feat(llm): closed theme vocabulary for comment tagging — themes.yaml v1, llm.vocab loader, stack llm vocab, P35 design spec (refs #574)

docs/superpowers/specs/2026-09-22-llm-tagging-design.md lays out P35:
a versioned hand-curated vocabulary, embedding shortlist + per-theme
yes/no judgement with an evidence chunk on the largest live host,
resumable docket-scoped runs with state in bib extra_json, write-back
that replaces only llm: tags, and a golden-set gate before fan-out.

Slice 1: src/llm/vocab/themes.yaml — 53 themes (conversion factor,
practice expense, telehealth, care management, behavioral health, Part
B drugs, MSSP, QPP, …), each with a label, one-sentence definition,
letter-language synonyms and the FR section stems it was seeded from
(dev/scripts/llm_vocab_candidates.py mines heading paragraphs from the
CY2018+ rules' fr_anchors). llm.vocab.load() validates unique
kebab-case slugs, non-empty definitions, integer version ≥ 1 and that
retired slugs are not still active; Theme.card is the text embedded
per theme; split_tags() separates llm: tags from hand tags for the
write-back. stack llm vocab [--path] [--slug].

dev/scripts/llm_vocab_candidates.py

@@ -0,0 +1,81 @@
+"""Seed a theme-vocabulary revision from Federal Register section headings (#574).
+
+Walks ``fr_anchors`` for heading-like paragraphs ("B. Determination of
+Practice Expense…") in every PFS rule whose rule year is at or after
+``--since`` (default 2018 — the regulations.gov comment dockets start in
+2017), stems the titles, and prints the stems ordered by how many rule
+years carry them, with an example title each. The output is raw material
+for a human editing ``src/llm/vocab/themes.yaml`` — it is not the vocab.
+
+Usage:
+    uv run python dev/scripts/llm_vocab_candidates.py [--since 2018] [--top 200]
+"""
+
+from __future__ import annotations
+
+import argparse
+import re
+from collections import defaultdict
+
+_HEADING = re.compile(r"^(?:[IVX]{1,4}|[A-Z]|\d{1,2}|[a-z])\. +([A-Z][^.]{6,120})$")
+_STOP = re.compile(
+    r"\b(cy|20\d\d|proposed|final|rule|for|the|of|and|to|in|a|an|under|pfs|physician fee schedule)\b"
+)
+_SKIP = ("authority citation", "on page", "column", "paragraph")
+
+
+def stem(title: str) -> str:
+    norm = re.sub(r"[^a-z0-9 ]", " ", title.lower())
+    norm = _STOP.sub(" ", norm)
+    return re.sub(r"\s+", " ", norm).strip()
+
+
+def candidates(store, *, since: int) -> list[tuple[int, str, str]]:
+    """``(n_rule_years, stem, example title)`` sorted by n desc."""
+    from pfs.descriptors import rule_year_of
+
+    con = store._con()  # noqa: SLF001
+    years: dict[str, int] = {}
+    for key, title, published in con.execute(
+        "SELECT key, title, date_published FROM items WHERE item_type = 'rule'"
+    ).fetchall():
+        # rule_year_of reads the payment year from the title, else falls
+        # back to publication year + 1 (PFS rules publish July–December).
+        years[key] = int(rule_year_of(title or "", published or ""))
+    recent = {k for k, y in years.items() if y and y >= since}
+    by: dict[str, set[int]] = defaultdict(set)
+    example: dict[str, str] = {}
+    for key, text in con.execute(
+        "SELECT item_key, text FROM fr_anchors WHERE length(text) < 140"
+    ).fetchall():
+        if key not in recent:
+            continue
+        m = _HEADING.match((text or "").strip())
+        if not m:
+            continue
+        title = m.group(1).strip()
+        s = stem(title)
+        if len(s) < 5 or any(s.startswith(x) or x in s for x in _SKIP):
+            continue
+        by[s].add(years[key])
+        example.setdefault(s, title)
+    return sorted(
+        ((len(v), k, example[k]) for k, v in by.items()), key=lambda r: (-r[0], r[1])
+    )
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser(description=__doc__.split("\n")[0])
+    ap.add_argument("--since", type=int, default=2018)
+    ap.add_argument("--top", type=int, default=200)
+    args = ap.parse_args()
+    from conf import connect
+
+    store = connect.bib()
+    for n, s, ex in candidates(store, since=args.since)[: args.top]:
+        print(f"{n:3d}  {ex}")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

docs/docs/cli/api-serve.md

@@ -1,6 +1,6 @@
 ---
 title: stack api serve
-sidebar_position: 61
+sidebar_position: 62
 ---
 
 # `stack api serve`

docs/docs/cli/api.md

@@ -1,6 +1,6 @@
 ---
 title: stack api
-sidebar_position: 60
+sidebar_position: 61
 ---
 
 # `stack api`

docs/docs/cli/db-comment.md

@@ -1,6 +1,6 @@
 ---
 title: stack db comment
-sidebar_position: 54
+sidebar_position: 55
 ---
 
 # `stack db comment`

docs/docs/cli/db-inspect.md

@@ -1,6 +1,6 @@
 ---
 title: stack db inspect
-sidebar_position: 55
+sidebar_position: 56
 ---
 
 # `stack db inspect`

docs/docs/cli/db.md

@@ -1,6 +1,6 @@
 ---
 title: stack db
-sidebar_position: 53
+sidebar_position: 54
 ---
 
 # `stack db`

docs/docs/cli/docs-build.md

@@ -1,6 +1,6 @@
 ---
 title: stack docs build
-sidebar_position: 57
+sidebar_position: 58
 ---
 
 # `stack docs build`

docs/docs/cli/docs-generate.md

@@ -1,6 +1,6 @@
 ---
 title: stack docs generate
-sidebar_position: 59
+sidebar_position: 60
 ---
 
 # `stack docs generate`

docs/docs/cli/docs-serve.md

@@ -1,6 +1,6 @@
 ---
 title: stack docs serve
-sidebar_position: 58
+sidebar_position: 59
 ---
 
 # `stack docs serve`

docs/docs/cli/docs.md

@@ -1,6 +1,6 @@
 ---
 title: stack docs
-sidebar_position: 56
+sidebar_position: 57
 ---
 
