cjm-transcript-correction-core

A frontend-agnostic core for the transcript correction workflow — the first downstream graph-extension core; composes the graph-storage capability worker into a headless pipeline that applies unified, non-destructive corrections (text, punctuation, segmentation) as a supersede-able overlay on a committed decomposition spine, recomputes the review worklist from deterministic signals, and exposes a CLI as its first driver.

Modules

• cjm_transcript_correction_core
• cjm_transcript_correction_core.cli — The CLI driver — the correction core's first (and currently only) frontend. run corrects the committed spine in the decomp graph DB, pointing the graph worker at that shared DB via load-time config, with optional session resume/reopen; review runs the interactive text-correction loop (the cross-transcriber diff is intra-graph since stage 5).
• cjm_transcript_correction_core.graph — The correction overlay's graph I/O: targeted (scale-shaped) reads of a committed spine via the graph-storage query action, construction of Correction / CorrectionSession nodes + CORRECTS / SUPERSEDES / DERIVED_FROM / REVIEWED edges, the in-core effective-spine projection (layer-0 + applied corrections), and commit through the job queue. Hand-rolled (revolution-1) = direct CR-18 spec material; append-only on layer-0 (never update/delete a Segment).
• cjm_transcript_correction_core.journal — Live append-through for the correction verbs — the workflow journal's domain half.
• cjm_transcript_correction_core.launch — The shared launch surface every correction shell drives through: the
• cjm_transcript_correction_core.models — Overlay data shapes for the transcript-correction workflow: the Correction / CorrectionSession graph nodes + their relation registry, the read view of a committed spine segment, the worklist item, run configuration, and the correction run manifest (proto-bundle that chains decomp -> correction).
• cjm_transcript_correction_core.pipeline — The headless correction workflow: load a decomp run manifest, resolve the shared graph DB, start/resume/reopen a CorrectionSession, recompute the worklist from deterministic signals + persisted review state, run the D14 empty-segment prune (first operation), and record a chainable correction run manifest — with a cheapest-form HITL approval seam.
• cjm_transcript_correction_core.signals — Pure deterministic Tier-1 signal functions (no capability calls): empty-segment detection, bidirectional boundary punctuation/capitalization heuristics, forced-alignment coverage flags, positional cross-transcriber diff, phonetic + edit-distance variant clustering, and the event-proposal overlay (leg 4: the finetuned detector's spans anchored onto the spine). The worklist is recomputed from these each session; revolution-1 builds ZERO new capabilities.
• cjm_transcript_correction_core.spine
• cjm_transcript_correction_core.state — Sidecar view-state + spine-picker helpers — the correction TUI's pure

API

cjm_transcript_correction_core.cli

• bench_command function — Execute the bench subcommand: the reserved-tail verdict join.
• build_parser function — Build the CLI parser (subcommands: run, review).
• export_command function — Execute export-wordless-propset: write one spine's effective wordless
• extract_command function — Execute the extract subcommand: fold the gated overlay into a manifested dataset.
• gate_command function — Execute the gate subcommand: show or assert per-spine extraction gates.
• load_capabilities function — Discover manifests + load each capability, passing per-capability config (CR-2 caller-wins).
• main function — CLI entry point (console script: cjm-transcript-correction-core).
• overlay_event_rows function — Overlay span records -> dataset event rows for ONE spine (pure).
• review_command function — Execute the review subcommand: interactive text corrections over the flagged worklist.
• run_command function — Execute the run subcommand: correct a decomp manifest's committed spine.
• run_extract function — The extract fold on an ALREADY-OPEN graph seat (flywheel build leg 2,
• scan_command function — Execute scan-mishomed: flag authoritative FA words stranded outside
• stats_command function — Execute the stats subcommand: flywheel accounting over the shared graph.
• transfer_command function — Execute transfer-wordless: replay wordless event inserts across sibling spines.
• wordless_donors function — The EFFECTIVE wordless layer of a spine: labeled, effectively wordless

