Versions of the format

Changelog

On this page

Versions of the format itself, not of this repository. Version 1 is the first published one. The three before it were private drafts used inside Cassini, and two of them also called themselves /1 on the wire; spec/drafts/ says how a reader tells them apart.

Specification changes that are not format changes

The wire format did not move. These change what the documents say about it, and what the schemas accept.

2026-11-09 — when the recording started

Version 1 had no member for when a meeting happened. createdAtUtc is when the file was stamped, which for a server recording is after transcription, and recordedAtLocal is a label in a zone the file does not name. A server that does not know the recording site's zone wrote its own clock's digits there.

  • meeting.recordedAtUtc, optional: the instant of playback position 0 of the audio, RFC 3339 UTC. A word's wall-clock time is recordedAtUtc plus its startMs. It may be added to an existing file from trustworthy records; the meeting id does not change. The schema accepts it.
  • CASSINI_RECORDED_AT_UTC, an optional mirror tag carrying the same string.
  • DATE is defined: when the recording started, RFC 3339, as the Vorbis comment field means it. A producer that does not know the start writes createdAtUtc, as every producer has until now.
  • Producers SHOULD NOT write recordedAtLocal unless they know the recording site's zone. Consumers still MUST NOT convert it.
  • When the meeting happened: a consumer uses recordedAtUtc, else a DATE with an offset that is not createdAtUtc, else recordedAtLocal as an unconverted label, and SHOULD NOT present createdAtUtc as the meeting time.
  • tools/cassini-pack.py --recorded-at-utc, a meeting-time line in tools/cassini-read-pure.py, meetingStart() in the browser reader, and conformance vectors 027 and 028.

The container, chunk transport, audio digest and /1 identifier are unchanged. A reader that predates this ignores the new member and tag, as it must.

2026-09-07 — simplification

Compared with gocassini main at e4856e9fac32c447d6c5e023681433319d7ea3c8. The audit records the evidence and the implemented Cassini follow-up.

  • Word-timed entries no longer require or define an origin role (raw-asr, human-corrected, translation, scripted) or derivation link. Readers select from transcripts[] by default and array order, ignoring old origin labels. The Python producer omits them.
  • Remove the withdrawn LLM cleanup slot and its provenance, unwritten chapters, meeting.summary, meeting.language, the retired meeting.roomName mirror, and the speculative payloads / CASSINI_PL_ reservation from the current contract. Unknown members in old files remain ignorable and must survive edits.
  • Keep the used display slot and its source link. Correct the claim that its body is word JSON: Cassini writes transcript.display.v1 with blocks.
  • Update the browser reader, demo metadata, producer guidance and conformance vectors. Existing vectors retain old labels to test compatibility.

The container, chunk transport, audio digest and /1 identifier are unchanged. Cassini’s packer, inspector, viewer and static exporter have a matching consumer update. Earlier Cassini builds require that update to read files without the old role.