 # `stack docs`

docs/docs/cli/llm-vocab.md

@@ -0,0 +1,21 @@
+---
+title: stack llm vocab
+sidebar_position: 53
+---
+
+# `stack llm vocab`
+
+```
+Usage: stack llm vocab [OPTIONS]                                               
+                                                                                
+ The closed theme vocabulary for comment tagging (#574): validate it and list   
+ its slugs, or show one theme's definition, synonyms and the FR section stems   
+ it was seeded from.                                                            
+                                                                                
+╭─ Options ────────────────────────────────────────────────────────────────────╮
+│ --path        TEXT  A vocabulary file to validate instead of the packaged    │
+│                     one.                                                     │
+│ --slug        TEXT  Show one theme in full.                                  │
+│ --help              Show this message and exit.                              │
+╰──────────────────────────────────────────────────────────────────────────────╯
+```

docs/docs/cli/llm.md

@@ -19,5 +19,9 @@ Usage: stack llm [OPTIONS] COMMAND [ARGS]...
 │ hosts    Show the Ollama fleet: declared VRAM, liveness, models, and which   │
 │          host + model would answer a chat right now.                         │
 │ serve    Serve the SSO-guarded chat UI (llm.api:app).                        │
+│ vocab    The closed theme vocabulary for comment tagging (#574): validate it │
+│          and list its slugs, or show one theme's definition, synonyms and    │
+│          the FR                                                              │
+│          section stems it was seeded from.                                   │
 ╰──────────────────────────────────────────────────────────────────────────────╯
 ```

docs/docs/cli/mail-attach-smarthost.md

@@ -1,6 +1,6 @@
 ---
 title: stack mail attach-smarthost
-sidebar_position: 102
+sidebar_position: 103
 ---
 
 # `stack mail attach-smarthost`

docs/docs/cli/mail-dkim-export.md

@@ -1,6 +1,6 @@
 ---
 title: stack mail dkim-export
-sidebar_position: 101
+sidebar_position: 102
 ---
 
 # `stack mail dkim-export`

docs/docs/cli/mail-dns.md

@@ -1,6 +1,6 @@
 ---
 title: stack mail dns
-sidebar_position: 100
+sidebar_position: 101
 ---
 
 # `stack mail dns`

docs/docs/cli/mail-down.md

@@ -1,6 +1,6 @@
 ---
 title: stack mail down
-sidebar_position: 98
+sidebar_position: 99
 ---
 
 # `stack mail down`

docs/docs/cli/mail-provision.md

@@ -1,6 +1,6 @@
 ---
 title: stack mail provision
-sidebar_position: 96
+sidebar_position: 97
 ---
 
 # `stack mail provision`

docs/docs/cli/mail-rotate-creds.md

@@ -1,6 +1,6 @@
 ---
 title: stack mail rotate-creds
-sidebar_position: 103
+sidebar_position: 104
 ---
 
 # `stack mail rotate-creds`

docs/docs/cli/mail-seed-mailboxes.md

@@ -1,6 +1,6 @@
 ---
 title: stack mail seed-mailboxes
-sidebar_position: 104
+sidebar_position: 105
 ---
 
 # `stack mail seed-mailboxes`

docs/docs/cli/mail-status.md

@@ -1,6 +1,6 @@
 ---
 title: stack mail status
-sidebar_position: 99
+sidebar_position: 100
 ---
 
 # `stack mail status`

docs/docs/cli/mail-up.md

@@ -1,6 +1,6 @@
 ---
 title: stack mail up
-sidebar_position: 97
+sidebar_position: 98
 ---
 
 # `stack mail up`

docs/docs/cli/mail-wire-git.md

@@ -1,6 +1,6 @@
 ---
 title: stack mail wire-git
-sidebar_position: 105
+sidebar_position: 106
 ---
 
 # `stack mail wire-git`

docs/docs/cli/mail.md

@@ -1,6 +1,6 @@
 ---
 title: stack mail
-sidebar_position: 95
+sidebar_position: 96
 ---
 
 # `stack mail`

docs/docs/cli/perf-show.md

@@ -1,6 +1,6 @@
 ---
 title: stack perf show
-sidebar_position: 63
+sidebar_position: 64
 ---
 
 # `stack perf show`

docs/docs/cli/perf.md

@@ -1,6 +1,6 @@
 ---
 title: stack perf
-sidebar_position: 62
+sidebar_position: 63
 ---
 
 # `stack perf`

docs/docs/cli/pfs-cpt-ingest.md

@@ -1,6 +1,6 @@
 ---
 title: stack pfs cpt-ingest
-sidebar_position: 77
+sidebar_position: 78
 ---
 
 # `stack pfs cpt-ingest`

docs/docs/cli/pfs-elements.md

@@ -1,6 +1,6 @@
 ---
 title: stack pfs elements
-sidebar_position: 69
+sidebar_position: 70
 ---
 
 # `stack pfs elements`

docs/docs/cli/pfs-exposure.md

@@ -1,6 +1,6 @@
 ---
 title: stack pfs exposure
-sidebar_position: 74
+sidebar_position: 75
 ---
 
 # `stack pfs exposure`

docs/docs/cli/pfs-families.md

@@ -1,6 +1,6 @@
 ---
 title: stack pfs families
-sidebar_position: 71
+sidebar_position: 72
 ---
 
 # `stack pfs families`

docs/docs/cli/pfs-guidance.md

@@ -1,6 +1,6 @@
 ---
 title: stack pfs guidance
-sidebar_position: 72
+sidebar_position: 73
 ---
 
 # `stack pfs guidance`

docs/docs/cli/pfs-lineage.md

@@ -1,6 +1,6 @@
 ---
 title: stack pfs lineage
-sidebar_position: 70
+sidebar_position: 71
 ---
 
 # `stack pfs lineage`

docs/docs/cli/pfs-reaction.md

@@ -1,6 +1,6 @@
 ---
 title: stack pfs reaction
-sidebar_position: 73
+sidebar_position: 74
 ---
 
 # `stack pfs reaction`

docs/docs/cli/pfs-review.md

@@ -1,6 +1,6 @@
 ---
 title: stack pfs review
-sidebar_position: 76
+sidebar_position: 77
 ---
 
 # `stack pfs review`

docs/docs/cli/pfs-utilization.md

@@ -1,6 +1,6 @@
 ---
 title: stack pfs utilization
-sidebar_position: 75
+sidebar_position: 76
 ---
 
 # `stack pfs utilization`

docs/docs/cli/pfs.md

@@ -1,6 +1,6 @@
 ---
 title: stack pfs
-sidebar_position: 68
+sidebar_position: 69
 ---
 
 # `stack pfs`

docs/docs/cli/prisma-eligible.md

@@ -1,6 +1,6 @@
 ---
 title: stack prisma eligible
-sidebar_position: 89
+sidebar_position: 90
 ---
 
 # `stack prisma eligible`

docs/docs/cli/prisma-export.md

@@ -1,6 +1,6 @@
 ---
 title: stack prisma export
-sidebar_position: 86
+sidebar_position: 87
 ---
 
 # `stack prisma export`

docs/docs/cli/prisma-extract.md

@@ -1,6 +1,6 @@
 ---
 title: stack prisma extract
-sidebar_position: 90
+sidebar_position: 91
 ---
 
 # `stack prisma extract`

docs/docs/cli/prisma-fetch.md

@@ -1,6 +1,6 @@
 ---
 title: stack prisma fetch
-sidebar_position: 92
+sidebar_position: 93
 ---
 
 # `stack prisma fetch`

docs/docs/cli/prisma-flow.md

@@ -1,6 +1,6 @@
 ---
 title: stack prisma flow
-sidebar_position: 91
