# Editor v2 RFP — FINAL judgment (3 of 4; deepseek a no-show)

deepseek: 5 launch attempts, all died to infra (latest: lease expired at seq 17, no output). Excluded.
In hand: nemotron (19/25 preliminary), glm53 (24/25 preliminary), qwen38 (scored below).

## qwen38 (qwen3.8-27b) — assessment

Strengths:
- Landscape-native schema from v1 (`aspect: '16:9' 720p | '9:16' 1080×1920`, `fit: cover|contain`) —
  satisfies the mandatory amendment glm53 needed, natively.
- Captions are schema-first: Speaches faster-whisper large-v3-turbo (correct model per research),
  words JSON as source of truth, SRT/VTT/ASS derived, ASS burn-in re-timed through the EDL.
- Container generalization is disciplined: the in-flight produce payload becomes a degenerate
  EdlProject → ONE validator, ONE render path; version discriminator + version-bump policy as the
  explicit anti-NLE-creep defense.
- Job model reuses the in-flight producer lease pattern (`kind='assemble'`, SKIP LOCKED).
- Render: one process / one filter_complex / one encode, measured 1.2× RT; stream-copy honestly
  rejected inside assemblies; cards as loop-PNG from the existing overlay renderer — no drawtext
  at all; pre-spawn trim validation against probe durationMs; retry/timeout/scratch-cleanup policy.
- AI: Speaches + SigLIP + job runner on gpu-01 over Tailscale, worker stays pure FFmpeg; VRAM budget
  table; license register incl. BLOCKED list (XTTS-v2 CPML, F5-TTS CC-BY-NC, RVM GPL-3, Ultralytics AGPL-3);
  auto-editor as SUGGESTED cuts with human confirm.
- UI: only new files, primitives named as reused, mobile single-column with chevron reorder, ≥44 px
  targets, one-thumb flow, 8 documented shortcuts, exact brand tokens.
- Risk honesty: 5 risks with mitigations; "preview ≈ final"; measured-vs-EST labeling.

Weaknesses: stale migration number (036 vs actual next); no explicit schema-drift guard; no
render-CPU-contention semaphore; thinner canvas/revision formalism than glm53.

## Final scoring (1–5 per criterion)

| Criterion | nemotron | glm53 | qwen38 |
|---|---|---|---|
| Correctness vs real constraints | 3.5 | 4.5 | 4.5 |
| Concreteness (files/jobs/models) | 4 | 5 | 5 |
| Speed-to-value for crews | 4 | 4.5 | 4.5 |
| Look-and-feel respect | 4 | 5 | 5 |
| Risk honesty | 3.5 | 5 | 5 |
| **Total** | **19/25** | **24/25** | **24/25** |

## Verdict

glm53 and qwen38 tie at 24/25. Tie-break on the deltas: glm53's only mandatory amendment was a SCHEMA
change (portrait-only canvas); qwen38's gaps are ADDITIVE process guards. 

**Decision: adopt qwen38 as the editor-v2 plan base, with two glm53 amendments folded in:**
1. glm53's mirror-file + mirror-assert convention for the EDL schema (API↔worker drift guard).
2. glm53's render semaphore (≤2 concurrent niced renders) so batch renders never starve the live compositor.

nemotron: not adopted (wrong transcription stack, captions missing, matting confusion).
deepseek: no-show; may be re-run later for a second opinion but is not blocking.

## Implementation status
- S1 EDL schema + /video/edl API + 043 store: SHIPPED (media-control/edl.ts, EdlController, 043)
- L1 assemble lane: SHIPPED (044 claim widening, runner probe+assemble lanes, assemble-builder,
  overlay-png cards, ffprobe port) — e2e verified (2 clips + card + overlay → derived asset)
- M1 editor UI: queued
- M2 transcription lane: sidecar deployed (Speaches-compatible faster-whisper on moxexleyhp1-1 GPU,
  http://100.70.240.7:9000) — hub wiring pending
- v1.5 search (SigLIP): sidecar deferred — TEI 1.8 cannot parse google SigLIP configs; needs a newer
  TEI or the planned FastAPI wrapper