. Here, keeping the failure visible made the next evidence requirement clearer than a broad green status would have.

Scope and remaining limits

This account reviews the earlier evidence packet, the October installation and scheduled-run acceptance, and the November 3 committed local-baseline closeout. It does not announce a deployed Action Gateway, current fleet operation, measured savings or improved reliability. The November 3 receipt-control amendment is identified separately from the earlier packet revision. Unreviewed draft claims and Action Gateway adoption remain excluded; the local commit does not establish a pushed repository or public source release. The original study and operational archive remain privately retained; publication of this account does not adopt or change the system. [C4]

Evidence notes

[C1] Private committed-evidence packet, revision 6e006c9: project charter, project schema and packet-state rules. Review verified all 110 packet-file hashes against its manifest. These establish the inspected derivative’s integrity, not execution.

[C2] Same packet, tests/test_effect_checks.py, EvidenceDeclarationTests, PathResolutionTests and DormancyTests. Test definitions were inspected; no passing-test total is claimed from their existence.

[C3] Privately retained October 2 Observation Evaluation and Addendum B, both dated October 10, 2026. Evaluator-output SHA-256: a9d9ec6f59ebd932d96466fbf9f335674033245a84a623c9aa96f510c90dda07. Addendum SHA-256: d35bbe83b9cd37363be2f5c0f581f9897297248e9e8bbf0a2796404d64872e52. The files were read and hashed for this account; historical observations were not rerun.

[C4] Packet README exclusions and October 10 owner closeout. This November account is a new, limited publication; the earlier private study is preserved. Underlying records contain operational detail and are not public downloads.

[C5] Private October 23 eleven-file installation receipt and October 30 natural scheduled acceptance: ACCEPTANCE.md, ACCEPTANCE.json, retained report, telemetry ledger and output/error logs. All four retained snapshot hashes were checked against the acceptance record in this November 5 review. Acceptance JSON SHA-256: 398359e41149244e03622e8c62d96eace5a1d858799b586b5c4f037f76585445. Historical acceptance is limited to receipt controls. No operational code or tests were run for this publication; private task and infrastructure details are omitted.

[C6] Private November 3, 2026 local-baseline closeout: CLOSEOUT.md, VERIFY_COMMITTED.json and CLOSEOUT_READBACK.json. Local commit dc77ad29117c4f022e15524f96df7dbb76bc86c0; complete tree 96536145b1c82831c25ae7dcc3ef2cba2591211f. November 5 read-only Git verification confirmed the branch reference, 21 changed paths and all committed file hashes against the retained verification record. The no-push/no-publication statement is scoped to the recorded closeout; no remote-state audit or test rerun was performed.

← All projects