PHOENIX OMEGA
Tech-lead onboarding

Video

CODE-WIRED

The beat-driven route is wired and regression-tested, but the supplied status evidence does not show a new end-to-end GPU render with operator acceptance.

What this pipeline produces

A final MP4 whose visuals match actual narration timing, with word alignment, beat manifest, frame provenance, audio mix, and operator-reviewed QA strip.

Where it really stands

Acceptance layer: CODE-WIRED

Evidence: scripts/video/run_pipeline.py

The beat-driven route is wired and regression-tested, but the supplied status evidence does not show a new end-to-end GPU render with operator acceptance.

Run it yourself

The context verifies this reference builder and that it renders, assembles, and mixes the existing ahjan_update v3.1 package. Rendering new frames requires Pearl Star ComfyUI.

python3 scripts/video/build_v3_1_yt_starseed.py
REQUIRES-LIVE-RUN
python3 scripts/video/run_pipeline.py --help

If help shows --pipeline-mode beat_driven and the required source/output flags, capture those flags and run one new artifact. If not, reconcile the live CLI with the landed route before claiming it is reachable.

The map

Verified path in supplied contextRole
scripts/video/run_pipeline.pyHigh-volume orchestrator
scripts/video/v3_1_beats.pyReference beat manifest
scripts/video/build_v3_1_yt_starseed.pyReference render/assemble builder
scripts/video/whisper_word_align.pyWord timing
scripts/video/build_frame_selector_v2.pyLocal-only best-of selector
scripts/video/assemble_v3_8.pySelection consumer
artifacts/video/yt_starseed_ahjan_update_20260610/Reference artifacts
config/video/render_params.yamlRender configuration

What is broken right now

landing

A real GPU-backed run and operator-reviewed artifact are still required to move beyond CODE-WIRED.

landing

The frame selector is local-only and hardcoded; its paths and version map are not generalized.

credential

New beat renders need the authorized image backend; the context routes NIM by default and Pearl Star only by explicit request for the current manga work, while the video reference spec itself names Pearl Star ComfyUI.

The traps

  • Do not time visuals from average WPM. Use word-level alignment; emotional delivery varies inside one recording.
  • Do not use stable_whisper on the documented macOS Python 3.9 setup; the authority records an ABI failure and uses vanilla whisper word_timestamps.
  • The final ffconcat manifest needs the terminating file reference, and final mixing must use -shortest.
  • A wired route and passing regression tests remain CODE-WIRED until a real artifact is rendered and reviewed.
  • A gate PASS is at most structurally clear; never collapse it into an unlayered “all checks pass” claim.

Who to ask

Pipeline owner

Pearl_Video

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

1
Run the reference builder and preserve the final MP4, timing JSON, concat manifest, and QA strip.
2
Scrub every beat against narration and repair one mismatch surgically through the beat manifest.
3
Verify the final manifest terminator and -shortest audio mix behavior.
4
Run the generalized entrypoint help check, then document the exact real-input invocation from live CLI output.