Michael Oakes
AI operations case study · November 2026

Office Space: making AI handoffs accountable

A supervised-work design that separates a request, its delivery, the worker’s result, and acceptance.

Knowledge operationsImplemented validatorSynthetic test evidence

The problem

A shared channel makes requests visible, but it does not prove that an assistant received them, accepted them or completed the work. Reposting an uncertain request can create duplicates. A successful file write can also be misattributed to a worker whose tools could not perform it. Office Space addresses those distinctions through bounded work packets and explicit records. [O1] [O3]

Role and deliverables

Michael is the owner and decision-maker in the operating records. The AI-assisted package supplies three portable operator guides, an offline packet validator, synthetic tests and saved verification receipts. The roles distinguish supervised coordination, review and records work. The guides are inspectable documents; their presence does not install an integration or grant access to an application. [O1]

A handoff with a defined boundary

Each packet identifies the task, source revision, owner, permitted sources and effects, result destinations and completion checks. The procedure calls for one verified session, a minimized input packet, a check for prior delivery, and independent readback of each saved result. The existing task record retains current state; message threads carry pointers rather than becoming additional queues. [O1]

  1. RequestDefine scope and owner.
  2. DeliveryVerify receipt; avoid replay.
  3. ResultInspect the actual output.
  4. AcceptanceCheck the agreed criteria.

What the validator actually does

The Python validator checks required fields and list types. It returns a hold when supplied fields indicate uncertain delivery, a previously sent packet, an unverified session, missing authority verification, source drift or a task that is not ready for initial handoff. Successful validation is labeled STRUCTURE_PASS_ONLY, with no authority or external-action effect. [O2]

This is deliberately a limited check. It reads the supplied fields; it does not independently establish permissions, inspect a provider conversation, detect a concurrent submission or authenticate an approval. Those remain operator responsibilities. Its duplicate-delivery check rejects a declared sent or unknown state; it is not a persistent, system-wide duplicate-prevention service. [O2]

The evidence

The retained test receipt records twelve synthetic tests passing. The complete set is one valid-packet case plus eleven negative cases: (1) already-sent delivery, (2) unknown delivery, (3) platform denial, (4) occupied session, (5) stale source, (6) unsuitable data class, (7) rejected task, (8) missing authority verification, (9) a string in place of a boolean, (10) a malformed list, and (11) missing source revision. All eight files listed in the package manifest matched their hashes in this review. The tests were not rerun, and these results concern the offline validator only. [O2]

A concrete attribution lesson

A saved October 4 receipt records a coordinator-assisted update to the existing Drive routing document, including a downloaded-byte comparison and preservation of the earlier content. It explicitly says this did not prove the other provider’s native write capability and that unit acceptance was still pending. The contemporaneous Slack checkpoint makes the same distinction. The result is useful precisely because successful delivery did not silently become a capability or completion claim. [O3]

October 6: one records unit reached acceptance

The October 4 pending receipt was not the last decision. For the bounded records-consistency task OS-GROK-20261004-01, the direct October 6 Slack thread records a coordinator correction and readback at 1:56 p.m. Central, followed by an acceptance sent using Claude at 2:02 p.m. The acceptance states that Michael delegated the decision to Claude. Its basis was two resolved gaps: attributed before-and-after wording for a capacity-note correction, and a stale sentence in the companion record labelled historical and verified. The Notion coordination record explicitly points to this acceptance as superseding the earlier pending status. [O4]

This is historical acceptance of one supervised records task after correction. It is distinct from the twelve offline validator tests, and it does not prove autonomous dispatch, provider-native file writing or acceptance of the broader Office system. The acceptance did not start unit 02.