+sidebar_position: 92
 ---
 
 # `stack prisma flow`

docs/docs/cli/prisma-init.md

@@ -1,6 +1,6 @@
 ---
 title: stack prisma init
-sidebar_position: 85
+sidebar_position: 86
 ---
 
 # `stack prisma init`

docs/docs/cli/prisma-ping-llm.md

@@ -1,6 +1,6 @@
 ---
 title: stack prisma ping-llm
-sidebar_position: 87
+sidebar_position: 88
 ---
 
 # `stack prisma ping-llm`

docs/docs/cli/prisma-run.md

@@ -1,6 +1,6 @@
 ---
 title: stack prisma run
-sidebar_position: 93
+sidebar_position: 94
 ---
 
 # `stack prisma run`

docs/docs/cli/prisma-screen.md

@@ -1,6 +1,6 @@
 ---
 title: stack prisma screen
-sidebar_position: 88
+sidebar_position: 89
 ---
 
 # `stack prisma screen`

docs/docs/cli/prisma-vpn.md

@@ -1,6 +1,6 @@
 ---
 title: stack prisma vpn
-sidebar_position: 94
+sidebar_position: 95
 ---
 
 # `stack prisma vpn`

docs/docs/cli/prisma.md

@@ -1,6 +1,6 @@
 ---
 title: stack prisma
-sidebar_position: 84
+sidebar_position: 85
 ---
 
 # `stack prisma`

docs/docs/cli/rec-list.md

@@ -1,6 +1,6 @@
 ---
 title: stack rec list
-sidebar_position: 65
+sidebar_position: 66
 ---
 
 # `stack rec list`

docs/docs/cli/rec-opps.md

@@ -1,6 +1,6 @@
 ---
 title: stack rec opps
-sidebar_position: 67
+sidebar_position: 68
 ---
 
 # `stack rec opps`

docs/docs/cli/rec-pfs.md

@@ -1,6 +1,6 @@
 ---
 title: stack rec pfs
-sidebar_position: 66
+sidebar_position: 67
 ---
 
 # `stack rec pfs`

docs/docs/cli/rec.md

@@ -1,6 +1,6 @@
 ---
 title: stack rec
-sidebar_position: 64
+sidebar_position: 65
 ---
 
 # `stack rec`

docs/docs/cli/zot-dump-schema.md

@@ -1,6 +1,6 @@
 ---
 title: stack zot dump-schema
-sidebar_position: 79
+sidebar_position: 80
 ---
 
 # `stack zot dump-schema`

docs/docs/cli/zot-fix-dates.md

@@ -1,6 +1,6 @@
 ---
 title: stack zot fix-dates
-sidebar_position: 80
+sidebar_position: 81
 ---
 
 # `stack zot fix-dates`

docs/docs/cli/zot-fix-fields.md

@@ -1,6 +1,6 @@
 ---
 title: stack zot fix-fields
-sidebar_position: 82
+sidebar_position: 83
 ---
 
 # `stack zot fix-fields`

docs/docs/cli/zot-fix-keys.md

@@ -1,6 +1,6 @@
 ---
 title: stack zot fix-keys
-sidebar_position: 81
+sidebar_position: 82
 ---
 
 # `stack zot fix-keys`

docs/docs/cli/zot-verify-parity.md

@@ -1,6 +1,6 @@
 ---
 title: stack zot verify-parity
-sidebar_position: 83
+sidebar_position: 84
 ---
 
 # `stack zot verify-parity`

docs/docs/cli/zot.md

@@ -1,6 +1,6 @@
 ---
 title: stack zot
-sidebar_position: 78
+sidebar_position: 79
 ---
 
 # `stack zot`

docs/superpowers/specs/2026-09-22-llm-tagging-design.md

@@ -0,0 +1,46 @@
+# P35 — closed-vocabulary RAG tagging of rulemaking comments into bib/Zotero
+
+Tracker: milestone P35 (#574 vocabulary, #575 chain + batch runner, #576 write-back + Zotero sync, #577 golden-set gate). Parent spec: `2026-07-16-llm-module-design.md` §P35. Builds on the comment pipeline (P47 dockets/seals, `.state/comments/<docket>/<id>/combined.md` bodies), the pgvector index (`comments` collection, chunk metadata `docket`/`comment_id`/`item_key`/`date`), `llm.classify.closed_vocab_classifier` (one slug or `none` from the largest live host), `llm.pool` (dead-host drop, #796), and the bib tag conventions (`bib.tag` namespaces, `Store.add_tags`, `remove_tag`, nightly `zotero-sync --tag`).
+
+## Goal
+
+Every regulations.gov comment on a PFS docket carries a small set of **theme tags** from a closed vocabulary — which parts of the rule the letter is about — written as `llm:<slug>` tags on its bib item so they reach Zotero with the nightly sync, are filterable in `stack bib query` and the search UI, and give #254–#256 (position classification, stakeholder segmentation, provision mapping) a stable thematic axis. Self-hosted inference only (the llm module never calls a cloud API).
+
+## Decisions
+
+1. **The vocabulary is a versioned YAML file in the repo**, `src/llm/vocab/themes.yaml`, hand-curated. It is *seeded* from the Federal Register section headings of the CY2018–CY2027 PFS rules (mined from `fr_anchors`, `dev/scripts/llm_vocab_candidates.py`) and from the hand code families, but the file is the artifact; a run records the `version` it was tagged with. Slugs are kebab-case, ~60 themes, each with `label`, `definition` (one sentence the model sees), `synonyms` (phrases that appear in letters), and `sections` (the FR heading stems it maps to). Mechanical bib namespaces (`docket:`, `year:`, `rule:`, `reg-docket:`, `enriched:`, `code:`, `family:`) are excluded by construction — they are provenance, not themes.
+2. **Multi-label, evidence-first.** A comment gets zero to `max_tags` (default 4) themes. The chain does not ask the model to pick from 60 slugs cold: it first shortlists candidates by embedding similarity between the comment's chunks and each theme's definition+synonyms (the theme "cards" are embedded once per vocab version), then asks the model a yes/no per candidate with the comment's most relevant chunk as evidence, on the largest live host, `temperature 0`. Anything outside `{yes, no}` counts as `no`. This is the closed-vocab discipline of `llm.classify` applied per theme, and the evidence chunk is stored so a human can audit any tag.
+3. **Resumable, docket-scoped, idempotent.** State lives in bib: `extra_json["llm_tags"] = {"version": v, "model": m, "tags": {slug: {"confidence": c, "evidence_chunk": seq}}, "tagged_at": iso}`. A comment is skipped when its stored `version`+`model`+`content_hash` match; `--force` redoes it. `stack llm tag --docket D [--limit N] [--force]` walks a docket newest-first through the pool (`HostPool` fan-out; a dead host is dropped, #796).
+4. **Write-back replaces only `llm:` tags.** `apply_tags(store, key, slugs)` removes every existing `llm:*` tag on the item and adds the new set — hand tags (`topic:`, `stance:`, …) are never touched. Confidence below `min_confidence` (default 0.6) is recorded in `extra_json` but not written as a tag. The nightly `zotero-sync` already carries tags; `--tag llm:` scopes a manual push.
