Chapter 52. The Five Layers of QC#

52.1 Quality problems must be classified first#

One symptom can arise from several layers. When a viewer says it is boring, the cause may be a weak story objective, shots with no consequence, flat voice performance, unvarying music, or repetitive cutting. Only after classifying can you find the responsible station.

52.2 Story QC#

Check comprehension in the first three seconds, character objective, pressure, strategy, power change, repayment of old debt, the new question, consistency with the promise, and honesty at the cut point. Evidence comes from cold-read retellings, the beat sheet, entry and exit states, and audience testing.

"I personally dislike the genre" is not a story defect. Noting that the protagonist has no visible strategy before E001_B03 is actionable.

52.3 Continuity QC#

Check the seven dimensions: identity, wardrobe, space, props, action, light and time, emotion and knowledge. Use golden assets, the state ledger and adjacent shots rather than memory.

Catch continuity problems at the still and small-batch generation stages rather than waiting for a finished episode.

52.4 Audiovisual technical QC#

Check deformation, flicker, black frames, frame rate, color, graphics, subtitles, sync, lip sync, noise, loudness, true peak, encoding and platform transcoding. Technical correctness does not mean the story works, and any blocking technical error can destroy the viewing.

52.5 Commercial QC#

Check target audience, the hook's promise, whether the episode matches it, the first paywall, post-payment repayment, ad versions, CTA and the metric path. High click-through does not pass automatically; look at episode retention and the post-conversion experience.

52.6 Compliance QC#

Check IP, likeness, voice, music, fonts, material provenance, platform rules, generation labelling, sensitive content and contract scope. Unclear rights is a P0 and cannot be offset by high aesthetic scores.

52.7 Five ways of watching#

Muted verifies action, subtitles and proof. Eyes closed verifies relationships, voices and the sound field. Frame by frame verifies technique and continuity. Normal speed on a phone verifies perception. A cold viewer verifies comprehension and commercial effect. Each answers a different question.

52.8 Staged QC#

Script, assets, shot design, keyframes, motion, rough cut and master each have a gate. The earlier something is found, the cheaper the repair. Final QC is not the first review; it confirms that each layer already passed and that the combination produced no new problems.

52.9 Severity#

P0 blocks the project or the release. P1 blocks the current gate. P2 must be fixed before scaling up. P3 is polish. Severity follows audience impact, propagation and risk — not how easy the repair is.

52.10 Separate the object under review from the questions asked#

Do not review everything at one stage against a single overall-impression sheet. The object at the keyframe gate is a still image, with questions concentrated on identity, state, composition and movability. The object at the rough cut is a scene sequence, with questions on causality, knowledge, performance, rhythm and audio continuity. The object at release candidate is an immutable set of files, adding rights, metadata and the platform package.

Each QC pass first writes artifact_type, version, gate, reference assets, and what is explicitly not being reviewed this round. The keyframe gate does not judge the final mix; the rough cut may use temporary music, marked as temporary. Clear boundaries stop review meetings from raising notes the current stage cannot resolve.

52.11 What a usable issue record contains#

An issue record carries at minimum location, observed fact, the standard violated, the audience or production impact, severity, evidence, the suspected earliest responsible layer, and how it will be re-verified.

issue_id: QC_E003_027
artifact: ROUGH_E003_V08
timecode: 00:41.12-00:42.06
observation: Lin Yun set the pen on the table in the previous shot; her right hand holds it again here
violates: CONTINUITY_PROP_OWNERSHIP
impact: an uncaused state jump immediately before the key evidence action
severity: P1
evidence: [FRAME_1012, FRAME_1034, STATE_E003_B06]
suspected_root_layer: motion_generation
verification: after repair, play two shots either side and confirm one pen position only

"Something feels off," "the character is inconsistent" and "the pacing is bad" cannot enter a rework queue. A reviewer need not prove the root cause, and must describe the symptom so another owner can reproduce it.

52.12 Automated checks, expert checks and cold-viewer checks#

Automation suits completeness and consistent rules: missing files, resolution, black frames, loudness, subtitle overflow, referenced versions, hashes, duplicate frames and some visual drift. Human experts judge performance, space, narrative causality, audio intent and aesthetics. Cold viewers verify whether comprehension, attention, emotion and the promise actually arrived.

None of the three can impersonate another. A similarity model rating two faces as close does not mean the audience believes it is one person. A director considering the evidence clear does not mean a first-time viewer read the amount. The system stores all three kinds of evidence and knows who arbitrates when they conflict.

52.13 Full inspection and sampling#

P0/P1 risks, hero shots, key evidence, first appearances of assets and new production routes are inspected in full. Low-risk background shots on a mature route can be sampled per batch, widening the scope when a sample fails. The sampling ratio follows historical defect rates rather than being cut when the schedule tightens.

