cjm-transcript-decomp-core

A frontend-agnostic core for the transcript decomposition workflow — composes isolated capability workers (forced alignment, VAD, graph storage) into a headless pipeline that decomposes transcription run manifests into a VAD-aligned context-graph spine with traceable provenance, and a CLI as its first driver.

Modules

• cjm_transcript_decomp_core
• cjm_transcript_decomp_core.alignment — Pure forced-alignment logic (no capability calls): map FA words back to character spans in the original text, assign words to VAD chunks by timestamp, and build one text segment per VAD chunk. Extracted from the page-centric ForcedAlignmentService (Tier-1 logic).
• cjm_transcript_decomp_core.cli — The CLI driver — the decomposition core's first (and currently only) frontend.
• cjm_transcript_decomp_core.discovery — Capability-role discovery by manifest surface match — the journaling-by-
• cjm_transcript_decomp_core.graph — Graph-spine EXTENSION + skeptical-lens verification (stage 5, CR-18 revolution 2). Decomp no longer creates a Document: it RECOMPUTES the transcription-emitted root's deterministic node ids from the consumed manifest (no search), verifies the root exists, and attaches the fine Segment spine under the existing AudioSegment nodes — PART_OF to the owning rendition, STARTS_WITH per rendition (the coarse-seam jump anchor), source-wide NEXT. Each Segment carries the audio TimeSlice ref plus per-transcriber CharSlice refs into the Transcript nodes (the D4/P10 framing, finally expressible). Commit goes through the layer's idempotent extend_graph.
• cjm_transcript_decomp_core.launch — The shared launch surface every decomp shell drives through: the argument
• cjm_transcript_decomp_core.models — Lean data shapes for the transcript-decomposition pipeline: in-core mirrors of the forced-alignment / VAD / text DTOs (no FastHTML deps), run configuration, the committed graph-segment carrier, and the decomposition run manifest (proto-bundle).
• cjm_transcript_decomp_core.pipeline — The headless decomposition pipeline (stage 5: decomp is an EXTENDER). Load a transcription run manifest, verify the transcription-emitted graph root exists (the graph begins at transcription), then per source per pipeline-segment run VAD + per-transcriber forced alignment, build one aligned segment per VAD chunk with per-transcriber text variants, and attach the fine spine under the existing AudioSegment nodes via the layer's idempotent extend_graph — with HITL approval seams between alignment, commit, and the next source.
• cjm_transcript_decomp_core.runs — Run-manifest indexes for the decomp-batch TUI (work item 0ff6bf0f): the
• cjm_transcript_decomp_core.segments — Fine-segment inspection for the decomp TUI (work item 166dd2b8, half a):
• cjm_transcript_decomp_core.state — Sidecar TUI state: last-used batch settings persisted across sessions (the

API

cjm_transcript_decomp_core.alignment

• assign_words_to_chunks function — Assign each FA word to a VAD chunk by timestamp overlap.
• build_segments_from_alignment function — Build a TextSegment per VAD chunk by grouping words by chunk assignment.
• carve_chunks_at_event_spans function — The event-carve stage (EVENT_SPLIT_POLICY, respine trial DEC 6cc10fb7):
• map_fa_words_to_text function — Map forced-alignment words back to character spans in the original text.
• rescue_gap_words function — The word-rescue stage (WORD_RESCUE_POLICY, 96edc646 verdict bc7ece7b):
• sentence_end_word_indices function — Map capability-delivered sentence boundaries onto FA words (B.5: the
• split_chunks_at_sentence_gaps function — The sentence-split stage (SENTENCE_SPLIT_POLICY, DEC f1024568): refine the
• tier1_alignment_checks function — Tier-1 deterministic pre-filters for the alignment-review seam (no AI).

cjm_transcript_decomp_core.cli

• build_parser function — Build the CLI parser (subcommands: run).
• load_capabilities function — Discover manifests + load each requested capability (default instance).
• main function — CLI entry point (console script: cjm-transcript-decomp-core).
• run_command function — Execute the run subcommand: extend transcription-run manifest(s) with the fine spine.

cjm_transcript_decomp_core.discovery

• discover_capability function — Pick a DEFAULT capability for a role by surface match.
• manifests_with_method function — Enumerate installed capabilities whose structural surface lists method.

cjm_transcript_decomp_core.graph

• SourceVerification class — Skeptical-lens verification of one Source's fine-spine extension under a
• build_extension_payload function — Build the fine-spine EXTENSION payload (pure; no capability calls).
• resolve_root_ids function — Recompute the transcription-emitted root node ids from manifest data.
• verify_source function — Verify a Source's committed extension via server-side AGGREGATES (D13/D19).

