# Publishing Desk correction and unseen-case result

Run date: September 20, 2026
Platform: Hermes Agent Python runtime
Provider: Nous
Requested model: `deepseek/deepseek-v4-flash`
Tools enabled: none
Human publication approval: not given

## Known fixture

The corrected Writer stopped padding the sparse repair-workshop source packet and returned an explicit source-shortfall draft. A later normally completed Writer produced a 77-word newsletter and a 175-code-point social draft using only supplied facts. Deterministic pre-segmentation supplied nine newsletter and five social assertions to a separately imported Editor.

The final Editor audit completed normally, matched every supplied boundary and attached literal evidence to every assertion. Both structural evidence gates passed. Codex semantic review found each publishable assertion to be a direct paraphrase of the approved sources. This is a pass for the known technical fixture, not human publication approval or proof of general crew reliability.

Provider-derived estimates for the final known Writer/Editor pair were `$0.00030788` and `$0.00038109`. These are estimates, not an invoice.

## Locked unseen fixture

A different fictional Tool Library packet was hashed before runtime. Scout, Researcher, Story, Writer and Editor then completed the full sequence from the current bundle. Every role used its imported profile, no tools, the original packet and the preceding untrusted handoff.

The unseen crew **failed acceptance**:

- Scout introduced an unsupported “without buying them” angle.
- Researcher missed that addition and supplied an irrelevant estimated count.
- Story introduced “trusted borrowing” and “watch for the catalog URL.”
- Writer returned fenced rather than strict JSON, declared a source shortfall despite a measured 261-word body, misstated the body as 245 words, used the wrong review-note type and added unsupported catalog/newsletter claims.
- Editor's prose noticed the important unsupported claims but omitted all required audit statuses, supplied malformed evidence values, used an invalid shortfall status and changed the deterministic social boundaries.

The Writer and Editor fenced responses were parsed only for defect inspection. That bounded transformation does not convert the original responses into accepted handoffs. The five stages reported provider-derived estimates totaling approximately `$0.001542834`; all passed the post-run credential-residue check.

## Current conclusion

Publishing Desk can pass the known correction fixture, but it has not generalized to the locked unseen case. It remains a failed crew demonstration.

A local executable handoff gate now combines bounded JSON parsing, Writer schema and length validation, deterministic assertion boundaries, Editor schema validation and literal-evidence checks. Replaying the preserved outputs through that gate accepts the known Writer/Editor pair and stops the original locked unseen Writer before Editor because its 261-word body contradicts its source-shortfall declaration and its review notes use the wrong type. That replay remains preserved evidence, not a new model run or a successful crew result.

The corrected Writer 2.0.2 contract was then tested in a new gated continuation using the same locked source packet. The Writer returned a 267-word newsletter and 162-code-point social draft. The gate removed only contradictory self-reported count metadata and passed the text unchanged to Editor with 23 newsletter and four social assertion boundaries. Editor completed its model turn but returned malformed JSON containing an unescaped control character. The parser stopped the sequence before schema or evidence acceptance. The Writer and Editor provider-derived estimates total `$0.000617754`; this is not an invoice. Both stages passed the credential-residue check.

Hermes schema-enforced output then produced a valid Editor audit without a retry. That audit correctly identified two unsupported catalog promises. One bounded Writer revision removed those claims and passed the strengthened Writer gate; the Editor re-audit returned valid structured output but altered punctuation in two passages presented as literal source quotations. The deterministic evidence gate therefore still rejected the handoff.

The next correction is one bounded Editor evidence correction using the exact gate errors, followed by the same deterministic acceptance check. A structural pass would still require semantic source review and a human publication decision.