+5. **Golden set gates fan-out.** `tests/llm/golden_themes.yaml` holds ~200 hand-labelled comments across ≥3 dockets (comment id, expected slugs). `stack llm eval-tags` computes per-slug precision/recall and the abstain rate; CI runs it mocked (the harness only), the nightly `llm-golden` job runs it live. Corpus-wide tagging beyond the pilot docket is blocked until micro-F1 ≥ 0.70 and no theme with support ≥ 5 has precision < 0.5.
+6. **Observability** reuses `llm.metrics` (`dispatch`, a `tagged` counter) — the P34 Grafana panel gains a tagging row later, not in this milestone.
+
+## Data flow
+
+```
+themes.yaml ──embed cards──▶ pgvector "themes" collection (version-stamped)
+comment chunks (pgvector) ──similarity vs cards──▶ shortlist (top 8 themes)
+shortlist × evidence chunk ──yes/no on largest host──▶ tags + confidence
+                                            └──▶ bib extra_json.llm_tags (state)
+                                            └──▶ bib tags llm:<slug> (write-back)
+                                                     └──▶ nightly zotero-sync
+golden_themes.yaml ──▶ stack llm eval-tags ──▶ precision/recall report (gate)
+```
+
+## Modules
+
+- `src/llm/vocab.py` — `Theme` dataclass, `load()` (validates unique slugs, kebab-case, non-empty definitions), `version()` (the file's `version` field), `cards()` (text per theme for embedding).
+- `src/llm/tagger.py` — `shortlist()`, `judge()`, `tag_comment()`, `run()` (batch), `apply_tags()`; pure pieces take injected callables so tests never touch Ollama or pgvector.
+- `src/cli/llm.py` — `vocab` (list/validate), `tag` (batch), `eval-tags` (gate).
+- `dev/scripts/llm_vocab_candidates.py` — the heading miner that seeds a vocab revision (not run in CI).
+
+## Slices
+
+1. **#574** vocabulary YAML + `llm.vocab` + `stack llm vocab` (this slice).
+2. **#575** theme cards in pgvector, shortlist + judge chain, batch runner with state.
+3. **#576** write-back + sync verification on a pilot docket (CMS-2026-2377).
+4. **#577** golden set + `eval-tags` + the gate; then corpus-wide run.
+
+## Out of scope
+
+Position/stance classification (#254's other half — reuse `stance:` from P43 tooling later), stakeholder segmentation (#255), cloud LLMs, tagging rules or the Zotero corpus (comments only).

src/cli/llm.py

@@ -203,3 +203,36 @@ def serve(
     import uvicorn
 
     uvicorn.run("llm.api:app", host=host, port=port, log_level="info")
+
+
+@app.command()
+def vocab(
+    path: str = typer.Option(
+        "", "--path", help="A vocabulary file to validate instead of the packaged one."
+    ),
+    slug: str = typer.Option("", "--slug", help="Show one theme in full."),
+) -> None:
+    """The closed theme vocabulary for comment tagging (#574): validate it
+    and list its slugs, or show one theme's definition, synonyms and the FR
+    section stems it was seeded from."""
+    from llm.vocab import VocabError, load
+
+    try:
+        v = load(path or None)
+    except VocabError as exc:
+        typer.echo(f"invalid vocabulary: {exc}")
+        raise typer.Exit(1) from exc
+    if slug:
+        if slug not in v:
+            raise typer.BadParameter(f"unknown slug {slug!r}")
+        t = v.get(slug)
+        typer.echo(f"{t.tag}  {t.label}")
+        typer.echo(f"  {t.definition}")
+        typer.echo(f"  synonyms: {', '.join(t.synonyms) or '—'}")
+        typer.echo(f"  sections: {' | '.join(t.sections) or '—'}")
+        return
+    for t in v.themes:
+        typer.echo(f"{t.slug:<34} {t.label}")
+    typer.echo(
+        f"vocabulary v{v.version}: {len(v)} themes, {len(v.retired)} retired  ({v.source})"
+    )

src/llm/vocab/__init__.py

@@ -0,0 +1,151 @@
+"""The closed theme vocabulary for comment tagging (P35, #574).
+
+``themes.yaml`` next to this file is the artifact: hand-curated, versioned
+(``version:``), seeded from the CY2018+ PFS rules' section headings
+(``dev/scripts/llm_vocab_candidates.py``). Every tagging run records the
+version it used, so a vocabulary change never silently re-labels old
+work — bump ``version`` when a slug, definition or synonym changes and
+retire slugs into ``retired:`` instead of deleting them.
+
+    load()      → Vocab (validated: unique kebab-case slugs, definitions)
+    Vocab.cards → {slug: text} — what gets embedded per theme (#575)
+    Vocab.tag(slug) → "llm:<slug>", the bib tag label
+"""
+
+from __future__ import annotations
+
+import re
+from dataclasses import dataclass, field
+from importlib import resources
+from pathlib import Path
+from typing import Any, Sequence
+
+import yaml
+
+TAG_NAMESPACE = "llm"
+_SLUG = re.compile(r"^[a-z0-9]+(?:-[a-z0-9]+)*$")
+
+
+@dataclass(frozen=True)
+class Theme:
+    slug: str
+    label: str
+    definition: str
+    synonyms: tuple[str, ...] = ()
+    sections: tuple[str, ...] = ()
+
+    @property
+    def tag(self) -> str:
+        return f"{TAG_NAMESPACE}:{self.slug}"
+
+    @property
+    def card(self) -> str:
+        """The text embedded for shortlisting: label, definition and the
+        phrases letters actually use."""
+        syn = ", ".join(self.synonyms)
+        return f"{self.label}. {self.definition}" + (f" Also: {syn}." if syn else "")
+
+
+@dataclass(frozen=True)
+class Vocab:
+    version: int
+    themes: tuple[Theme, ...]
+    retired: tuple[str, ...] = ()
+    source: str = ""
+    _by_slug: dict[str, Theme] = field(default_factory=dict, repr=False, compare=False)
+
+    def __post_init__(self) -> None:
+        object.__setattr__(self, "_by_slug", {t.slug: t for t in self.themes})
+
+    @property
+    def slugs(self) -> tuple[str, ...]:
+        return tuple(t.slug for t in self.themes)
+
+    def get(self, slug: str) -> Theme:
+        return self._by_slug[slug]
+
+    def __contains__(self, slug: object) -> bool:
+        return slug in self._by_slug
+
+    def __len__(self) -> int:
+        return len(self.themes)
+
+    @property
+    def cards(self) -> dict[str, str]:
+        return {t.slug: t.card for t in self.themes}
+
+    def tag(self, slug: str) -> str:
+        return self.get(slug).tag
+
+
+class VocabError(ValueError):
+    """The YAML is not a valid vocabulary — say exactly what is wrong."""
+
+
+def _theme(raw: dict[str, Any]) -> Theme:
+    slug = str(raw.get("slug") or "").strip()
+    if not _SLUG.match(slug):
+        raise VocabError(f"slug {slug!r} is not kebab-case")