cjm_transcript_decomp_core.launch

• batch_argv function — Render one hand-off group as headless decomp-core argv.
• build_parser function — The TUI driver's argument surface (batch-setup options + core passthrough).
• event_split_batch_error function — One propset carves ONE source: the core applies --event-propset to every
• hand_off function — The shared driver tail: persist the confirmed choices, adopt the in-app
• resolve_settings function — Resolve the batch-setup settings every shell shares (flags > persisted
• resolve_split_flags function — Resolve the post-parse split-flag contract (pure; main() calls it

cjm_transcript_decomp_core.models

• DecompConfig class — Configuration for one transcript-decomposition run.
• DecompManifest class — Durable record of one decomposition run (proto-bundle; see CR-20).
• DecompSegment class — One fine spine segment (stage 5: shared audio-side skeleton + per-transcriber variants).
• DecompSourceRecord class — Record of one Source whose fine spine this run committed (stage 5:
• FAWord class — One word-level forced-alignment result (segment-local times).
• SegmentVariant class — One transcriber's text + char range for one fine segment (stage 5).
• TextSegment class — A text segment produced by alignment, before graph commit.
• VADChunk class — A voice-activity time range within one pipeline segment (segment-local).
• new_run_id function — Generate a unique, sortable decomposition run id.

cjm_transcript_decomp_core.pipeline

• build_alignment_composition function — Build the whole-source M×(VAD ∥ T×FA ∥ SEG) composition (D8 fan-in, stage-5 variants).
• collect_capability_info function — Record capability identity + data-DB pointers for the run manifest (provenance).
• compute_skeleton_hash function — Skeleton identity, pure (DEC f1024568 + 9241564f + 6cc10fb7).
• confirm_seam function — HITL approval seam in its cheapest viable form (log + optional CLI prompt).
• decomp_replay_handlers function — The decomp core's replay vocabulary (DEC 426658f1, replay stays DOMAIN-OWNED).
• decompose_source function — Decompose one source into aligned fine segments with per-transcriber variants.
• event_spans_from_propset function — Load a proposal set BY POINTER and select its carve spans (respine trial
• fa_words_from_result function — Normalize a typed forced-alignment result into FA words (pure; stage 3).
• load_source_manifest function — Load + lightly validate a transcription-core run manifest.
• resolve_event_propsets function — Join each proposal set to ITS source by the set manifest's own source
• run_decomp function — Extend every source in a transcription run manifest with its fine spine.
• sentence_spans_from_result function — Normalize a typed sentence-segmentation result into char-span tuples (pure; B.5).
• submit_and_wait function — Submit one capability job, wait for it, and return its result (raise on failure).
• vad_chunks_from_result function — Normalize a typed VAD result into segment-local VAD chunks.

cjm_transcript_decomp_core.runs

• DecompIndex class — Decomp-core run manifests read back: coverage chips for the batch stage
• PropsetIndex class — Proposal-set manifests under the workspace proposals/ dir — the model's
• SourceRunIndex class — Transcription-core run manifests — the decomp workflow's SOURCES — plus
• TrainingRunIndex class — Training-run manifests under the workspace training-runs/ dir — the
• group_batches function — Fold an ordered batch selection into headless hand-off groups.

cjm_transcript_decomp_core.segments

• SegmentStack class — One lazily-opened READ-ONLY graph seat, keyed by db path.
• aseg_index_for function — Which coarse AudioSegment a source-coordinate time falls in (pure;
• build_display function — Interleave gap markers into the paintable entry list (pure).
• capability_config_schema function — A capability's config_schema off its installed manifest (json read, pure).
• find_gaps function — Uncovered timestamp spans between committed segments (pure).
• fmt_ts function — Source-coordinate timestamp for listing rows (pure).
• locate_span function — Resolve a source-coordinate span onto its owning coarse WAV (pure).
• ordered_rows function — Re-impose the manifest's spine order on fetched rows (pure).
• predicted_rows function — Synthesize segment-shaped rows for a probe's predicted skeleton (pure).
• probe_compare function — Compare a VAD probe against the committed skeleton (pure).
• realign_rows function — Re-run the decomp pipeline's own text fold over a PROBE skeleton (pure).
• split_predicted function — Run the decomp pipeline's own SENTENCE-SPLIT stage over a probe skeleton
• vad_summary function — The decomp run's recorded VAD config, one status line (pure).

cjm_transcript_decomp_core.state

• load_state function — Read this project's persisted TUI state.
• save_state function — Merge updates into the persisted state and write it back (best-effort:
• state_path function — Where this project's TUI state lives.

Dependencies

Depends on: cjm-capability-primitives, cjm-context-graph-layer, cjm-context-graph-primitives, cjm-substrate, cjm-transcript-graph-schema, pyyaml Used by: cjm-transcript-decomp-qt, cjm-transcript-decomp-tui