Locked fictional Tool Library packet
The approved launch notice, inventory report, operations FAQ, superseded draft and assignment limits supplied to every role.
Back to Test Records
Runtime failed acceptance
Publishing Desk runs a five-role newsletter handoff
Five roles reach a source-checked draft and the required human-review hold
Material supplied
The Bot begins with disclosed inputs
The Riverside Library, Tool Library, dates, inventory and operating rules are fictional. The fixture deliberately mixes current, superseded and missing information to test source discipline.
Download file
Sentence-level audit fixture
Every sentence in the failed Writer body, its bounded status and the original source passages used by the deterministic gate.
Recorded failed request
The assignment used for the failed handoff
Prepare a 250–350-word newsletter for adult Riverside Library members and an X draft of at most 280 Unicode code points. Use only S1–S4. Preserve current versus superseded details and the distinction among tools, categories, kits and borrowers. Do not invent a catalog URL, contact, placeholder, partners, sponsors, demand, savings, popularity, usage, outcomes, quotations, guarantees, environmental claims, staff names, delivery options, registration availability or publication approval.
Conversation evidence
What the record currently contains
These lines summarize the preserved stage outputs and Codex reviews. Provider-internal reasoning and identifiers are not published. The linked failed draft, audit fixture, deterministic result and sanitized summary are the public evidence.
Person · excerpt
Use the original fictional packet through Scout, Researcher, Story, Writer and Editor. Preserve every handoff and block unsupported public claims.
Bot · excerpt
Editor held three invented missing-information directions. Writer removed only those claims; the fresh Editor audit accepted all 25 corrected boundaries and retained the human-review hold.
Finished deliverable
What a person can inspect
Inspect the latest simplified-Writer resultThe full five-role result, exact hashes, correction loop, bounded cost, final gates and remaining proof limits.
Crew and evidence files
The bundle and records behind this handoff
These records separate the downloadable setup bundle from the outputs and checks produced during the test.
Inspect the original failed newsletter
The earlier repair-workshop regression fixture remains preserved rather than silently replaced.
Inspect the deterministic gate result
Twelve sentences covered; six unsupported and four ambiguous; semantic support, factual accuracy and human approval remain false.
Read the sanitized runtime failure summary
What ran, what failed, recorded cost limits and the proof still required.
Read the correction and locked unseen-case result
The known technical pass, the separate unseen five-role failure, bounded cost estimates and the next executable correction.
Read the Publishing Desk 1.0.2 gated result
Exact bundle and member hashes, deterministic Writer metadata normalization, Editor JSON failure and bounded cost estimates.
Read the structured-output and revision result
Schema enforcement, the bounded Writer–Editor feedback loop, removed unsupported claims and the remaining literal-evidence failure.
Read the fresh five-role result
Exact tested bundle, role sequence, gate result, bounded costs and the semantic sentence that still holds publication.
Read the independent reproduction failure
The same package and fixture produced one unsupported missing-information implication; the preserved output now fails the revised deterministic guard.
Download the Publishing Desk Crew Kit
The current guided bundle. Import each profile separately; this failed run does not establish crew reliability.
Human decisions
The Bot prepares work; a person remains responsible
- 1
Review every ambiguous or unsupported sentence against the original source packet.
- 2
Approve or reject a corrected draft only after semantic source review.
- 3
Keep publishing, sending and any public claim under human control.
Evidence ledger
Separate checks for separate claims
Package, import, role-run, reproduction, and human-review status are reported independently.
Package checksPassed
The tested bundle and exact Writer 2.0.5 and Editor 2.0.3 archives were generated, fingerprinted and included in the run.
Checked 2026-09-20Profile importPassed
The current Writer and Editor archives were imported through Hermes native profile import. This does not establish output quality.
Checked 2026-09-20Role runPassed
All five roles completed. Editor caught three unsupported directions; one bounded Writer correction and fresh Editor re-audit reached the required human-review hold with zero gate issues.
Checked 2026-09-20ReproductionFailed
An independent second run completed on the same package and fixture but added `catalog URL coming`, which the source did not establish. The preserved output now fails the revised missing-information projection guard.
Checked 2026-09-20Human technical reviewUnavailable
A human technical review is a possible long-term goal. Bot Cabinet cannot offer one at this time.