dramapack

Acceptance QC: grade finished short-drama episodes green/yellow/red against market benchmarks - duration (TikTok Creator Rewards needs > 60 s), average shot length, long-shot share, mean volume, seam level jumps (give splice points in `bounds`), frozen tail / end-card length, black frames (mid-episode black >= 0.5 s is red), and with faster-whisper installed: dialogue density (`speech`) and subtitle-to-speech alignment (`align`: cue offsets, cues with no speech, speech with no cue, likely typos when subtitles are in the spoken language; uses `srt` or the same-name .srt next to the file; speech=false disables both). Pass exactly one master in `files` plus its raw `cut` to grade the cut on pacing/speech/seams and the master on duration/first line/tail/subtitles. `vlm=true` adds vision-model checks on sampled frames (visual defects, platform-policy flags, first-frame hook) ONLY if the server's environment sets DRAMAPACK_VLM_BASE_URL / DRAMAPACK_VLM_API_KEY / DRAMAPACK_VLM_MODEL; frames then go to that endpoint (it may be a local model). Otherwise no network access. Lip-sync is not checked. Does not modify files.

Rights hygiene check for overseas short-drama distribution. Grades every music cue declared in rights.json (source: generated | procedural | public_domain | licensed | unknown) and every font burned into the picture. Red: undeclared/unknown music, major-label publishers (Sony / Universal / Warner and imprints), public domain published <= 50 years ago, fonts not free for commercial use. Yellow: licensed without proof, font not installed (substitute is checked too). Give `rights` (path to rights.json) and/or `input_dir` (uses its rights.json, scans *.ass). Not legal advice. Read-only.

Measure a video/audio file: duration_s, width/height, fps, codecs, file size, integrated loudness (LUFS), true peak (dBTP), loudness range, and the constant gain needed to reach the -14 LUFS short-drama target. Read-only. Use it to check an episode before or after mastering.

Master ONE short-drama episode into a platform-ready file: burns `srt` subtitles inside the safe area (18% bottom margin clear of the player bar, text box <= 70% of frame width for TikTok Drama Center), normalises loudness to -14 LUFS with a constant gain + peak limiter (no dynamic loudnorm), optionally lays a procedural music bed (`bed`: one of tension, warm, suspense, dread, hope) that ducks under dialogue, trims `head_trim` seconds (subtitles shift with it) and appends an end card. Set burn=false for a clean master without subtitles. Writes `output`; returns its measured spec. For a whole series use `package` instead.

Build a complete short-drama delivery package (platform submission pack) from `input_dir`, which holds episodes 01.mp4, 02.mp4, ... (un-subtitled cuts by default; set masters_ready=true if they are finished subtitled masters) plus optional 01.srt subtitles, 01.jpg covers, 01.timeline.json edit timelines, show.json and rights.json. Writes to `output_dir`: episodes/NN.mp4 + same-name NN.srt, clean/NN.mp4 clean masters (no subtitles), covers/NN.jpg (3:4, >=350x500), timeline renders + EDL CSV, docs/CHAIN-OF-TITLE.md (blank, to be signed by a person), docs/AI-DISCLOSURE.md, manifest.json with measured specs and a rights check. Returns status (red = do not ship), gaps, notes and file paths. target='tiktok-dc' also finishes every video for TikTok Drama Center: lifts dark frame edges (its 'black bars' rejection) and prepends a 1 s still of the episode cover so frame 0, its default episode cover, is the cover (SRTs shifted to match); edge_lift / cover_head turn either off. Slow: two encodes per episode (one more each with tiktok-dc).