+    definition = str(raw.get("definition") or "").strip()
+    if not definition:
+        raise VocabError(f"{slug}: definition is empty")
+    label = str(raw.get("label") or "").strip() or slug
+    return Theme(
+        slug=slug,
+        label=label,
+        definition=definition,
+        synonyms=tuple(
+            str(s).strip() for s in (raw.get("synonyms") or []) if str(s).strip()
+        ),
+        sections=tuple(
+            str(s).strip() for s in (raw.get("sections") or []) if str(s).strip()
+        ),
+    )
+
+
+def parse(text: str, *, source: str = "") -> Vocab:
+    data = yaml.safe_load(text) or {}
+    if not isinstance(data, dict) or "themes" not in data:
+        raise VocabError(
+            f"{source or 'vocabulary'}: expected a mapping with a `themes` list"
+        )
+    try:
+        version = int(data.get("version", 0))
+    except (TypeError, ValueError) as exc:
+        raise VocabError(f"{source}: version must be an integer") from exc
+    if version < 1:
+        raise VocabError(f"{source}: version must be >= 1")
+    themes = tuple(_theme(t) for t in data.get("themes") or [])
+    if not themes:
+        raise VocabError(f"{source}: no themes")
+    seen: set[str] = set()
+    for t in themes:
+        if t.slug in seen:
+            raise VocabError(f"duplicate slug {t.slug!r}")
+        seen.add(t.slug)
+    retired = tuple(
+        str(s).strip() for s in (data.get("retired") or []) if str(s).strip()
+    )
+    clash = seen & set(retired)
+    if clash:
+        raise VocabError(f"retired slugs still active: {sorted(clash)}")
+    return Vocab(version=version, themes=themes, retired=retired, source=source)
+
+
+def load(path: Path | str | None = None) -> Vocab:
+    """The packaged ``themes.yaml`` (default) or a file on disk."""
+    if path is None:
+        res = resources.files(__name__).joinpath("themes.yaml")
+        return parse(res.read_text(encoding="utf-8"), source=str(res))
+    p = Path(path)
+    return parse(p.read_text(encoding="utf-8"), source=str(p))
+
+
+def is_llm_tag(label: str) -> bool:
+    return label.startswith(f"{TAG_NAMESPACE}:")
+
+
+def split_tags(labels: Sequence[str]) -> tuple[list[str], list[str]]:
+    """``(llm tags, everything else)`` — the write-back only ever touches the first."""
+    ours = [t for t in labels if is_llm_tag(t)]
+    return ours, [t for t in labels if not is_llm_tag(t)]

src/llm/vocab/themes.yaml

@@ -0,0 +1,276 @@
+# Closed theme vocabulary for tagging PFS rulemaking comments (P35, #574).
+#
+# Seeded from the section headings of the CY2018–CY2027 PFS proposed and
+# final rules (dev/scripts/llm_vocab_candidates.py over fr_anchors) and the
+# hand code families; curated by hand. Slugs are stable once published —
+# retire a theme by moving it to `retired:` rather than deleting it, and
+# bump `version` whenever a slug, definition or synonym list changes, since
+# every tag run records the version it used.
+version: 1
+themes:
+  - slug: conversion-factor
+    label: Conversion factor and payment update
+    definition: The annual PFS conversion factor, statutory update percentage, budget-neutrality adjustment, or the overall payment cut or increase.
+    synonyms: [conversion factor, CF, payment update, budget neutrality, statutory update, payment cut, physician payment cut]
+    sections: [Conversion Factor, Physician Fee Schedule Update, Budget Neutrality]
+  - slug: practice-expense
+    label: Practice expense RVUs
+    definition: Practice expense methodology, direct and indirect PE inputs, clinical labor pricing, supply and equipment pricing, or the PE/HR survey data.
+    synonyms: [practice expense, PE RVU, clinical labor, supply pricing, equipment pricing, indirect practice expense, PE/HR, AMA PPIS survey]
+    sections: [Determination of Practice Expense (PE) Relative Value Units (RVUs), Practice Expense Methodology]
+  - slug: work-rvu
+    label: Work RVUs and code valuation
+    definition: Physician work RVUs, RUC recommendations, time assumptions, or the valuation of specific CPT/HCPCS codes.
+    synonyms: [work RVU, RUC recommendation, valuation of specific codes, intraservice time, survey data, code valuation]
+    sections: [Valuation of Specific Codes, Work RVUs]
+  - slug: malpractice-rvu
+    label: Malpractice RVUs
+    definition: Malpractice (MP) RVU methodology, premium data, or specialty risk factors.
+    synonyms: [malpractice RVU, MP RVU, professional liability insurance premium, risk factor]
+    sections: [Determination of Malpractice Relative Value Units (RVUs)]
+  - slug: misvalued-codes
+    label: Potentially misvalued codes
+    definition: Identification, review or revaluation of potentially misvalued services, including the statutory misvalued-code target.
+    synonyms: [misvalued, potentially misvalued codes, misvalued code target, revaluation]
+    sections: [Potentially Misvalued Services Under the PFS]
+  - slug: gpci-localities
+    label: GPCIs and payment localities
+    definition: Geographic practice cost indices, locality definitions, or geographic adjustment of payment.
+    synonyms: [GPCI, geographic practice cost index, payment locality, locality, geographic adjustment, work GPCI floor]
+    sections: [Geographic Practice Cost Indices (GPCIs), Payment Localities]
+  - slug: telehealth
+    label: Medicare telehealth services
+    definition: The Medicare telehealth services list, originating-site and geographic restrictions, audio-only, virtual direct supervision, or telehealth flexibilities after the public health emergency.
+    synonyms: [telehealth, telemedicine, audio-only, originating site, distant site, virtual supervision, telehealth list, home as originating site]
+    sections: [Medicare Telehealth Services, Payment for Medicare Telehealth Services Under Section 1834(m) of the Act]
+  - slug: communication-technology-services
+    label: Communication technology-based services
+    definition: Virtual check-ins, e-visits, remote evaluation of images, interprofessional consultations and other communication technology-based services that are not telehealth.
