Chapter 53. Fault Trees and Rework Routing#

53.1 A symptom is not a root cause#

"The face changed" may come from contradictory identity assets, a missing angle reference, a wrong keyframe, motion that ran too long, lip sync damage, or a mismatched composite. Regenerating the video directly only helps with some of those.

53.2 Trace to the earliest deviating layer#

Check in order: facts and brief, script, assets, state, shot design, keyframes, motion, compositing, editing, output. Find the first layer that departed from its approved input. Downstream errors are usually amplifications of an upstream one.

Figure 53-1 The fault triage tree. Character drift starts by checking whether the still identity was already wrong. Action that does not connect starts by comparing handoff phases. Text errors return to the authoritative source of fact. Inaudibility starts by judging the dry recording itself. Re-accept after repair; repeated failure escalates to a systemic issue.

The yellow questions in the diagram are triage points, not automatic conclusions. The same symptom may keep tracing upward: if the still is already wrong, you still have to decide whether the fault is in the golden asset, the angle reference, contaminated state, or the keyframe. Only when the dry recording is clear and it is still hard to hear do you move to orchestration, frequency, ducking and playback device. The value of triage is stopping every problem from landing on "generate it again."

53.3 The identity fault tree#

If the still is already wrong, check golden references, angle, wardrobe and prompt. If the still is correct and motion breaks it, shorten the action, reduce the expression, or change route. If only the composite is wrong, fix the matte, light and color. If only one edit version is wrong, check which source take was used.

53.4 The pacing fault tree#

If it is slow at cold read, return to beats and scenes. If it is slow at storyboard, check for repeated shots. If it is slow with scratch audio, check dialogue and pauses. If it is slow at rough cut, delete process and repeated reactions. If it only becomes slow after music, check cue duration and drops.

53.5 The comprehension fault tree#

Ask first whether the entrance promised something mismatched. Then whether relationships and goals were established. Then whether the rules of evidence are legible. Then whether subtitles and audio are synchronized. Then whether the edit removed necessary establishing. Do not answer everything by adding narration.

53.6 The audio fault tree#

For inaudibility, separate recording quality, articulation, performance, frequency conflict, loudness and device. Simply pulling the music down may still leave sibilance or a wrong name unintelligible.

53.7 The economics of rework#

Among several viable repairs, choose the route with the lowest future cost that does not damage the promise. Local repair suits local problems. Changing the shot design can eliminate several failures at once. Rewriting the script is expensive and may resolve a requirement that is fundamentally unproducible.

repair_decision:
  issue: ISSUE_E006_019
  symptom: two characters jump position after the blackout
  root_layer: motion_generation
  options:
    regenerate_motion: cost_3_risk_2
    insert_phone_closeup: cost_1_risk_1
    rewrite_scene: cost_5_risk_3
  choice: insert_phone_closeup
  reason: preserves state and suspense, and the insert carries information of its own

53.8 Regression scope#

Fixing a subtitle needs only the related time range and the other-language versions. Replacing a character's golden image may affect every keyframe. Changing a fact field affects graphics, dialogue, subtitles and advertising. Use the dependency graph to determine the minimum sufficient regression.

53.9 Systemic problems#

The same class of P2 recurring should be escalated to a systemic defect. When twenty document shots have drifting text, stop fixing them one by one and change the shot template and the graphic compositing route.

53.10 Find the first deviation by comparing versions#

Root-cause analysis is not a discussion of who made a mistake. It compares approved inputs against actual outputs along the dependency chain. Is the fact field correct? Does the script reference the correct fact? Does the shot design preserve action and state? Does the keyframe match the shot design? Does the video preserve the keyframe? Does the composite reference the right graphic? Does the edit use the approved take? The first layer that deviated is the primary rework entry point.

Downstream may also have produced secondary problems. A wrong keyframe leaves an extra document in the character's hand, the edit covers it with a reaction shot, and the subtitle therefore moves onto a face. Fixing only the subtitle does not restore the document state, and regenerating only the video does not automatically remove the editorial patch. Issue records distinguish the root defect from secondary defects.

53.11 Compare repair options on four axes#

Compare direct cost, probability of success, dependency propagation and loss of promise. The fastest repair is not always the cheapest overall: cutting away from a changed face takes minutes and may delete the antagonist's reaction to seeing the evidence. Regenerating is more expensive and preserves the story function. Changing the shot design may eliminate ten similar failures at once.

repair_option:
  action: replace_with_document_insert
  direct_cost: 1.5h
  success_probability: 0.9
  affected_assets: [edit, foley, subtitle_position]
  promise_loss: none
  new_risks: [reading_time]
  rollback: restore_ROUGH_E003_V08

Rework decisions record future cost, and never let money already spent drive the choice.

53.12 Impact analysis and the regression set#

Before repairing, query direct and semantic dependencies. Direct dependencies come from asset references — a subtitle referencing an audio hash. Semantic dependencies come from a shared story fact: even with no filesystem link, the amount in an advertisement must stay in sync. Both enter the impact list.

