Tech-lead onboarding

Manga

EXECUTED-REAL

No manga sample has a landed blind-judge scorecard. The current composited episode has 35 panels and zero passing L2 validations, so it has not reached PROVEN-AT-BAR.

What this pipeline produces

A byte-verified, composited manga episode: authored panels assembled into final page images, with lettering, QC, and continuity memory—not merely a series plan.

Where it really stands

Acceptance layer: EXECUTED-REAL

Evidence: artifacts/reports/production_status_20260806/FINDINGS.md

No manga sample has a landed blind-judge scorecard. The current composited episode has 35 panels and zero passing L2 validations, so it has not reached PROVEN-AT-BAR.

Run it yourself

REQUIRES-LIVE-RUN

python3 scripts/manga/run_series_setup.py --help
python3 scripts/manga/run_manga_chapter.py --help

If both commands expose required inputs and output directories, record those exact flags and run series setup before the resumable chapter DAG. If either command is absent or its help conflicts with the authority docs, stop and reconcile the live CLI before attempting production.

The supplied context verifies these entrypoints and that the chapter runner supports --from-stage / --to-stage; it does not provide a verified full invocation for a real series.

The map

Verified path in supplied context	Role
`specs/AI_MANGA_PIPELINE_SUMMARY.md`	System and artifact contracts
`docs/MANGA_IMPLEMENTATION_OUTLINE.md`	Operational bridge
`config/manga/gate_registry.yaml`	QC gates
`scripts/manga/run_series_setup.py`	Series setup CLI
`scripts/manga/run_manga_chapter.py`	Resumable chapter DAG
`scripts/manga/assemble_from_bank.py`	Layered assembly
`config/source_of_truth/manga_series_plans/`	Series-plan SSOT
`config/source_of_truth/manga_book_plans/`	Episode-plan SSOT

What is broken right now

landing

35 composited panels, zero passing L2 validations; re-render head-and-shoulders or medium-shot L2 assets on NIM and byte-verify from R2, not the TSV.

operator-action

Blind judging still needs at least three judges and real scorecards.

The traps

A RENDER_PROGRESS.tsv row marked ok/done with bytes under 50,000 is a stub, not a panel. Render the asset; never edit the byte column.
A series_plan is catalog metadata, not an authored story. Rendering requires a chapter_script_writer_handoff with at least six authored panels and no stub markers.
An INTERIM layer is a labelled stand-in and is never final art. Layered assembly must preserve REAL|INTERIM provenance.
A gate PASS is at most structurally clear; never collapse it into an unlayered “all checks pass” claim.

Who to ask

Pipeline owner

Pearl_Dev

Coordination

Pearl_PM owns scope, sequencing, and status claims through artifacts/coordination/ACTIVE_WORKSTREAMS.tsv.

Architecture

Pearl_Architect owns architecture decisions and cap rulings through docs/PEARL_ARCHITECT_STATE.md.

Your first week

Reproduce one replay-backed chapter through the resumable DAG and retain every stage manifest.

Byte-check every emitted panel and page; reject missing images and any ok/done ledger row below 50,000 bytes.

Re-render the failing character L2 shots on NIM, then land evidence for at least one passing validation.

Run a three-judge blind review on the repaired sample and land the scorecards.