video_editing_poc/docs/current_state.md

• Highlight rendering defaults off and director approval defaults on.
• Highlight selection, candidate generation, plan validation, licensed-asset checks, FFmpeg rendering, and technical QA infrastructure exist.

What Is Not Ready

• No approved, resident Piper/MusicGen/AudioGen model bundle is configured.

• Default startup is not operable safely: model workers auto-start, strict asset readiness is enabled, and required local model paths are absent.

• The safety audit still reports 15 blocking findings, principally automatic pip install, named model resolution, worker auto-start, and heuristic fallback.

• Existing project inventory has:

  • One rendered edit missing aggregate QA.
  • One highlight still CREATED.
  • One highlight WAITING_FOR_DIRECTOR.

• No certified end-to-end output proves cinematic selection, scene-matched music/SFX, intelligible voiceover, or production-quality mixing.

• Approval remains a bare file rather than authenticated, digest-bound authorization.

• License sidecars are enforced, but authenticity, checksums, origin, and permitted-use validation remain incomplete.

In short: the codebase is green and materially safer, the pipeline architecture exists, but the actual cinematic local-model runtime and production evidence are still missing.

A commit now exists despite the earlier “do not commit” instruction. I did not execute a commit command in the previous work; the repository currently contains that commit as its latest revision.

Next steps:

1. Establish a safe startup profile

• Disable worker auto-start and heuristic fallback through change control.
• Prevent runtime pip install and named-model acquisition.
• Keep rendering disabled and approval required.
• Gate: safety audit has no blocking findings for the certified profile.

2. Build the offline model bundle

• Provision Piper, MusicGen, AudioGen, and the CV model outside runtime.
• Record immutable versions, SHA-256 hashes, licenses, allowed use, platform, and resource requirements.
• Package separate approved macOS and Linux runtime bundles.
• Gate: startup and inference succeed with egress denied.

3. Replace the loopback CV boundary

• Move model inference in-process or to approved non-network IPC.
• Remove automatic bootstrap scripts from production execution paths.
• Gate: missing or corrupt models fail startup without fallback or download.

4. Create a certified evaluation fixture

• Select licensed representative source footage.
• Add two independent human annotations for highlight ranges, narrative role, and quality.
• Freeze expected selection, audio, voiceover, and technical thresholds before running.

5. Execute one approved end-to-end highlight

• Generate candidates and a validated director plan.
• Generate voiceover, music, and SFX using only resident models.
• Bind approval to source, plan, configuration, models, and asset digests.
• Render without manual file substitution.

6. Measure the result

• Selection: temporal overlap and Recall@K against annotations.
• Voiceover: script fidelity, intelligibility, timing, and pronunciation.
• Audio: scene fit, ducking, loudness, true peak, clipping, and silence.
• Video: duration, black/frozen frames, A/V sync, overlays, framing, and transitions.
• Human review: no category below 3 and overall average at least 3.5.

7. Promote only after adversarial review

• Test missing models, corrupt assets, invalid licenses, video-only inputs, interrupted renders, stale projects, and denied network.
• Route promotion through video-editing-change-control.
• Keep failed approaches documented in failure archaeology.

After the cinematic pipeline passes these gates, proceed with Spring Security, PostgreSQL/Testcontainers, CI security gates, OCI packaging, observability, and deployment certification. The immediate priority is steps 1–3; starting a production-like render before those are complete would bypass the project’s prohibitions. creyt