2026-09-01

  • cassini.words.v1, the transcript body, is written down for the first time: spec/cassini-words-v1.md and spec/cassini-words-v1.schema.json. One schema covers both a v1 file's inline manifest.transcript and a v2/v3 chunk-set body, because they are the same document type. Verified against a real file of each.
  • Correction: items are in speaker-turn order, not sorted by startMs. Across a speaker change startMs goes backwards, because overlapping speech keeps each speaker's words together. Sorting by time destroys every turn. The first draft of the body spec said the opposite and was wrong.
  • The v1 manifest schema declares displayTranscript. Without it the schema rejected all nine shipped v1 files, which is a schema that cannot validate the files it describes.
  • spec/cassini-opus-audio-integrity-v1.md defines finalGranule and what makes one invalid. Both were used in the normative formula and defined nowhere. The same file now cites RFC 6716 sections 3.1 and 3.2.5 for packet duration and for the frame-count rules that make a packet malformed, states the length prefix exactly (marker byte, then the length, then the packet, with the length counting the packet only), and says milliseconds truncate toward zero. Its one contradiction is gone: the final granule is not excluded from the digest, it bounds the playable sample count that is hashed.
  • A website: site/. It renders these documents rather than copying them, ships a real v3 file everything on it is measured from, and publishes the whole specification flattened as /llms-full.txt.
  • SPEC.md rewritten and reordered. It is now sorted into the machinery every version shares (identify a file, read the comment vector, reassemble a chunk set, decide whether to trust it), then the current manifest, then the earlier versions, and no section states a v1-only rule without v1 in a heading above it. The substantive changes:
    • A chunk set is defined once, over a tag prefix, so the manifest and every transcript body are one mechanism instead of a v1 rule plus a v2 special case. The index is zero-padded to a minimum of three digits, which is what the producer has always written.
    • Implementing this: the specifications this format inherits, with section numbers, and the Ogg page CRC parameters that RFC 3533 leaves out.
    • Reading the comment vector: which packet the tags are in, that it spans pages, that the separator is the first =, that names are case-insensitive, and what to do with a repeated name.
    • One integrity rule instead of three, with six named trust states — plain-audio, unknown-cassini-format, invalid-cassini-metadata, unverified, stale-audio, ok. None of them discards the transcript; the hard check moves to the producer, which is where it already was.
    • Embedded manifest points at the schema for the file's own version instead of always at v1, states the value of kind, and requires transcripts rather than a singular transcript no v3 file carries.
    • The transcript index, the per-transcript chunk sets and the reserved ids move out of the section marked legacy and into the current version, where a v3 implementer can find them.
    • base64url is stated as unpadded, and a consumer is required to bound its inflate against the declared RAW_BYTES before performing it.
    • Transcript ids may not contain _, because the tag prefix collapses it and one transcript's body is then lost with no error. The reserved list is ten, matching the code.
    • exact-pcm is no longer an unconditional MUST to run an algorithm nobody published. A consumer that cannot reproduce it reports unverified.
    • Extensibility, private use under X_ and x, and payloads/CASSINI_PL_ reserved for later.
    • The rollout plan moved to design/rollout-plan-v1.md; it was a CLI delivery plan inside a format specification, and it named commands that do not exist.
  • spec/conformance/: 22 vectors, a language-agnostic index and a harness, so "does my reader conform" has an answer. Running it found three disagreements between the reference implementation and these documents, two of them in files this repository's own readers decode fine.

Version 1 — org.cassini.portable-meeting/1 (published 2026-09-02)

The published format. Everything below this heading that was written as "v3" describes it: the last private draft became version 1 unchanged, plus the scripted role.

Private draft 3 — org.cassini.portable-meeting/1 (2026-08-29)

Current producer format. The number is contested: the format freeze written the day before proposed publishing the renamed format as <name>/2, which /3 now cuts across. See freeze item 6.

  • Recording identity moves from decoded PCM to the compressed Opus audio: integrity.matchPolicy = exact-opus-audio-v1, digest in integrity.opusAudioSha256, byte stream defined in spec/cassini-opus-audio-integrity-v1.md.
  • A consumer no longer has to decode the audio to verify identity.
  • The digest excludes OpusTags and Ogg framing, so it survives a metadata-only remux and is not self-referential.
  • Manifest and OpusTag layout are unchanged from v2.
  • containerSha256 dropped. matchPolicy survives and the freeze proposal wants it gone.

Private draft 2 — org.cassini.portable-meeting/2 (2026-05-12)

Legacy writer format; readers still support it.

  • One file can carry several transcripts of the same audio: transcripts and readableTranscripts arrays replace the single transcript / readableTranscript objects.
  • Each transcript body moves into its own OpusTag chunk set with its own gzip envelope and SHA-256, addressed from the manifest by payloadRef. A consumer decompresses only the body it displays.
  • provenance.speechToText and provenance.readableCleanup become maps keyed by transcript id.
  • CASSINI_TRANSCRIPT_IDS and CASSINI_TRANSCRIPT_DEFAULT make the set discoverable from ffprobe alone.

Private draft 1 — org.cassini.portable-meeting/1 (2026-03)

The original: Ogg Opus, one transcript, manifest gzipped and base64url-encoded across CASSINI_PAYLOAD_000..N, identity as a SHA-256 of decoded s16le PCM (integrity.matchPolicy = exact-pcm).

Rendered from CHANGELOG.md in the repository.