Co-work Truth Analysis Job Submit

Submit one immutable typed candidate set for the exact Co-work Truth-analysis run without writing to the Truth ledger.

MCP name: cowork_truth_analysis_job_submit

Category: cowork

Mutates state (retry policy: manual)

Parameters

Name	Type	Required	Description
payload	dict	Yes	Output conforming exactly to the schema returned by cowork_truth_analysis_job_get, including truthful source coverage and run-owned evidence references.
run_id	str	Yes	Server-issued Truth-analysis run id. It must match the gateway-injected worker session.

Details

This is the only authoritative delivery path for the exact <run-id>-cowork-truth-analysis worker. Stdout, files, terminal output, and the hosted agent's final response do not stage candidates. The server normalizes the typed payload, validates allowed claim kinds and expression roles, reanchors every expression to the immutable target, verifies existing-claim IDs against the supplied bounded context, and replaces worker-reported source coverage with coverage derived from durable search and fetch receipts.

Evidence references are source-specific. A recorded Truth span must name an admitted span receipt. A web item must name a completed run-owned fetch_id, pass its captured-text digest, and contain an exact uniquely anchored quote. A passage citation remains a non-attachable citation cue. Search hit IDs and snippets cannot be submitted as evidence.

An identical normalized payload replay returns the same completed run. A different second output is rejected. Successful submission writes only durable operational analysis output and prepared candidates: it cannot create or confirm a claim, connect an expression, attach evidence, edit the document, or mint a human decision. Those consequences require a later explicit human action through the ordinary Truth surface.

The response is deliberately compact: ok, schema wb.cowork.truth-analysis-submit-receipt/v1, analysis_run_id, public status, and output_sha256. It does not echo candidate content or imply that ledger consequences have occurred.

Provenance keeps preparation and acceptance distinct. The staged output is AI-prepared and bound to the run, provider/model authorization, and output hash. If a person later chooses Add as proposed or connects an existing claim, the claim/expression consequence is authored by that human actor while retaining analysis-run and candidate metadata. A selected web source retains its agent_run acquisition actor and external-quarantined provenance; the human authors the later support decision rather than laundering the source's origin.