P366 · Tool use

Process Success and Semantic Completion Handshake

Require both successful process exit and a task-completion record before acknowledging background work.

Editorially reviewed

These examples and illustrative results are independently authored teaching materials, not measured model results.

Use case

A background analyzer processes an observation snapshot. Teaching run R7 reports an unreadable input but exits 0. Treating that as success and archiving the only pending input loses work.

Mechanism

Bind the run to a fixed input snapshot and identity. Its contract emits exact {“status”:”analysis_complete”} as the final nonempty line only after required reads, analysis and writes succeed. A completed no-findings run emits it too; blocked runs do not. The parent checks both exit 0 and exactly one final record from this run, then independently verifies required artifacts. Acknowledge only that snapshot and retain later inputs; keep pending work and diagnostics on failure.

Bad example

R7 exited 0, so analysis completed. Archive all current observations even though its log says the input was unreadable.

Good example

Exit 0 for R7 establishes process completion only. Check exactly one final analysis_complete record and required artifacts from this run. Missing/duplicate records, failed writes or snapshot mismatch leave input pending. Valid no-findings analysis can complete, but never archive new observations this run did not read.

Why the change matters

Process and task states answer different questions. An explicit completion protocol supplies a checkable signal; snapshot binding prevents acknowledgment from expanding to unprocessed data.

Observable expectation

Teaching matrix: exit 0 with a blocked read, absent/duplicate record or later nonempty output is unacknowledged. Nonzero exit is unacknowledged even with a record. Exit 0, the exact final record and verified artifacts permit acknowledgment of R7’s snapshot; newly added observations remain pending.

Limits

A self-reported marker can be wrong and proves neither analytical quality nor every write, so independent artifact checks remain necessary. This exact record is an illustrative contract; hosts may use other protocols. Archival and acknowledgment also require concurrent-update protection rather than a model’s verbal completion claim.

Sources and evidence

Read the editorial criteria