Skip to content

Examples

The moment someone noticed.

A near-miss. A correction too far. A person who withheld approval. Reports and examples are labelled separately. Noticing is not independent verification.

Fact director-reported; not independently verifiedInterpretation our reading of itIllustration constructed scenarioPlanned not yet done

November 11, 2026 · Creative / educational

A homepage changed without being asked

Fact

Director-reported: on November 11, 2026, retained Main and Alternative candidates showed unrequested homepage design drift. These candidates have not been independently verified here.

Fact

This report concerns homepage design drift, not an anatomical error.

Fact

The user reported a meaningful review burden. Hours, counts and rates were not measured.

Interpretation

Without a record of what was approved, each new version becomes the standard. That makes drift harder to see.

Interpretation

Keep named candidates to compare requested changes with unrequested drift.

Interpretation

Check unrequested changes against the approved brief before accepting, revising or rejecting a candidate.

Planned

Record review time and the problems caught in future work, to show the burden with evidence.

November 12, 2026 · Independent hero comparison

A comparison image appeared to have two left hands

Fact

Director-reported: on November 12, 2026, a creative-director comparison variant appeared to have two left hands, after a different hero had been selected for another site.

Fact

This is separate from the November 11 homepage design review. It is a reported apparent flaw. The source image has not been independently verified here.

Interpretation

Inspect the source image before using the variant or calling the apparent anatomical error a fact.

Interpretation

Reviewing a later comparison variant does not reverse the prior hero selection.

Director-reported film review · November 12, 2026 · Human decision

A director-reviewed animated film

Not accepted.

Director-reported: an unreleased, approximately 60-second review cut on November 12, 2026. It was not finally approved. The director’s 5.2/10 rating is subjective, not a formal audience measure.

  1. Artistic intent →

    Characterization, gestures, an expressive speaking shot, precise timing and an exact ending.

  2. AI execution →

    Generated clips and music do not prove the work was accepted.

  3. Human review / veto →

    The director rejected the reviewed output against the intended performance.

  4. New quality gate

    Proposed: review one revised speaking shot before further bulk generation.

Evidence

Director-reported observation · owner-supplied research account

The director reported absent or inadequate mouth movement and lip sync, insufficient acting chemistry, quiet or flat pacing, and missing or inaudible background music in the reviewed version. This page has not independently inspected the cut.

The account notes 10 completed Seedance jobs and music generation. Completion does not establish finished quality or final acceptance.

Director decision

Not accepted. The director set characterization, gestures, timing and the exact ending. Then the director checked whether the work met that intent. No final MP4 acceptance is verified.

Proposed next check

Analytical interpretation · proposed corrective action

These are proposed choices, not a recorded Decision Gate verdict:

  • ACCEPT: only after review against the stated criteria; not established here.
  • REVISE: the speaking shot; check mouth movement, lip sync, acting chemistry, music, timing and ending.
  • HOLD: bulk generation until that shot is reviewed.
  • REJECT: the claim that a final master is already accepted.

Unknown cost

Film spend, human hours and lost revenue are unknown. Workspace balance changes, replacement quotes and separate site or API use are not a film total. Some clips may be reusable. Rejection does not mean all spend was wasted.

Suggested case: Can one revised speaking shot meet the director’s criteria before bulk generation resumes? Record what was checked and what remains unknown. Improvement is not guaranteed.

Review this case in the Decision Gate

Director-reported follow-up · November 13, 2026

The file passed. The film did not.

Fact

Director-reported: the 58-second film had the correct format and audio, but remained REJECTED, with no accepted master. During “Look at them, they’re in love,” the two-person drawing was cropped away. Earlier artwork was upside down; the transition was abrupt and the wedding/bar punchline weak. These observations have not been independently verified here.

Fact

The supplied forensic inventory records 31h56m of calendar elapsed time, including breaks (not active labor) and 43 unique altered named long-form MP4 hashes, excluding picture/delivery intermediates and the canonical source. That is not 43 approved creative decisions or paid generations. Full active hours and tool-credit totals are unknown. The historical 5.2/10 rating belongs to a different, earlier cut above, not this film.

Fact

The director wanted viewers to imagine attraction through subtle expressions, not forced explanations. The pictured subject and cut no longer supported that intent. The director withheld release.

Interpretation

Automated stream checks can confirm duration, format and audio, not artistic acceptance. Before release, a qualified person checks the pictured couple, the transition and the intended performance. Technical checks do not give permission to release. This is not an automatic Decision Gate verdict.

Source research record ↗ · Owner-supplied account; source access may be required.

Director-reported connector review · November 12, 2026 · Support loop

The same screen kept returning. The person stopped the loop.

Evidence

Director-reported observation

The human reported five unproductive Claude–Notion connection attempts and halted repeated guidance. The reported confirmation dialog still asked for a larger window, with Continue disabled. Five is a self-reported count, not an independently measured attempt log.

Human intervention

The human saw no progress, stopped repetition and asked for a different method. That is supervision, not successful AI execution. The connection outcome and technical cause remain unverified.

Proposed next check

Consider HOLD when equivalent attempts bring no observed progress or new evidence. Try a different method. Read back the connection state to check for change. This is an interpretation, not a recorded tool verdict. Attempt count alone does not trigger a veto.

Unknown cost

Time, money and mental effort were not measured. Root cause and final outcome remain unverified. Time savings have not been shown.

Illustration · Not a reported restaurant incident

Undercooked, then burnt

Illustration

A diner says the steak is undercooked. It comes back burnt. The correction goes too far. It still misses what the diner wanted.

Illustration

Reading “not acceptable” as “do the opposite, harder” misses the target. Ask what the diner meant. Agree on the target, then check the result. Do not assume your idea of “right” is theirs.

Interpretation

Someone who knows the target needs authority to say “too far the other way”. They can choose HOLD or reframe, instead of approving or repeating the request.

Interpretation

A second reviewer with separate criteria can help catch a missed standard. A second undirected agent may keep costly rework going. The qualified reviewer need not be the engineer running the agents. It may be an editor, clinician, sound director, designer or teacher.

Interpretation

This example makes no general claim about what systems can do. No cost, duration or rate is measured. The Decision Gate does not detect or enforce a hold for correcting too far. A person decides.

Before asking for another pass, listen and record the intended standard: work the overshoot through the Decision Gate.

Non-artistic, higher stakes

Where the stakes are people

02Illustration

Clinical note summary

A model condenses a patient history. It reads well but omits an allergy mentioned once.

Judgment · Hold. A clinician verifies against the source record before anything enters the chart. Omission risk is a veto condition.

03Illustration

Credit pre-screen

An assistant ranks applicants. The ranking correlates with postcode.

Judgment · Reject the ranking as a basis for decision. Proxy discrimination is a fairness veto; applicants retain the right to human review and explanation.

04Illustration

Hiring shortlist

AI summaries of CVs are concise and confident, with one summary inventing a qualification.

Judgment · Revise. Summaries are checked against source documents; candidates are not filtered on unverified claims.

05Illustration

Public service notice

A drafted multilingual flood advisory is fluent in English, uncertain in two other languages.

Judgment · Hold the uncertain translations for a native reviewer; publish verified versions. Speed matters, but misinformation in an emergency harms.

Work a case through the Decision Gate