Skip to content

Reported cases & illustrations

Finished is not the same as right.

The director reported these separate cases. They are not independently verified. Invented examples and proposed steps are labelled separately.

Historical reviews

Two separate design reviews

Fact

Director-reported observations

  • A website design review (11 Nov 2026): the director found unrequested design changes among retained candidates.
  • A separate comparison hero illustration on another site (12 Nov 2026): a different hero had already been selected. The director then found what appeared to be two left hands in a creative-director comparison image.
  • The user reported a review burden. It was not measured.
Interpretation

What it suggests

In two separate cases, the director challenged unrequested design changes and an apparent anatomy error. This shows the right to question or reject output. Savings were not measured. Corrections were not demonstrated. Plan for the cost of reviewing drafts.

Planned

What comes next

Future work could record review time and correction counts. These have not yet been measured.

Director-reported animated film review · November 12, 2026

The director stops the render loop

An unreleased, roughly 60-second animated review cut. Source: director’s own review notes (private research record). Not independently verified.

Fact

Director-reported observation

  • The director set personalities, original art, an arm-cross gesture and the exact ending.
  • Production delivered a review cut. It was not approved as final.
  • Self-rated 5.2/10. A personal score, not an audience measure.
  • Noted: poor mouth movement and lip sync, little chemistry, flat pacing, music missing or inaudible.
Interpretation

Rejection, not final delivery

Generated clips are not accepted work. The director set the standard and rejected the cut. Some clips may still be reusable. Financial impact and human time were not measured.

Planned

Proposed corrective action

No more bulk paid generations until one speaking shot shows lip sync, audible audio and consistent acting. The rejection was observed. A production fix is not yet confirmed.

A proposed proof-shot check. Not an automated tool
  1. Define what good is

    Character, gesture, ending, sound.

  2. Interrupt repetition

    Stop re-rendering without progress.

  3. Change the method

    Fix one speaking shot first.

  4. Verify before paying again

    Director approves the proof shot.

Map the human owner and stop rule →

Director-reported film loop & version audit · November 13, 2026

Stop the loop. Preserve the source.

Fact

What the record shows

The supplied forensic inventory records 31h56m of calendar elapsed time, including breaks. This is not active labor. It lists 43 unique altered named long-form MP4 content hashes, excluding picture/delivery intermediates and the canonical source. These are not 43 approved creative decisions or paid generations. The latest film is rejected. There is no accepted master. The historical 5.2/10 rating belongs to a different, earlier cut.

The director heard “Look at them, they're in love” while the two-person drawing was cropped away; earlier artwork was upside down. An abrupt transition and a weak wedding/bar punchline also missed the intended experience: attraction imagined by the audience through subtle expressions, not forced explanations.

Interpretation

The director's authority

The director can stop repeated fixes, audit versions and preserve the original 60-second source. They can refuse further unauthorized substitutions. Technical completion does not override their rejection or mean the work is accepted.

Planned

A stop rule, not a cost estimate

After two materially equivalent attempts without observable progress, pause. Choose a genuinely different method or qualified reviewer. Define what would be accepted. Check that the result has changed before resuming. A human applies this rule. This site does not enforce it. Full active hours and tool-credit totals remain unknown. Measure time, credits and cognitive cost separately. Do not invent rates or infer totals from elapsed time or file counts.

Director-reported connector loop · November 12, 2026

A human interrupts the support loop

Fact

Director-reported observation

A connector-confirmation dialog kept asking for a larger window. The Continue button stayed disabled. The director reported five unsuccessful attempts and directed a change of reasoning. The cause and final connection outcome are unverified. Time, financial impact and cognitive cost were not measured.

Interpretation

The contribution was interruption

The human saw that equivalent instructions were not changing the screen. They stopped the repetition and asked for a different method. That supervision is a human contribution. It does not show successful AI execution.

Planned

Pause, reframe, verify

After two equivalent failures without progress, a named human pauses the work. They choose a genuinely different method or escalation. They require an observable result before continuing. This is a proposed procedure for a human to apply. This site does not enforce AI behavior.

Illustration

The steak that came back burnt

A customer says a steak is undercooked. It comes back burnt. The kitchen missed the intended middle and went to the opposite extreme. Before another attempt, ask what the customer meant. The reviewer’s first assumption may be wrong too.

The answer is not another instruction. It is these four moves:

  • Pause the loop. Stop re-issuing equivalent fixes that keep missing the middle.
  • Choose the right reviewer. A chef, editor, art director or subject specialist who knows the target. Not necessarily the engineer running the agents.
  • State the intended middle. Describe how the result should look, read or behave. Do not just reverse the first complaint.
  • Verify against it. Check the new result against that stated standard before spending another pass.

Several AI agents reviewing each other can add cost without adding judgment. Each pass adds work and review. None owns the standard. Independent AI checks can help when a qualified human can inspect the standard. The standard and reviewer matter, not the number of agents.

This is an invented example. No outcome, cost or time is claimed.

Illustration

The approval that wasn't

An assistant may send customer replies "after approval." The team presses one approval button hundreds of times a day. Authority exists on paper but is lost in practice. This is an invented example, not a reported case.

Map your own workflow →