The revision trail and its limit

The October 4 delivery receipt identifies the Notion result version and the companion-file version it used. The subsequent change record is explicitly tagged OS-GROK-20261004-01-NEEDS-CORRECTION-20261005, superseding the prematurely ready version. The October 6 correction and acceptance identify the same task, destination and outstanding defect. They do not provide a final accepted content hash or an immutable snapshot of both record surfaces. The evidence therefore supports the recorded acceptance decision and its stated basis, without claiming a byte-exact reconstruction of the final accepted artifact. [O4]

Three failure cases the packet makes explicit

Historical synthetic cases and their operational limits
Supplied conditionImplemented checkStill needs real-world evidence
Delivery is SENT or UNKNOWNThe validator holds the initial handoff.The operator must inspect the actual conversation; the validator cannot discover a prior submission.
The source has driftedA declared drift or unknown drift state produces a hold.The source must be compared with the approved revision; a supplied flag is not that comparison.
Session exclusivity is the text “true”The validator rejects it because it requires the boolean value true.A correctly typed value still does not prove exclusive use of a provider session.

These examples come from the validator and matching synthetic test definitions. The retained twelve-test receipt is historical offline execution evidence. It is not a live submission test, a permission check or proof that duplicate work was prevented in practice. [O2]

Delivery, saved output and acceptance are separate

The procedure requires a task ID, revision and owner before a handoff; inspection of the provider conversation before submission; observation of actual acceptance and output; and independent readback of each changed destination. It then calls for an independent checker against the completion tests. An uncertain delivery stays on hold rather than prompting a repeat submission. These are instructions in the operator guide, not a claim that the complete sequence was demonstrated. [O1]

The recorded coordinator-assisted write establishes a narrower result: the saved content matched a downloaded copy and the prior content was preserved. At that October 4 checkpoint, it left provider-native writing unproved and independent unit acceptance pending. The October 6 decision later accepted the corrected records unit; the earlier receipt remains a historical delivery checkpoint. A successful save and an acceptance decision are separately evidenced events. [O3]

Evidence needed for a stronger result

These are publication evidence requirements derived from the guide’s completion procedure and the validator’s limits. The operator-skills package alone does not supply these broader operational results. The separately retrieved October 6 acceptance supports one records-consistency unit; final artifact hash binding and broader operational measurements remain missing. [O1] [O2] [O3]

What remains unproved

This package demonstrates a documented handoff method, implemented structural checks and recorded synthetic test results. It does not demonstrate automatic assignment, installed provider controls, end-to-end acceptance of the broader Office system, customer adoption or measured savings. Historical review and delivery states are not represented as current provider status.

The lesson

Coordination becomes easier to inspect when every transition has its own evidence. A request can be valid but undelivered; a result can be saved but unaccepted. Naming those states gives the next person a specific check to perform and reduces the temptation to repeat uncertain work.

Evidence notes

[O1] Privately retained Office Space operator-skills v1 package: README and common supervised-provider procedure, especially steps 1, 4, 6, 8–10 and its completion-evidence section. The Notion operating record supplies dated context and owner roles. No private task links or content are reproduced here.

[O2] Same package: tools/check_packet.py, tests/test_packet.py, manifest and synthetic-test receipt. Validator SHA-256: 371f7abd528a5d2a5f4920c150f99c729825376f82abab96aa57afabe817c9f0. Receipt SHA-256: 664ccea75ec44d25df4a9323bdc225289422efed49932f4559cd39ee185c0a86. Hashes identify the reviewed files, not independent reproduction of the historical run.

[O3] Package Drive-delivery verification receipt and the October 4, 2026 Office skills checkpoint in Slack. They support historical attribution and record-write verification; they do not prove provider-native delivery or final unit acceptance. Raw records remain private.

[O4] Privately retained Office task record: October 4 coordinator-delivery receipt, October 5 named correction revision, and October 6 Slack correction/readback and acceptance messages. Direct thread read confirmed the acceptance message’s “Sent using Claude” attribution; the Notion worker-state correction identifies the same unit and superseding decision. The delivery receipt names Notion source version 2026-10-04T17:56:11.796Z and companion source version 2026-10-04T17:18:10.864Z. These are historical inputs, not the final accepted revision. The October 6 acceptance contains no final artifact hash. Retrieved November 5, 2026; no provider work or current cloud-byte comparison was performed.

← All projects