Chapter 110. Generation Operator Practicum: Every Attempt Must Add Information#

A generation operator does not click repeatedly until a good image appears. They read the shot packet, select the right references, control variables, log attempts, recognize stop conditions, and submit candidates with evidence for review. The practicum shots are the two-person document handoff and the rainy-night arrival.

110.1 The operator's boundary, preflight, and the baseline attempt#

The boundary.

The operator may create candidates, adjust parameters within an approved range, request missing context and propose alternative routes. They may not change locked facts, character identity, dialogue, budget caps or final approval.

Inputs are the prompt manifest, the reference bundle, state, budget and acceptance criteria. Outputs are candidates, an attempt log, a technical report and a recommendation — not a decision.

Preflight.

Verify the task ID, model capability, input versions, reference angles, rights, budget, start and end states, and the stop conditions. A left-hand versus right-hand conflict blocks immediately; the operator does not pick whichever looks plausible.

Upload the minimum references, confirming no retired or quarantined asset has slipped in. Each reference records why it was included.

The baseline attempt.

The first attempt establishes a baseline with the standard recipe, rather than making ten "improvements" at once. Record identity, action, space, performance, technical quality and cuttability. Classify failures against specific invariants.

If the handoff baseline duplicates the prop, the next attempt adjusts only the action keyframes or reduces concurrency — it does not change model, prompt, references and duration together.

110.2 Controlled variables, re-anchoring, and stop rules#

Controlled variables.

attempt_plan:
  hypothesis: reducing Gu Zhou's head turn preserves identity and hands
  change: head_rotation_20_to_8_degrees
  hold: [model, seed_family, keyframes, references, duration, camera]
  expected: identity_pass_and_same_action_state
  stop_if: handoff_peak_lost

A result that contradicts the prediction is still valuable. The attempt log records observations, not "this one wasn't good."

References and re-anchoring.

Choose identity images by shot angle, inherit state from the previous shot, and re-anchor identity to the golden references periodically. Do not derive continuously from the newest generated frame. Multi-person shots use slots and feature attribution.

Wet hair on the rainy night is only a state overlay; it does not change the character reference. If the wet-hair state alters the face shape, reduce it or composite in layers.

Stopping and alternatives.

Stop after three failures on the same invariant, when projected total cost exceeds the alternative, or when a fix introduces a bigger problem. Propose splitting the shot, a hand insert, reaction coverage, local compositing or a live pickup, stating what the narrative keeps and what it costs.

Continuing to sample requires a new falsifiable hypothesis. "Maybe next time" does not earn budget.

110.3 Candidate submission, injected failures, and acceptance#

Candidate submission.

Candidates enter quarantine and, once technically validated, carry first and last frames, the usable range, cost, inputs, failed items and a recommended use. The operator may recommend; they cannot mark their own output approved.

A candidate with a beautiful background and failed identity can be tagged "background only" — and never becomes a character reference.

Injected failures.

The facilitator injects network timeouts, duplicate responses, a late-arriving old pack and budget exhaustion. The operator follows the runbook rather than renaming files by hand or repeating calls, and distinguishes creative failure from infrastructure failure.

Stopping safely with complete evidence is itself part of acceptance. A shot that is not yet finished does not mean the operating process failed.

Acceptance and deliverables.

Preflight 15 percent; variable control 20 percent; identity and state 15 percent; attempt log 15 percent; stop judgment 15 percent; candidate evidence 10 percent; failure handling 10 percent.

• Each attempt changes only its declared variable.
• Failures are classified against invariants.
• Identity is re-anchored to the golden references.
• Stop conditions were actually enforced.
• No candidate was approved beyond the operator's authority.

Deliverables: preflight_check.json, attempt_plans/, attempt_logs.jsonl, candidate_manifests/, cost_ledger.csv and escalation_packet.md.

110.4 A simulated shift, candidate review, and the retrospective#

Simulated shift: four attempts, four different questions.

The handoff baseline produces a duplicated folder and drift in Gu Zhou's face. The first improvement reduces only his head turn: identity passes, the folder still duplicates. The second splits the handoff into start, two-handed peak and end keyframes: the prop passes, the performance is stiff. The third adds only a small eye shift and breathing: all three pass. The fourth was planned as "one more, prettier" — and the operator cancels it under the stop rule, because acceptance is already met and there is no hypothesis left to test.

Those four records tell the team that head rotation affects identity, that state keyframes solve the prop, and that micro-motion improves the performance. Changing seed, references and prompt together each time would have produced four inexplicable results.

That afternoon, the rainy-night task times out remotely. The operator does not resend immediately; they check the call intent and the vendor job ID and find the remote job still running. It returns twelve minutes later, avoiding a duplicate charge. A pack update then makes the result stale and the system quarantines it automatically — and the operator does not request an exception to "use it anyway" because the image is good.

Candidate review.

The operator recommends the third attempt, showing the two failed versions, the changed variables, the cost and the usable range. The reviewers ask why not the first version, which scored higher on identity. The operator points out that its action end state failed, and a composite metric cannot offset that.

The facilitator's note: an operator's professional value is reducing attempts that carry no information, and switching to a different narrative route before hitting the model's capability wall. A beautiful, inexplicable accident can be used — but it does not automatically become the project's recipe.

The live retrospective.

The facilitator supplies a failed video with the prompt hidden. The operator classifies it by identity, state, action, space, performance and technical quality, and states which inputs are needed to determine the root cause — saying "add negative words" immediately is prohibited. They then open the manifest and may propose exactly one next variable with a falsifiable prediction.

The second question presents two candidates: A scores high on identity with the wrong action end state; B clears identity by a small margin and passes action and performance. The operator should recommend B or justify rejecting both, not average the scores. The third simulates having budget for one more call, and asks them to compare the expected value of generating again against splitting the shot as insurance.

Evidence review opens the attempt log, input checksums, remote jobs and costs — not the operator's memory of the settings. Acceptance is about whether the operator knows when to generate, why to change, how to stop, and whether the candidate can be handed safely to the next workstation.

After the retrospective, a second operator re-runs the baseline from the records. Pixel identity is not required, but they must obtain the same semantic inputs, the same stop rules and the same acceptance route. If it cannot be re-run, the evidence chain is incomplete.

A note on sources#

This is a rehearsal design rather than a vendor-specific procedure. What transfers is changing one declared variable per attempt, classifying failures against invariants, enforcing stop rules, and keeping recommendation separate from approval.