# Publishing Desk structured-output and revision result

Run date: September 20, 2026
Locked source SHA-256: `ba8faf8619afebe2077c570608565ea1f94cf34e967edbd3d1560dcf10ed4cca`
Writer 2.0.2 archive SHA-256: `80b2d7be68b0d57b834157d2a3749c77f8728114e7098f886b92d73dad3c2a5d`
Editor 2.0.1 archive SHA-256: `6e0141767028d36d78fd363f61942163d18383fefe377f349ee16a1c6522b9f0`

Hermes's machine-validated output contract corrected the prior formatting failure: Editor returned valid JSON on its first response, with no schema retry. The deterministic evidence gate then rejected the handoff because Editor correctly marked two unsupported future-facing catalog claims and supplied one non-literal combined source quotation. Its provider-derived estimate was `$0.00106967`.

That defect report was returned to Writer for one bounded revision. Writer removed the unsupported catalog promises and passed a strengthened gate that requires one source check for every deterministic newsletter and social boundary. Editor then re-audited the revised handoff under the same output contract. Both responses were valid JSON without a schema retry.

The revised newsletter and social draft contained no assertion that Editor labeled unsupported or ambiguous. Acceptance still failed because two Editor evidence quotations changed source punctuation instead of quoting literal contiguous passages. Writer's revision estimate was `$0.0007036`; Editor's re-audit estimate was `$0.00135525`. These are provider-derived estimates, not invoices.

One prompt-only evidence correction then failed: its first response corrupted JSON text, and its single schema retry incorrectly claimed that no sources were supplied. The gate rejected it. Its provider-derived estimate was `$0.003106868`.

The interface was then changed so Editor selected from 24 deterministic evidence options containing immutable source IDs and exact source sentences. The resulting audit was valid on its first response and the complete revised Writer → Editor continuation passed the executable structure-and-evidence gate with zero issues. Its provider-derived estimate was `$0.00442156`.

Current result: **bounded Writer–Editor continuation passed; full crew still failed acceptance**. This proves the corrected inter-role mechanism on the locked source packet, not a successful new Scout → Researcher → Story → Writer → Editor run, general crew reliability, semantic correctness or human publication approval.