When a batch reveals a systemic error — a wrong voice card, an expired authorization template — do not fix only the sampled instances. Query every dependency. Sampling saves repeated confirmation; it does not exempt you from systemic tracing.

52.14 Triage across five layers#

Pilot review returned the note that the authorization reversal did not land. Splitting it across the five layers showed: at the story layer the rules were established; at the continuity layer the document's position was correct; at the technical layer the title held too briefly to read on a phone; at the commercial layer the ad had already shown the full identity, weakening the surprise; compliance was clean. The team therefore extended the proof insert, adjusted subtitle avoidance and re-cut the ad — rather than rewriting the scene or generating a more exaggerated expression.

One vague note became two fixable problems, and the story and performance that were already right were preserved.

52.15 SOP for QC#

First, establish the current artifact and gate. Second, load versioned references. Third, run automated technical checks. Fourth, review by hand across the five layers. Fifth, spot-check using the five viewing modes. Sixth, record evidence, severity and root-cause candidates. Seventh, block on P0/P1. Eighth, regress the relevant scope after repair.

52.16 Fault tree#

Symptom: the notes contradict each other. They belong to different layers and goals. Record story, technical and commercial separately rather than mixing them in one verdict.

Symptom: problems explode at final review. There were no upstream gates. Move QC earlier, to assets, keyframes and the rough cut.

Symptom: automation is all green and it still looks bad. Automation covers format and some anomalies only. Add human judgment on performance, emotion and commerce.

52.17 Checklist, exercises and deliverables#

Check that problems are layered; that references are versioned; that all five viewing modes are used; that severity follows impact; that P0/P1 block; and that repairs are regressed.

Exercise one: sort twenty vague notes into the five layers. Exercise two: produce muted, eyes-closed and phone reports for one episode. Exercise three: define one blocking item for each production stage.

Deliverables for this chapter: the five-layer QC report, viewing-mode reports, the severity matrix, the gate decision, and regression results.

52.18 The review matrix: object, question, evidence, owner#

A QC plan puts objects in rows and the five layers of question in columns. The script pack is reviewed for causality, promise and compliance. Character assets for identity, rights and production stability. Shots for continuity, technique and legibility. The episode for rhythm, commercial cut points and viewing experience. The release package for files, licences and platform requirements. Not every object needs everyone to look at it again.

Each cell defines the question, the evidence of passing, the sampling ratio, the owner and the gate. Commercial QC on an evidence insert asks whether a first-time viewer can read the key amount, evidenced by the phone test record. Compliance QC on a music asset asks whether the licence covers advertising and the target territory, evidenced by the rights record. A check with no evidence format eventually degenerates into verbal opinion.

An owner may delegate the checking and never the final judgment. Automated tools report technical anomalies, legal confirms rights, the director confirms performance, the producer decides waivers and cost. A composite score cannot override any blocking item.

52.19 The division of labor between cold viewers and professional QC#

A professional team knows the correct answer and will auto-complete missing causality. Cold viewers do not know the script and will find confused characters, evidence they never read, misunderstood space and cut points with no pull — while not being responsible for diagnosing a specific model or a compliance root cause. Test questions should ask for retelling and prediction rather than whether they liked it.

Ask what the protagonist is trying to do now; what the authorization proved; who controls the meeting; what they think happens next; and where their attention wandered. Compare the answers against the design intent, then let the professional team route the problems. One viewer misunderstanding does not change the cut; several independent viewers misunderstanding at the same point is a strong signal.

Record the version tested, the sample conditions and the order of questions. Explaining the plot before testing contaminates the result, and passing on a large screen in a quiet room does not substitute for phone viewing on a commute.

52.20 QC noise, false positives and quality debt#

Automated checks and multiple reviewers generate noise. When the false-positive rate is high the team starts ignoring all alerts; when the standard is too strict, time goes into one finger inside motion blur while a critical story problem has no owner. Track per rule the confirmation rate, misses, average handling time and the actual blocking value.

Unrepaired problems enter a quality debt ledger recording the reason, the audience risk, propagation, the batch by which it expires and the owner. If P2 waivers accumulate into visible style drift ten episodes later, escalate to a systemic fix. Local waivers cannot be treated as unrelated to one another.

QC health metrics are issue reopen rate, recurrence by root cause, average rework radius and gate escape rate — not how many problems were found. A good system gradually reduces repeated defects so review attention returns to new creative risks.

A note on sources#

Human review is a permanent station in AI microdrama production. The value of layering is preventing every problem from being handed, incorrectly, to "generate it again."