Reference

fxdub-dialogue

fxdub-dialogue <scene.json> <words.json> [--only-speaker NAME] [--json OUT] [--debug]

Argument	Meaning
scene	Authored scene JSON. See Scene scripts.
words	Word-level diarized transcript JSON.
--only-speaker NAME	Narrow the contract to one character — the mode for per-character stems. Case-sensitive.
--json OUT	Write the full receipt.
--debug	Re-raise on error instead of printing the structured shape.

Checks

line_present:N:SPEAKER · no_invented_speech · no_overlap · no_internal_straggle · turn_gaps · one_voice_per_character · characters_are_distinct · fits_clip

Error codes

code	Cause
scene_not_found	no file at the scene path
scene_malformed	scene is not valid JSON
empty_contract	the scene has no lines
unknown_speaker	--only-speaker names a character absent from the scene
words_not_found	no file at the transcript path
words_malformed	transcript is not valid JSON
receipt_unwritable	--json path is not writable

fxdub-receipt

fxdub-receipt <run_dir> [--bed-gain-db N] [--json OUT] [--debug]

Argument	Meaning
run_dir	Directory holding the downloaded artifacts.
--bed-gain-db N	Mix gain applied to the bed, in dB. Required for a correct ducking figure. Default -9.0.
--json OUT	Write the full receipt.

Expected files

Matched by filename_prefix stem, whatever counter suffix the platform appends:

mix*.flac   stem_bed*.flac   stem_vo*.flac
mix_lufs*.txt   vo_lufs*.txt   bed_lufs*.txt
caption*.txt   dubbed*.mp4

The bed_lufs manifest is required — R128 gates a quiet bed out of the mix master, so ducking depth is unmeasurable without a meter on the bed stem.

Contract values

Constant	Value
Runtime sample rate	48 000 Hz
Target loudness	−18.0 LUFS, ±2.0
Dialogue over bed	8–20 LU

Exit codes

Code	Meaning
0	every check passed
1	contract failure — a finding, with a receipt to read
2	runtime error — the tool could not run

3 (partial success) is unused: a receipt is all-or-nothing, and there is no partial verification.

Python API

from fxdub import audition_receipt, dialogue_receipt, media_probe, vo_graphs

dialogue_receipt

Function	Returns
check_dialogue(scene, words, only_speaker=None)	{"checks": [...], "measured": {...}}
align(scene_lines, words)	(matched, unconsumed) — per-line spans, and every token no line claimed
normalize_words(raw)	coerce raw node output into the internal word shape
load_words(path)	the same, from disk
normalize(text)	comparable word tokens
render(result)	the markdown table

check_dialogue accepts raw transcript JSON directly — no conversion step needed.

audition_receipt

Function	Returns
check_run(run_dir, bed_gain_db=-9.0)	{"checks": [...], "measured": {...}}
render(result)	the markdown receipt

Never raises on a bad artifact — a malformed file is a failed check, not a crash.

media_probe

Function	Returns
parse_streaminfo(data)	FLAC sample rate, channels, bit depth, total samples
probe_mp4(data)	track handlers, frame count, duration, header validity
parse_lufs(text)	the float out of a loudness manifest

Standard library only — no ffprobe subprocess.

vo_graphs

See Graph builders. Every function returns an API-format dict; none of them submit.

Supported platforms

Python 3.10+ on any OS. CI runs 3.10 and 3.12 on Linux. There is no compiled extension, no subprocess, and no network call, so platform differences are limited to path handling.

Security

No network egress, no telemetry, no credential access, zero runtime dependencies — and CI fails the build if that dependency list ever becomes non-empty. Full threat model in SECURITY.md.