+    synonyms: [virtual check-in, e-visit, online digital E/M, interprofessional consultation, remote evaluation, CTBS, G2012, 99421]
+    sections: [Communication Technology-Based Services (CTBS)]
+  - slug: remote-monitoring
+    label: Remote physiologic and therapeutic monitoring
+    definition: Remote physiologic monitoring (RPM) or remote therapeutic monitoring (RTM) codes, device supply days, or monitoring time requirements.
+    synonyms: [remote physiologic monitoring, RPM, remote therapeutic monitoring, RTM, 99454, 99457, 98977, 16 days]
+    sections: [Remote Physiologic Monitoring]
+  - slug: evaluation-management
+    label: Evaluation and management visits
+    definition: Office/outpatient, hospital, nursing facility or home E/M visit coding, documentation guidelines, level selection, prolonged services, or split/shared visits.
+    synonyms: [E/M, evaluation and management, office visit, 99213, 99214, prolonged services, split/shared, level selection, documentation guidelines]
+    sections: [Evaluation & Management (E/M) Visits, Evaluation & Management (E/M) Guidelines and Care Management Services]
+  - slug: visit-complexity-add-on
+    label: Visit complexity add-on (G2211)
+    definition: The office/outpatient visit complexity add-on code G2211, its use with modifier 25, or its budget impact.
+    synonyms: [G2211, visit complexity, complexity add-on, longitudinal relationship, modifier 25]
+    sections: [Visit Complexity]
+  - slug: care-management
+    label: Care management services
+    definition: Chronic care management, principal care management, transitional care management, advanced primary care management, or other monthly care-management codes and their consent, staffing and time requirements.
+    synonyms: [chronic care management, CCM, principal care management, PCM, transitional care management, TCM, advanced primary care management, APCM, 99490, 99439, G0556, care management]
+    sections: [Care Management Services, Advanced Primary Care Management]
+  - slug: behavioral-health
+    label: Behavioral health integration and services
+    definition: Behavioral health integration, collaborative care, psychotherapy, marriage and family therapists and mental health counselors, or crisis and safety-planning services.
+    synonyms: [behavioral health, mental health, collaborative care, BHI, psychotherapy, MFT, MHC, safety planning, crisis psychotherapy, community health integration]
+    sections: [Behavioral Health Services]
+  - slug: opioid-treatment
+    label: Opioid treatment programs and substance use disorder
+    definition: Opioid treatment program bundled payments, medications for opioid use disorder, or substance use disorder treatment via telehealth.
+    synonyms: [opioid treatment program, OTP, medication for opioid use disorder, MOUD, buprenorphine, methadone, substance use disorder, SUD]
+    sections: [Requirements for Opioid Treatment Programs (OTP)]
+  - slug: caregiver-training
+    label: Caregiver training services
+    definition: Caregiver training services furnished to a patient's caregivers, with or without the patient present.
+    synonyms: [caregiver training, CTS, 97550, 96202]
+    sections: [Caregiver Training Services]
+  - slug: community-health-social-needs
+    label: Community health integration and social needs
+    definition: Community health integration, principal illness navigation, social determinants of health risk assessment, or community health worker services.
+    synonyms: [community health integration, CHI, principal illness navigation, PIN, social determinants of health, SDOH risk assessment, community health worker, G0019, G0136]
+    sections: [Community Health Integration, Principal Illness Navigation]
+  - slug: palliative-care
+    label: Palliative and serious-illness care
+    definition: Community-based palliative care, advance care planning, hospice interaction, or serious-illness care management.
+    synonyms: [palliative care, advance care planning, ACP, serious illness, hospice, end-of-life]
+    sections: [Request for Information on Community-Based Palliative Care]
+  - slug: supervision
+    label: Supervision requirements
+    definition: Direct, general or virtual supervision of auxiliary personnel, incident-to services, or supervision of diagnostic tests and residents.
+    synonyms: [direct supervision, general supervision, incident to, incident-to, supervision of residents, virtual presence, teaching physician]
+    sections: [Direct Supervision by Interactive Telecommunications Technology, Teaching Physician Documentation Requirements]
+  - slug: non-physician-practitioners
+    label: Non-physician practitioners and scope
+    definition: Nurse practitioners, physician assistants, clinical nurse specialists, therapists or other non-physician practitioners' billing, supervision or scope-of-practice policies.
+    synonyms: [nurse practitioner, physician assistant, PA, NP, advanced practice, scope of practice, CRNA, non-physician practitioner, NPP]
+    sections: [Physician Assistants, Nurse Practitioners]
+  - slug: therapy-services
+    label: Physical, occupational and speech therapy
+    definition: Therapy services, therapy assistants and the CQ/CO modifiers, the KX modifier threshold, plan-of-care certification, or therapy supervision.
+    synonyms: [physical therapy, occupational therapy, speech-language pathology, therapy assistant, CQ modifier, CO modifier, KX modifier, therapy threshold, plan of care]
+    sections: [Therapy Services, Therapy Caps]
+  - slug: drugs-biologicals
+    label: Part B drugs and biologicals
+    definition: Average sales price methodology, drug add-on percentages, biosimilars, discarded-drug refunds, or drug payment under Part B.
+    synonyms: [average sales price, ASP, WAC, biosimilar, discarded drug, JW modifier, JZ modifier, Part B drug, drug payment, 340B]
+    sections: [Part B Drug Payment, Payment for Biosimilar Biological Products]
+  - slug: vaccines-preventive
+    label: Vaccines and preventive services
+    definition: Vaccine administration payment, preventive services coverage, screening services, or the Medicare Diabetes Prevention Program.
