Output files

Every artefact a run produces is colocated with output.video_path. The project names them; the caller decides where they land, via {output} and --output-root.

The artefacts

File	Written when	Contents
<name>.cvbs	signal_type: composite	Composite samples
<name>.cvbsy	signal_type: yc	Luma samples
<name>.cvbsc	signal_type: yc	Chroma samples
<name>.meta	Always	SQLite metadata sidecar
<name>.dropouts.meta	Any section has dropouts	SQLite dropout sidecar
<name>_audio_<pair>.wav	Any section declares that pair	48 kHz/24-bit stereo audio
<name>.efm	laserdisc.efm_audio is present	Laserdisc EFM channel stream
<name>.efm.meta	With the above	SQLite frame index for the EFM stream
<name>.truth	output.truth: true	Picture truth: planar float32 Y/U/V at 4fsc
<name>.truth.meta	With the above	SQLite sidecar: layout and conventions of the truth file

<name> for the audio, EFM and truth files is output.video_path with its trailing .cvbs or .cvbsy removed. The .meta path is derived from the video path by replacing that suffix — it is never specified in the project.

Parent directories are created automatically, so an output path can point anywhere writable.

Sample files

The .cvbs (or .cvbsy/.cvbsc) file is a flat, headerless stream of samples in the project's sample encoding, little-endian, with no per-frame framing of any kind. Frame boundaries are implicit in the sample count:

Standard	Samples per frame at 4fsc
PAL / SECAM	709,379 (1135 per line, plus two extra samples on each of lines 313 and 625)
NTSC	477,750 (910 × 525, orthogonal)
PAL-M	477,225 (909 × 525)

At 16 bits per sample that is roughly 1.4 MB per PAL frame and 0.95 MB per NTSC frame — so about 2.1 GB per minute of PAL. Plan storage accordingly, and use --output-root to send bulk runs somewhere with room.

The metadata sidecar

<name>.meta is a SQLite database following the CVBS File Format Specification. Its cvbs_file table describes the file as a whole:

Column	Contents
preset	PAL, NTSC, PAL_M or SECAM
sample_encoding_preset	The encoding the samples are in
signal_state_preset	STANDARD_STABLE_LOCKED; STANDARD_STABLE_UNLOCKED for SECAM
sequence_continuous	Always TRUE — the synthesised signal is one unbroken sequence
signal_type	composite or yc
decoder	The producing tool
git_branch, git_commit	The build the file came from
number_of_sequential_frames	Frame count
black_level	The black level in use — relevant for the NTSC setup option
has_nonstandard_values	Whether the file deliberately departs from the standard
capture_notes	Free text

When any section declares audio, an audio_channel_pair table carries one row per emitted WAV track, with the description from the project.

Recording the git commit in the sidecar is deliberate: a generated file always says which build produced it, so a decoding result can be traced back to a specific state of videosynth.

The dropout sidecar

<name>.dropouts.meta (schema version 5) carries one dropout_run row per contiguous dropout, listing the frame, first sample, length and a severity of 25 (non-visible VBI) or 75 (intersecting the active picture). It is ground truth against which a dropout detector can be scored directly. Dropouts with flagged: false are applied to the signal but have no rows, so they are absent from the sidecar; the sidecar itself is still written whenever any section has dropouts. See Impairments.

Audio and EFM

Audio tracks are stereo, 24-bit signed little-endian PCM (RIFF/WAVE) at 48 000 Hz, one file per declared channel pair, each spanning the whole output. EFM output is the two-file pair defined by the EFM extension format — the channel stream and its own frame-index sidecar. See Audio.

Picture truth

With output.truth: true, <name>.truth holds the component picture that went into the modulator: for every frame, a Y, a U and a V plane, each one row per active line (in progressive order) by one sample per 4fsc sample of the active window, as little-endian float32 millivolts. U and V are the filtered colour-difference components exactly as modulated, with the PAL V-switch removed. The SQLite sidecar <name>.truth.meta gives the dimensions, the first sample and line of each field, the levels and the conventions. The truth is taken before every impairment, so a capture variant with channel effects, noise or dropouts shares its truth with the clean run.

Determinism

The same project produces byte-identical files on every run, on any machine, at any thread count, with or without the frame template cache. Where randomness is involved — noise and dropouts — a fixed seed makes it reproducible too.

This is a testable property, not just an intention: the repository ships scripts/output-hashes.sh, which records SHA-256 manifests of the sample, audio, picture-truth and .meta files (using SQLite's canonical .dump for the databases; the .efm stream is not included) and fails if a later build differs from a recorded baseline.

# On a known-good build
scripts/output-hashes.sh --record

# After a change: lists any artefact that differs
scripts/output-hashes.sh

Cancellation

A cancelled run aborts every artefact it was writing — video, chroma, metadata, WAV tracks, dropout sidecar, EFM stream and picture truth — so it leaves no partially written files behind. A cancelled run is reported as cancelled, not as a failure.