Chapter 73. QC and Release: Every Problem Must Return to the Right Workstation#

Episode one had no single point of QC where a final reviewer watched it through on instinct. The team checked five layers — story, continuity, audiovisual, technical and rights — and bound every problem to a severity, evidence, a root layer and a rework route.

73.1 Five-layer QC, issue records, and rework routing#

The order of the five layers.

The story layer checks whether the audience understands that Lin Xia was struck off, why she returns, why the document changes the power, and what question the closing line raises. The continuity layer checks identity, wardrobe, props, axis, knowledge and time. The audiovisual layer checks performance, lip sync, music, spatial sound and subtitles. The technical layer checks resolution, frame rate, color space, loudness, bitrate and file integrity. The rights layer checks images, voice, music, fonts, client material and synthetic content labelling.

If the story layer does not hold, do not spend time repairing a background reflection frame by frame. If rights block, technical perfection cannot ship.

An issue record.

issue:
  id: issue_e001_042
  detected_at: 00:58.440
  layer: continuity
  severity: blocker
  observation: after the handoff, the shot still shows the folder in Lin Xia's left hand
  expected: holder_gu_zhou
  evidence: [frame_00_58_440.png, event_evt_e001_s014_gu_takes_folder]
  likely_root_layer: edit_selection
  route_to: editor
  proposed_fix: replace with the end section of take_015_04; no regeneration required
  regression_scope: [shot_015, shot_016, prop_state_timeline]

An issue describes observable fact, never "this looks odd." The initial root-cause guess may change; the observed evidence must stay stable.

Three real rework routes.

The first problem was the folder's holder teleporting. The generated shot had a correct candidate, and the edit had used a more attractive take with the wrong end state — so it routed to the editor, with no regeneration.

The second was Gu Zhou's closing line arriving about 140 milliseconds late in the mouth. The audio was natural and the picture had sufficient handles, so it routed to lip sync and micro-timing. The line was not rewritten.

The third was the contract amount reading incorrectly under OCR. The root cause was glyph deformation during graphic tracking, so it routed to the graphics department to adjust perspective and motion blur. Regenerating the contract shot would only have widened the uncertainty.

73.2 Severity, regression sets, and the rights snapshot#

Severity and the release gate.

blocker covers factual errors, obvious identity changes, missing rights, corrupt files and unintelligible key dialogue. major covers noticeable lip sync problems, spatial jumps, damaged pacing and subtitles covering critical content. minor covers local blemishes that affect neither comprehension nor brand. Severity follows audience and commercial consequence, not how hard the repair is.

A release candidate requires zero blockers; zero majors, or a signed time-limited waiver; and an explicit acceptance record for minors. Waivers may never cover missing licences, wrong amounts, or a platform's hard technical specification.

Regression is not watching it again.

After repairing the folder shot, the regression scope covers the action interfaces of the shots either side, the prop state, the music entry and the total duration. After repairing the contract graphic, check the amount, the date, the four-corner tracking, motion blur, subtitle occlusion, and every derived campaign version. Impact analysis determines the regression set, which avoids both blind full-episode review and checking only the edit point.

Automated checks cover file checksums, black frames, frozen frames, peaks, loudness, subtitle overflow, OCR, missing fonts and manifest completeness. Human checks handle story comprehension, performance nuance, rhythm and brand judgment. An automated composite score never replaces the gate rules.

The rights snapshot.

When the release package freezes, a rights snapshot is created: characters are synthetic with no dependency on a real person's likeness; voices use synthetic profiles licensed for the project; the music theme was commissioned with commercial all-media scope; the font licence covers video embedding; the rain material's source and licence are archived; the client's corporate mark is used only in approved versions.

Each entry records territory, platform, term, whether adaptation for campaigns is permitted, and the checksum of the evidence file. Writing only "licensed" cannot support later multilingual versions, additional platforms or re-cuts.

73.3 Release operations, acceptance evidence, and the red-team drill#

The release day runbook.

The release owner generates the platform file from a read-only RC and verifies title, episode number, cover art, subtitles, synthetic content labelling and the visible area. A second person independently re-checks the checksum and the platform preview. Publish to a restricted test scope first, confirm playback, sound, subtitles and data events, then widen traffic gradually.

Rollback triggers include a wrong file, missing critical subtitles, a rights complaint, platform transcoding anomalies and the wrong factual version. A rollback is not deleting source assets: it stops distribution, preserves evidence, switches to the previously approved version and opens an incident record.

The acceptance evidence package.

The client receives more than a review link: the RC video, the textless version, subtitles, the five-layer QC report, the closed issue list, accepted exceptions, the rights snapshot, the technical report, the version differences, checksums and approval signatures. Acceptance notes must locate a timecode and a target; "make the whole thing more premium" does not enter rework directly.

SOP and deliverables.

Check in the order of the five layers. Register issues in structured form. Find the root cause at the earliest deviating layer. Choose the lowest-cost route that does not lower the standard. Regress by dependency. Freeze the RC and the rights snapshot. Have two people verify the release. Validate at small traffic. Widen distribution. Retain post-release evidence.

Deliverables: qc_report.json, issues.csv, regression_sets.yaml, rights_snapshot.json, release_candidate_manifest.json, technical_report.pdf, approval_signatures.json, release_runbook.md, rollback_record_template.md.

A red-team release drill.

Before the real release, copy an RC and deliberately plant six problems: a wrong amount, mirrored handedness, a missing subtitle, an expired music licence, loudness over specification, and an outdated cover. Have someone who was not involved in production follow the runbook, and record which problems automation caught, which a human caught, and which escaped. Escapes are written back into the gate — never left as a reminder to be careful next time.

Then simulate three incidents: audio collapsing to mono after platform transcoding, subtitles covered by the bottom buttons, and the wrong episode uploaded. The person on duty must be able to find the steps to stop distribution, confirm the version, roll back and notify owners without consulting the creative leads. The product of an effective drill is not that it succeeded; it is a clear record of detection time, decision time, recovery time, evidence gaps and the runbook changes made.

Afterwards, have a different owner verify that the revised gate really does stop the same class of error. Adding words to a checklist without an automated test sample, an owner and evidence of execution does not close the problem.

A note on sources#

The QC pass here is reconstructed to show routing rather than to report real defect rates. What transfers is checking in layer order, describing issues as observable facts, and letting impact analysis — not habit — define the regression set.