The regression set has three rings. The first verifies the problem itself. The second verifies adjacent shots, sound and state handoffs. The third verifies every distant reference sharing that asset or fact. After fixing a document's text, check not only the current insert but the handheld version, later flashbacks, acquisition creatives and other-language versions.

53.13 When to escalate to a systemic defect#

Escalate when any of these hold: the same root cause recurs across more than three shots; a P2 appears in two batches; one error affects several episodes; manual repair takes longer than updating the template or route; or the error stems from a missing field, permission or gate. A systemic defect needs an owner, a preventive measure, and a scan of historical assets.

Closing a systemic defect needs three kinds of evidence: the root rule or asset has changed, affected existing material has been handled, and subsequent new batches pass regression. Writing a lessons-learned note does not prevent recurrence.

53.14 The root cause behind the face changes#

Lin Yun's nose bridge and jaw drifted across six shots of her looking down at a document. Initially each shot was sent for regeneration on its own, with a low success rate. Version comparison showed the frontal golden reference was complete while the downward-angle pack contained only one heavily retouched image. Deviation had already begun at the keyframe stage; motion merely made it obvious.

The team stopped retrying video, produced a neutral-light downward reference, updated the angle authorities, and scanned every keyframe referencing the old image. Two of the six shots were repaired locally, three had keyframes redone, and one was waived on the record as a wide. The repair happened at the earliest deviating layer, which also stopped the retry cost recurring in later episodes.

53.15 SOP for fault routing#

First, describe the symptom and its evidence precisely. Second, find the earliest deviating layer. Third, list the repair options. Fourth, compare cost, risk and propagation. Fifth, choose the responsible station. Sixth, define acceptance and regression. Seventh, count recurring root causes and change the system.

53.16 Fault tree#

Symptom: repaired many times and it keeps recurring. Only the presentation layer is being fixed. Look for the shared upstream asset or rule.

Symptom: one change breaks the whole episode. No impact analysis was done. Restore the previous version and migrate in batches by dependency.

Symptom: the team blames each other. The problem has no first deviation point and no evidence. Compare versioned inputs.

53.17 Checklist, exercises and deliverables#

Check that symptoms are specific; that the root cause is the earliest layer; that repair options were compared; that the responsible station is explicit; that acceptance and regression are written; and that recurring problems are escalated.

Exercise one: draw fault trees for face changes, slowness and inaudibility. Exercise two: compare regenerating, inserting and rewriting. Exercise three: compute the regression scope for one character asset change.

Deliverables for this chapter: the fault tree, the root-cause report, repair options, the route decision, the regression scope, and the systemic issue log.

53.18 Choose repairs by expected total cost#

The cheapest single repair is not necessarily the most economical. Compare each option's direct cost, probability of success, retry cost on failure, scope of impact, and probability of introducing new errors. Covering a changed face in the edit takes ten minutes and may cost a critical reaction. Redoing the keyframe costs more and preserves the performance while reducing recurrence.

repair_option:
  option: regenerate_keyframe_and_motion
  direct_cost_minutes: 95
  pass_probability: 0.78
  retry_cost_minutes: 55
  regression_radius: adjacent_2_shots
  narrative_loss: none
  new_risk: hand_contact_medium

The producer selects the lowest expected total cost that meets the quality threshold — not the option the technical staff finds most interesting. If a simpler design can carry the shot's function, restaging may reduce cost and risk together. For a hero shot at a paywall, preserve the narrative value and invest in the repair.

53.19 Repair recipes must record their preconditions and failure boundaries#

Teams say a face swap fixed it last time, or that a sound bridge will do — and forget last time's angle, lighting and shot function. A repair recipe records the symptoms it suits, the confirmed root cause, input conditions, the steps, acceptance criteria, prohibited conditions and typical side effects.

Horizontal flipping applies only to shots with no text, no asymmetric markers, no handed props and no directional light source; failing any one of those disqualifies it. A sound bridge can connect a spatial elision and cannot prove a character completed an action the audience had to see. A recipe with boundaries becomes organizational knowledge instead of a dangerous shortcut.

Record the result and cost of each application and update the success rate. A recipe that fails three times in a row is automatically demoted and requires fresh root-cause analysis. Do not keep burning budget to protect a method's reputation.

53.20 Write prevention back from a single incident#

After closing an issue, ask whether it could have been stopped at an earlier, cheaper station. Repeated profile drift means adding character angle assets and a preflight rule. Subtitles covering evidence means updating the avoidance map. Missing music licences means blocking asset promotion. Fixing only the finished episode without changing the system is booking the next incident.

A systemic fix creates an owner, a deadline batch, a verification metric and a rollback plan. Run the new rule on a small batch first, watch for false positives and side effects, then extend it to the season. Rule changes can also harm creative freedom, so the benefit must be demonstrated.

Organize the postmortem library by root cause rather than stacking meeting notes by project date. When the next show starts, model evaluation, asset stress testing and the QC plan read the high-frequency root causes, turning experience into constraints applied up front.

A note on sources#

The randomness of AI production easily conceals structural errors. A fault tree turns "try again" into an engineering decision supported by evidence.