+    synonyms: [vaccine administration, preventive services, screening, colorectal cancer screening, Medicare Diabetes Prevention Program, MDPP, hepatitis B vaccine, HIV PrEP]
+    sections: [Medicare Diabetes Prevention Program, Preventive Services]
+  - slug: skin-substitutes
+    label: Skin substitutes and wound care
+    definition: Payment or coding for skin substitute products, cellular and tissue-based products, or wound care services.
+    synonyms: [skin substitute, cellular and tissue-based product, CTP, wound care, Q4 code, amniotic, dermal substitute]
+    sections: [Skin Substitutes]
+  - slug: dental-services
+    label: Dental services
+    definition: Medicare payment for dental services inextricably linked to covered medical services.
+    synonyms: [dental, oral health, dental services, inextricably linked]
+    sections: [Dental Services]
+  - slug: global-surgery
+    label: Global surgical packages
+    definition: Global surgery periods, post-operative visit reporting, or the transfer-of-care modifier for global packages.
+    synonyms: [global surgery, global period, 010-day, 090-day, post-operative visit, modifier 54, modifier 55, transfer of care, G0559]
+    sections: [Global Surgery]
+  - slug: anesthesia
+    label: Anesthesia services
+    definition: The anesthesia conversion factor, anesthesia base units, or anesthesia billing and supervision.
+    synonyms: [anesthesia, anesthesia conversion factor, base units, CRNA, medical direction]
+    sections: [Anesthesia Services, Anesthesia Conversion Factor]
+  - slug: imaging-radiology
+    label: Imaging and radiology
+    definition: Diagnostic imaging payment, appropriate use criteria for advanced imaging, or radiology assistants and supervision.
+    synonyms: [imaging, radiology, appropriate use criteria, AUC, clinical decision support, CT, MRI, radiologist assistant]
+    sections: [Appropriate Use Criteria for Advanced Diagnostic Imaging Services]
+  - slug: laboratory
+    label: Clinical laboratory fee schedule
+    definition: The clinical laboratory fee schedule, private payer rate reporting, specimen collection, or laboratory test payment.
+    synonyms: [clinical laboratory fee schedule, CLFS, private payor rate, PAMA, laboratory test, specimen collection]
+    sections: [Clinical Laboratory Fee Schedule]
+  - slug: ambulance
+    label: Ambulance services
+    definition: The ambulance fee schedule, ground or air ambulance payment, or ambulance cost data collection.
+    synonyms: [ambulance, ambulance fee schedule, ground ambulance, air ambulance, ambulance cost collection]
+    sections: [Ambulance Fee Schedule]
+  - slug: dme-supplies
+    label: Durable medical equipment and supplies
+    definition: DME infusion drugs, DMEPOS competitive bidding, or supplies furnished with physician services.
+    synonyms: [durable medical equipment, DME, DMEPOS, infusion drugs, competitive bidding]
+    sections: [Payment for DME Infusion Drugs]
+  - slug: esrd-dialysis
+    label: ESRD and dialysis services
+    definition: Monthly capitation payment for ESRD-related services, home dialysis, or kidney disease education.
+    synonyms: [ESRD, dialysis, monthly capitation payment, MCP, home dialysis, kidney disease education]
+    sections: [Monthly Capitation Payment (MCP) for ESRD-Related Services]
+  - slug: rhc-fqhc
+    label: Rural health clinics and FQHCs
+    definition: Payment or care-coordination policies for rural health clinics and federally qualified health centers.
+    synonyms: [rural health clinic, RHC, federally qualified health center, FQHC, G0511, all-inclusive rate]
+    sections: [Rural Health Clinics (RHCs) and Federally Qualified Health Centers (FQHCs)]
+  - slug: rural-access
+    label: Rural and underserved access
+    definition: Access to care in rural or underserved areas, workforce shortages, or health equity concerns raised about a policy's effect on access.
+    synonyms: [rural, underserved, health equity, access to care, workforce shortage, health professional shortage area, HPSA]
+    sections: [Health Equity]
+  - slug: shared-savings-program
+    label: Medicare Shared Savings Program
+    definition: Shared Savings Program ACO participation, benchmarking, risk tracks, quality reporting, beneficiary assignment, or advance investment payments.
+    synonyms: [Shared Savings Program, MSSP, accountable care organization, ACO, benchmark, ENHANCED track, BASIC track, beneficiary assignment, advance investment payment]
+    sections: [Medicare Shared Savings Program]
+  - slug: quality-payment-program
+    label: Quality Payment Program and MIPS
+    definition: MIPS categories, MIPS Value Pathways, scoring, performance thresholds, or the QPP reporting requirements.
+    synonyms: [Quality Payment Program, QPP, MIPS, MIPS Value Pathway, MVP, performance threshold, promoting interoperability, cost category, improvement activities]
+    sections: [Updates to the Quality Payment Program]
+  - slug: advanced-apm
+    label: Advanced APMs and APM incentives
+    definition: Advanced alternative payment model participation, qualifying participant thresholds, the APM incentive payment, or the APM conversion factor.
+    synonyms: [advanced APM, alternative payment model, qualifying participant, QP threshold, APM incentive, APM conversion factor]
+    sections: [Advanced Alternative Payment Models]
+  - slug: quality-measures
+    label: Quality measures and reporting
+    definition: Specific quality measures, measure specifications, electronic clinical quality measures, or measure reporting burden.
+    synonyms: [quality measure, eCQM, measure specification, reporting burden, measure set, benchmarking of measures]
+    sections: [Quality Measures]
+  - slug: innovation-models
+    label: Innovation Center models
+    definition: CMS Innovation Center models and demonstrations, including primary care, kidney, oncology or radiation models.
+    synonyms: [Innovation Center, CMMI, model, demonstration, Primary Care First, ACO REACH, oncology care model, radiation oncology model]
