Chapter 74. Agentic Audit and Postmortem: Turning One Pilot into Capability#

Publishing episode one did not end the project. The team rebuilt the entire production chain from events, budget, human decisions and quality data, to judge which methods could enter the first ten episodes and which had merely worked once.

74.1 Production traceability, cost variance, and attempt data#

One traceable production chain.

The root trace ID for this workflow is trace_ep001_pilot_001. It connects the greenlight decision, the episode pack lock, asset versions, 166 model calls, 38 video candidates, 18 approved shots, 4 edit versions, 27 QC issues, the rights snapshot and the final release record.

Any timecode in the finished cut can be queried backward: which take was used, which task generated it, which identity and state were input, which prompt version applied, who approved it, what it cost and what was repaired. An asset that cannot be traced does not enter a scaling template, however correct the picture looks.

Actual cost against the original budget.

Judged by model calls alone, the project looks cheap. Real cost appears only once research, approvals, rework, rights and delivery are added. Video generation came in under budget not because the model was more stable, but because the animatic removed shots that were never going to be used.

What the attempt data showed.

Of 38 video candidates, 18 reached the finished cut, 7 failed on identity, 5 on action state, 3 on spatial drift, 2 on performance intent, and 3 were kept as backups. The high-risk folder handoff took 4 attempts; simple inserts averaged 1.3; close shots with lip sync averaged 2.8.

So the budget model for the first ten episodes no longer assumes two attempts per shot uniformly; it sets prior attempt counts by shot type. Action failures clustered in shots demanding a head turn, a line and a handover simultaneously, so the rule library gained an entry: a key handoff does not share a shot with a long line by default.

74.2 Human overrides, feedback speed, and the scale gate#

Where humans overrode the automated recommendation.

Automated evaluation preferred entrance candidate B for a marginally higher facial similarity score. The visual director chose A, because A's stopping and eyeline better matched the intent of already holding the procedural advantage, and its identity still cleared the threshold. That decision was not filed as a subjective human exception; it entered the evaluation calibration set, establishing that small identity differences above threshold should not outweigh a performance objective.

The music evaluator preferred the version with a continuous bed for emotional continuity. The humans chose the version with three silences. Later automated evaluation gained metrics for whether key Foley remains audible and whether music masks dialogue — without attempting to judge all musical taste automatically.

Two-speed feedback.

The fast loop executed immediately in the next batch: correct the hand fields in the shot packet; add an automatic OCR check for the contract; make phone-speaker playback a mandatory mix test device; supply a default shot-splitting template for key handoffs.

The slow loop needs evidence from more episodes: whether the semi-photoreal route can hold up in crying scenes; whether Gu Zhou's voice reads too cold in intimate scenes; whether the two-note motif fatigues with repetition; whether the first paywall belongs at episode six or eight. None of those becomes a permanent rule from one episode's data.

The scale gate for the first ten episodes.

scale_gate:
  scope: episodes_002_010
  decision: approve_batch_2_with_limits
  evidence:
    narrative_clarity_test: pass
    identity_closeup_pass_rate: 0.91
    approved_seconds_per_workday: 19.4
    traceability: 1.0
    rights_blockers: 0
  limits:
    max_parallel_episodes: 3
    max_unapproved_video_spend_usd: 900
    freeze_character_identity_version: lin_v03
  required_experiments:
    - rain_exterior_identity_test
    - three_character_argument_previs
    - long_narration_audio_test

The team did not approve a whole season at once. Three parallel episodes expose cross-episode state and asset bottlenecks while keeping the rework radius survivable.

74.3 System changes, incident review, and reuse boundaries#

Changes to the workflow system.

The pilot exposed three systemic defects. First, approving an asset did not automatically invalidate older context, so one task referenced a retired wardrobe image; a dependency invalidation event was added. Second, the budget service recorded call costs without human repair time; labor_minutes was added. Third, the approval interface showed candidates without their end states; adjacent frames and a state diff were added.

Each system change has an owner, a test, a release batch and a rollback plan. A postmortem cannot conclude with "communicate better," which is not a verifiable improvement.

Incident review: how a wrong take nearly shipped.

The wrong folder holder originated in the edit agent selecting take_015_02 by picture score, while that take's asset metadata was marked action_state: fail. The selector filtered only review_status and not the acceptance dimensions. Human QC caught the error, and the systemic gate had a hole.

The fix records approval state per dimension; the edit candidate query now requires identity, action and continuity all to pass; the offending candidate joined the regression tests; and historical projects were scanned for the same pattern. Responsibility does not attach to whoever acted last — it attaches to repairing the interface that let the error through.

What is reusable and what is not.

Reusable: the episode pack structure, the identity stress test, camera zones, the attempt log, five-layer QC, the rights snapshot and the release runbook. Reusable with parameters: the meeting room space, Lin Xia's identity, the wardrobe base layer and the music themes. Not mechanically reusable: episode one's beat tempo, the length of Gu Zhou's closing line, the number of handoffs, and where the silences sit in this episode's score.

A template's value is preserving constraints and checks, not copying surface rhythm. If every episode runs hand over the document, look up, cut to black, the system will reliably manufacture fatigue.

74.4 The postmortem meeting, the end-to-end checklist, and final deliverables#

A complete postmortem meeting.

The meeting reviews the greenlight assumptions against evidence, then budget variance and quality escapes, then the automated recommendations that were overruled, and finally decides what enters the rule library, the experiment pool, or the discard list. Speaking order runs from fact to interpretation, so the most senior person does not characterize the result first.

Every conclusion belongs to one of five categories: an immediate process change; an experiment needing more evidence; a decision applicable to this episode only; something falsified and stopped; or no conclusion yet. By the end, every change is in the task system rather than only in a deck.

End-to-end replay checklist.

• The business model and audience hypothesis carry evidence and stop conditions.
• The episode contract defines this episode's payoff and its new debt.
• Facts, character knowledge and audience knowledge are locked separately.
• Characters, locations, wardrobe and props passed stress tests.
• Every shot has a responsibility, a state interface, a risk and a fallback route.
• Generation attempts changed controlled variables only, with cost recorded.
• Voices, music themes and acoustic space are continuous across shots.
• Edit, subtitles, graphics and grading all trace to a source of fact.
• QC issues returned to the earliest deviating layer, with regression by impact.
• The release carries rights, technical, approval and checksum evidence.
• Any shot in the finished cut traces to its task and input versions.
• Postmortem conclusions entered rules, experiments or a stop list.

Final deliverables.

This run produced: the complete E001 episode pack; the locked asset library; shot and sound project files; five-layer QC and the rights snapshot; the platform release package; the cost ledger; the event log; the human decision log; evaluation calibration samples; the scale gate for the first ten episodes; and the system improvement tasks.

Replaying this yourself does not require copying the subject matter. It requires being able to deliver the same complete chain of evidence. Real capability in commercial AI microdrama is not happening to produce one good episode — it is knowing why it works, how to reproduce it, where it will fail, and how to return to the right workstation when it does.

A note on sources#

The figures in this postmortem are a teaching reconstruction. What transfers is the structure: trace every finished second to its inputs, separate cost from spend, treat human overrides as calibration data, and end with changes that someone owns.