cjm_transcript_correction_core.graph

• active_corrections function — Filter to the effective correction set (the layer's resolve_active over a read superseded set).
• active_speaker_assignments function — Project the ACTIVE speaker assignment per segment (latest-wins).
• active_speech_overlays function — The surviving speech-overlay corrections (supersession applied; pure).
• aggregate_session_purposes function — Fold sessions into a per-source purpose mix (pure; d915d545 picker rung).
• apply_chunk_inserts function — Synthesize inserted chunks into the effective spine (DEC 3d3fa2a8).
• apply_time_nudges function — Apply timing corrections onto segment times (latest-wins per edge).
• bench_event_proposals function — The reserved-tail bench join (leg 4, DECs 8e05b87b + 8cf12c22) — pure.
• build_boundary_shift_correction function — Build a grouping Correction that moves text across one segment boundary.
• build_chunk_insert_correction function — Build an insertion Correction that adds a chunk the skeleton never cut (DEC 3d3fa2a8).
• build_chunk_split_corrections function — Compose a chunk SPLIT from the EXISTING verbs (work item 99c1d2ba) — no
• build_correction_node function — Construct a Correction overlay node (pure; commit happens separately).
• build_extraction_gate_assertion function — Build one extraction-gate ASSERTION (DEC 8e05b87b — flywheel build leg 1).
• build_mark_correction function — Build a NON-MUTATING mark Correction (DEC 2a231843: routed attention).
• build_prune_amendment function — Build a grouping Correction that supersedes a prune with a REDUCED set (unprune).
• build_prune_correction function — Build one batch grouping Correction that prunes empty segments (D14).
• build_reject_review function — Build a review Correction that REJECTS a prior correction (reject-as-supersede).
• build_speaker_assign_correction function — Build a speaker Correction — the assignment op envelope (DEC d6df3a8e).
• build_speech_overlay_correction function — Build a NON-MUTATING speech-overlay Correction (check fc42614d, DEC 4e05a066).
• build_text_correction function — Build a text_content Correction + its CORRECTS (+ optional SUPERSEDES) edges.
• build_time_nudge_correction function — Build a timing Correction that nudges segment boundary TIMES (node + CORRECTS edges).
• commit_boundary_shift_correction function — Commit a boundary-shift correction (node + CORRECTS x2 [+ SUPERSEDES]) + REVIEWED markers on both segments.
• commit_chunk_insert_correction function — Commit a chunk insertion (node + CORRECTS per flank).
• commit_chunk_insert_removal function — Remove an inserted chunk WITHOUT touching layer-0 (reject-as-supersede).
• commit_chunk_split_correction function — Commit a chunk split: three composed nodes in ONE atomic batch + ONE journal op.
• commit_chunk_split_removal function — UNSPLIT: remove a split's right half AND its whole group (one review
• commit_extraction_gate function — Commit one extraction-gate assertion (node + GATES edge) — the journaled
• commit_mark_correction function — Commit a mark (node + CORRECTS per anchored segment [+ SUPERSEDES]).
• commit_mark_dismissal function — Dismiss an open mark WITHOUT a correction (reject-as-supersede).
• commit_nodes_edges function — Commit overlay nodes/edges through the layer's idempotent extend_graph.
• commit_prune_amendment function — Commit an unprune amendment (node + DERIVED_FROM edges + SUPERSEDES).
• commit_speaker_assign_correction function — Commit a speaker assignment (node + CORRECTS per segment + ASSIGNS).
• commit_speaker_entity function — Mint a source-spanning Entity into the shared registry (DEC 4ec6a49c).
• commit_speech_overlay_correction function — Commit a speech overlay (node + CORRECTS [+ SUPERSEDES]).
• commit_speech_overlay_removal function — Remove a speech overlay WITHOUT a replacement (reject-as-supersede).
• commit_text_correction function — Commit a text_content correction (node + CORRECTS [+ SUPERSEDES]) + a REVIEWED marker.
• commit_time_nudge_correction function — Commit a time-nudge correction (node + CORRECTS per touched segment).
• correction_stats function — Fold one Source's ACTIVE overlay into flywheel-accounting counts (pure).
• corrections_to_edits function — Map this core's Correction payloads onto the layer's spine-edit vocabulary.
• count_source_segments function — Count a Source's segments server-side under its chosen rendition + skeleton (typed count mode).
• extract_spine_dataset function — Fold ONE spine's overlay into its v1 insert-span dataset slice (pure).
• fa_words_for_transcript function — One transcript's FA words in source coordinates (the scan-mishomed join,
• find_active_text_correction function — Single-segment convenience over the batch read (cross-session; latest wins).
• find_active_text_corrections_batch function — Active text corrections for MANY segments in TWO round-trips (C17).
• find_chunk_split_group function — Resolve a split right-half's GROUP — the ac84360a group marker cashed in.
• find_corrections_for_session function — List corrections recorded in a session (typed property filter).
• find_prior_corrections_by_hash function — Cross-transcript correction-cache lookup (targeted; the graph IS the lexicon).
• get_session function — Fetch a CorrectionSession node by id (resume/reopen) — typed get, dict shape preserved.
• labeled_insert_spans function — Fold the ACTIVE chunk inserts into span records with op-chain provenance (pure).
• latest_extraction_gates function — Fold gate assertions into the live per-spine gate state (pure; latest-wins).
• list_source_spines function — The SPINES coexisting under a Source's chosen rendition (DEC f1024568).
• list_speaker_entities function — Read the source-spanning entity registry (the picker's registry tier).
• load_empty_segments function — Load ONLY a Source's empty-text segments under its chosen rendition (D14 prune).
• load_extraction_gates function — Read one Source's live per-spine extraction gates (typed property filter).
• load_review_markers function — Load a session's review markers (typed edge projection over REVIEWED edges).
• load_source_corrections function — Load every Correction targeting a Source (across sessions) + the superseded-id set.
• load_source_segments function — Load a Source's fine Segment spine under its chosen rendition + skeleton (typed query surface).
• load_variant_texts function — Resolve per-transcriber chunk texts from the segments' CharSlice refs.
• mark_anchor_segments function — Validate a mark anchor and list the Segment ids it touches.
• negative_regions function — Complement the accounted spans below the watermark (pure; DEC 8e05b87b).
• open_marks function — Filter to the OPEN marks — the pass-2 worklist ('query open marks, walk them').
• project_effective_spine function — Project the effective spine = layer-0 + applied corrections.
• reanchor_span function — Re-locate a span anchor in text that may have been edited since mark time.
• record_review_markers function — Persist per-(session, segment) review markers as REVIEWED edges.
• rename_speaker_entity function — Rename an Entity on its STABLE id — identification IS a rename (DEC 484e2d74).
• reorder_gap_inserts function — Restore TIME order among stacked inserts after the nudge stage
• resolve_source_renditions function — Pick the AudioRendition set whose fine Segment spine correction operates on.
• session_purposes_by_source function — Read every CorrectionSession once and fold the per-source purpose mix.
• set_session_purpose function — Update a session's purpose + updated_at (the test-session hygiene tag, DEC c86714a4).
• set_session_status function — Update a session's status + updated_at.
• skeleton_hash_for function — Resolve a skeleton selector to the chosen spine's HASH (pure) — the gate's
• source_audio_segment_ids function — The Source's coarse spine (one small typed read; ordered by index).
• speech_overlay_spans function — Fold the ACTIVE speech overlays into extraction-facing span records (pure).
• spine_where_for function — Resolve a skeleton selector against the observed spine set (pure).
• start_session function — Create + commit a new CorrectionSession node.
• submit_and_wait function — Submit one capability job, wait for it, return its result (raise on failure).

