From operator interventions to one unattended-work test

Commissioned by the operator on 2026-10-18: develop the consequences of the interaction-function inventory for making Commonplace more automatic, retaining only gaps in existing work. This investigation belongs to the existing workshop; it does not add a new failure catalogue or authorize changes to processing procedures.

What already has a home

Consequence	Existing home and remaining work
Learn from interventions; measure the work they support	Human-effort note: now distinguishes missing information, activation failures, and reserved decisions before choosing a repair.
Detect when support is needed	Activation and trigger dependency already carry this argument. Identify the actual trigger in the selected workflow.
Handle consequential conditions discovered during execution	Delegation and productive deferral already govern choices and return conditions. Inspect unresolved branches rather than write another general claim.
Continue after a human decision	The workshop-layer note supplies the place for working state. Test what a fresh executor needs to resume; durable storage alone does not establish successful continuation.
Evaluate the process as well as the artifact	Reuse the operator-work experiment template, adding missed conditions and recovery effort for this local comparison. Its separate outside-use commission remains unchanged.

Bounded investigation

The intended result is one evidence-backed diagnosis of a recurring operator dependency in note revision, followed by a testable replacement or an explicit reason to retain human involvement. Choose the revision workflow from accessible records; a named skill alone does not establish how an actual run proceeded.

Start with up to three recorded episodes. For each intervention retain the record locator, condition noticed, contribution supplied, evidence available to the agent before intervention, and resulting change. Distinguish unavailable information from available but unused information, reserved authority, and an unresolved diagnosis. If the records do not distinguish them, report that limit; do not infer a cause from the final diff.

The catalogue's premature-polishing entry is an available starting lead. It records substantial discarded work but explicitly leaves whether the later rewrite was foreseeable unresolved. It therefore cannot yet establish a missed correction or stopping opportunity. Seek the contemporaneous commission and review; absent those, retain the gap and select another episode.

For one recurring dependency, specify what would detect the condition, who or what would respond, and when work must stop. If the answer belongs to the operator, retain the current artifact, unresolved choice, relevant evidence, permitted alternatives, and next action. A continuation test starts a fresh executor with that state and the decision; measure reconstruction and repair work rather than assuming the checkpoint suffices.

Evaluation and closure

Before trials, fix inputs, acceptance criteria, model and harness, repetitions, and resource ceilings using the existing template. Compare fuller initial instructions with those instructions plus the proposed detection-and-response mechanism. Keep task evidence and authority equal. Use additional cases beyond the originating episodes to assess transfer.

Record accepted outcomes, necessary and unnecessary escalations, conditions missed according to review, and total setup, review, recovery, and repair effort. A justified stop can be correct behavior while the task remains incomplete; report both. Fewer questions with worse outcomes is not an improvement.

This investigation closes with a tested disposition or a documented evidence gap that prevents a test. Apply the parent workshop's admission and evaluation rules before proposing a countermeasure. Feed findings into its existing catalogue; create a new theory note only if the findings require a claim the linked notes do not already carry. No comparison has been run yet.