Paste or edit the transcript, generate an audit-ready SOAP with Linked Evidence, then sign off with edits.
Hierarchy
Classic CAT and New (modern) are vertical flows that fan out into S/O/A/P — not a metrics table. Flat stays a thin one-shot. Level-1 summarize is chunk compression, not L1-reg.
New (modern)
Chunk → Level-1 summarize → parallel sections → SOAP
live-hierarchical-llm
Transcript
Utterance hierarchy + CAP-lite
Chunk transcript
Utterance windows (chunkTranscript)
Level-1 summarize
LLM compress each chunk (not L1-reg)
Enrich hierarchy
CAP-lite bundles + L1 summaries
Parallel section generators
Together chat · stub if no key
Subjective (S)
Patient-reported
Objective (O)
Observed findings
Assessment (A)
Formulation
Plan (P)
Interventions
SOAP note
Assembled clinical note
Flat (baseline)
Transcript → one-shot SOAP (no hierarchy for generation)
direct-prompt
Transcript
Full session as single chunk
One-shot SOAP prompt
direct-prompt · no Level-1 summarize
Parse S / O / A / P
Shared hierarchy only for Linked Evidence
SOAP note
Assembled clinical note
Classic (CAT)
Context Aggregation of Transcripts → SOAP
classic-cat
Transcript
Raw encounter lines
Normalize transcript
Line clean · speaker turns
Extract key phrases
CAT-keyphrase inventory
Clinical embeddings
BiomedBERT-class (demo: local heuristic)
Hierarchical clustering
Phrase clusters + subclustering
Section ensemble
Classify into SOAP sections
Subjective (S)
Patient-reported
Objective (O)
Observed findings
Assessment (A)
Formulation
Plan (P)
Interventions
SOAP note
Assembled clinical note