cjm_transcript_correction_core.journal

• correction_replay_handlers function — The correction core's registered replay vocabulary (replay stays DOMAIN-OWNED).
• journal_correction_op function — Append one correction op — envelope + semantic args + the EXACT wires committed.
• segment_anchor function — The run-independent anchor stamped on every correction op (DEC ccbab9f5 point 5).

cjm_transcript_correction_core.launch

• build_parser function — The TUI driver's argument surface (mirrors correction-core's run/review args).
• resolve_settings function — Resolve the launch settings every shell shares (5daadfc4 workspace

cjm_transcript_correction_core.models

• Correction class — A single non-destructive correction over the committed spine (overlay node).
• CorrectionConfig class — Configuration for one correction run.
• CorrectionManifest class — Durable record of one correction run (proto-bundle; chainable, CR-20).
• CorrectionRelations class — Registry of edge types the correction overlay adds to the spine graph.
• CorrectionSession class — A resumable, reopen-able correction review over one or more sources.
• DatasetManifest class — Durable record of one dataset extraction (chainable; DEC 16159e09).
• Entity class — A source-spanning identity in the shared entity substrate (DEC 4ec6a49c).
• ExtractionGate class — One per-spine extraction-gate ASSERTION (DEC 8e05b87b — flywheel build leg 1).
• SpineSegment class — A committed layer-0 Segment loaded from the graph (read view).
• WorklistItem class — One spine segment surfaced for review, with its deterministic Tier-1 flags.
• new_dataset_id function — Generate a unique, sortable dataset id (the new_run_id pattern, dataset kind).
• new_run_id function — Generate a unique, sortable correction run id.

cjm_transcript_correction_core.pipeline

• collect_capability_info function — Record capability identity + data-DB pointers for the run manifest (provenance).
• compute_worklist function — Recompute the worklist from layer-0 + signals + review state (only decisions persist).
• confirm_seam function — HITL approval seam in its cheapest viable form (log + optional CLI prompt).
• load_decomp_manifest function — Load + lightly validate a decomp-core run manifest (untyped JSON; CR-20 interchange).
• prune_empty_segments function — First operation: prune empty (silence) segments as one grouping correction (D14).
• resolve_graph_db_path function — Resolve the graph DB path: explicit override > the decomp manifest's recorded db_path.
• review_worklist function — Interactive review loop -> text_content corrections (cheapest HITL seam).
• run_correction function — Correct every source in a decomp run manifest (prune + worklist surfacing).
• run_review function — Interactive review pass over a decomp manifest's flagged worklist (text corrections).

cjm_transcript_correction_core.signals

• boundary_punct_caps_flags function — Bidirectional boundary punctuation/capitalization heuristics (in-segment only).
• cluster_variants function — Cluster word variants by phonetic key + edit distance (fix-one-fix-all).
• compute_signal_flags function — Combine all deterministic Tier-1 signals into per-segment flags.
• detect_empty_segments function — Find empty-text segments (silence VAD chunks with no aligned words; decomp D14).
• event_span_proposals function — Anchor pending event proposals onto the spine — the propose lane's paint.
• fa_coverage_flags function — Flag segments whose forced-alignment coverage looks suspect (Tier-1).
• levenshtein function — Levenshtein edit distance (pure, in-core; variant-clustering primitive).
• load_event_proposal_set function — Find the latest proposal set for a source (the turns-artifact discovery
• phonetic_key function — Compute a coarse phonetic key for a word (groups like-sounding variants).
• speaker_turn_proposals function — Dominant diarization cluster per segment — the assign lane's proposal paint.
• variant_divergence function — Within-segment cross-transcriber divergence (stage 5: intra-graph).

cjm_transcript_correction_core.spine

• ChunkRef class — Where one Segment's VAD-chunk audio lives: the model-input WAV + the chunk-local span.
• SeamRef class — A source-coordinate audio span across one fine-spine boundary (the g/G
• SpineView class — One Source's effective correction spine, cursor-windowed for the TUI.
• list_sources function — Enumerate the graph's Source nodes (the discovery corpus, 2ce81638).
• load_source_slice function — Decode a source-coordinate slice of the ORIGINAL media to playable samples.
• match_sources function — The --source selector (pure; shared by direct open and the picker's seed).
• neighbor_word_bound function — The adjacent word's FA boundary facing an overlay span (pure).
• open_stack function — Bootstrap the graph capability stack, resolving the db path (2ce81638).
• parse_entity_input function — Parse the new-speaker editor line (pure). A leading ? marks the entity
• parse_mark_input function — Parse the M-editor mark grammar (pure; the DEC 2a231843 TUI gesture).
• plan_boundary_shift function — Plan a ONE-WORD boundary shift (the [ / ] gesture unit).
• plan_chunk_insert function — Plan a chunk insertion into the seam after the cursor (the i gesture unit; pure).
• plan_chunk_split function — Plan a chunk split at a caret position (the S gesture unit; pure).
• plan_gate function — Plan an extraction-gate assertion (the F gesture unit; pure — DEC 8e05b87b).
• plan_time_nudge function — Plan a boundary-time nudge (the ,/. and </> gesture unit; pure).
• resolve_mark_class_token function — Resolve a leading digit token to its menu class (the M picker; pure).
• segment_word_tokens function — Tokenize a segment's text into words WITH character offsets (pure).
• snap_word_span function — Derive a word range's TIME span from FA word timestamps (pure; fc42614d).
• source_status function — Correction-status-at-a-glance for one Source (the picker's detail row).

cjm_transcript_correction_core.state

• load_tui_state function — Read the per-graph TUI sidecar state (last-focused positions).
• save_tui_state function — Merge one source's view state into the sidecar state file.
• selector_for_spine function — The selector value a picker choice persists (pure): the full skeleton
• spine_label function — One picker row's config summary for a skeleton spine (pure).

Dependencies

Depends on: cjm-capability-primitives, cjm-context-graph-layer, cjm-context-graph-primitives, cjm-substrate, cjm-transcript-graph-schema, numpy Used by: cjm-transcript-correction-qt, cjm-transcript-correction-tui, cjm-transcription-core, cjm-workflow-hub-qt, cjm-workflow-hub-tui