+    sections: [Innovation Center Models]
+  - slug: self-referral
+    label: Physician self-referral law
+    definition: Stark law exceptions, the annual code list update, or self-referral compliance.
+    synonyms: [self-referral, Stark, physician self-referral law, designated health services]
+    sections: [Physician Self-Referral Law]
+  - slug: enrollment-program-integrity
+    label: Enrollment and program integrity
+    definition: Provider enrollment requirements, revocations, prior authorization, fraud, waste and abuse controls, or medical review.
+    synonyms: [enrollment, revocation, program integrity, fraud waste and abuse, prior authorization, medical review, overpayment, audit]
+    sections: [Medicare Provider Enrollment, Program Integrity]
+  - slug: coverage-policy
+    label: Coverage and benefit categories
+    definition: Whether a service is a covered Medicare benefit, benefit-category determinations, or national coverage policy raised in the rule.
+    synonyms: [coverage, benefit category, covered service, national coverage determination, NCD, reasonable and necessary]
+    sections: [Coverage]
+  - slug: beneficiary-cost-sharing
+    label: Beneficiary cost sharing
+    definition: Coinsurance, deductibles, or beneficiary out-of-pocket effects of a payment policy.
+    synonyms: [cost sharing, coinsurance, deductible, out-of-pocket, beneficiary liability, copay]
+    sections: [Telehealth Modalities and Cost-sharing]
+  - slug: administrative-burden
+    label: Administrative burden and documentation
+    definition: Documentation, paperwork, reporting or compliance burden imposed on practices, including collection-of-information estimates.
+    synonyms: [administrative burden, documentation burden, paperwork, compliance burden, collection of information, patients over paperwork, prior authorization burden]
+    sections: [Collection of Information Requirements]
+  - slug: regulatory-impact
+    label: Regulatory impact and specialty impacts
+    definition: The regulatory impact analysis, specialty-level payment impact tables, or estimated effects on practices and beneficiaries.
+    synonyms: [regulatory impact analysis, impact table, specialty impact, Table 118, estimated impact, small entities]
+    sections: [Regulatory Impact Analysis]
+  - slug: medicare-economic-index
+    label: Medicare Economic Index and cost weights
+    definition: The Medicare Economic Index, its rebasing or revised cost-share weights, and their effect on RVU allocation.
+    synonyms: [Medicare Economic Index, MEI, cost share weights, rebasing]
+    sections: [The Percentage Change in the Medicare Economic Index (MEI)]
+  - slug: efficiency-adjustment
+    label: Efficiency adjustment to work RVUs
+    definition: An across-the-board efficiency adjustment to work RVUs or intraservice time for non-time-based services.
+    synonyms: [efficiency adjustment, efficiency gains, across-the-board reduction, non-time-based services]
+    sections: [Efficiency Adjustment]
+  - slug: site-of-service
+    label: Site of service and facility payment
+    definition: Facility versus non-facility payment, site-neutral policies, hospital outpatient department comparisons, or indirect PE by setting.
+    synonyms: [site of service, site neutral, facility rate, non-facility rate, hospital outpatient department, HOPD, office-based]
+    sections: [Site of Service]
+  - slug: hospice-home-health
+    label: Hospice and home health interaction
+    definition: Hospice face-to-face encounters, homebound status, home health certification, or physician services interacting with hospice and home health benefits.
+    synonyms: [hospice, home health, face-to-face encounter, homebound, certification of home health]
+    sections: [Clarification of Homebound Status Under the Medicare Home Health Benefit, Telehealth and the Medicare Hospice Face-to-Face Encounter Requirement]
+  - slug: interoperability-health-it
+    label: Health IT and interoperability
+    definition: Certified EHR technology, promoting interoperability requirements, data exchange, or the EHR incentive programs.
+    synonyms: [electronic health record, EHR, certified EHR technology, CEHRT, interoperability, health information exchange, promoting interoperability]
+    sections: [Medicare EHR Incentive Program, Medicaid Promoting Interoperability Program Requirements]
+  - slug: price-transparency
+    label: Price transparency
+    definition: Public disclosure of prices, charges or payment rates to beneficiaries.
+    synonyms: [price transparency, charge transparency, disclosure of prices]
+    sections: [Request for Information on Price Transparency]
+  - slug: covid-phe
+    label: COVID-19 public health emergency flexibilities
+    definition: Policies tied to the COVID-19 public health emergency and their extension or expiration.
+    synonyms: [public health emergency, PHE, COVID-19, pandemic, flexibilities, waiver]
+    sections: [Counting of Resident Time During the PHE for the COVID-19 Pandemic]
+  - slug: data-requests-rfi
+    label: Requests for information
+    definition: Responses to a request for information or comment solicitation rather than a specific proposal.
+    synonyms: [request for information, RFI, comment solicitation, seeking comment, solicit feedback]
+    sections: [Requests for Information]
+  - slug: coding-descriptors
+    label: Coding and code descriptors
+    definition: Creation, deletion or crosswalk of CPT/HCPCS codes, code descriptors, or bundling of services into other codes.
+    synonyms: [new code, deleted code, HCPCS code, G code, crosswalk, bundled, code descriptor, CPT Editorial Panel, unbundle]
+    sections: [Proposed Valuation of Specific Codes]
+retired: []

tests/llm/test_vocab.py

@@ -0,0 +1,114 @@
+"""llm.vocab — the closed theme vocabulary (#574)."""
+
+from __future__ import annotations
+
+import pytest
+
+from llm.vocab import Theme, Vocab, VocabError, is_llm_tag, load, parse, split_tags
+
+GOOD = """
+version: 3
+themes:
+  - slug: telehealth
+    label: Telehealth
+    definition: Medicare telehealth services.
+    synonyms: [telehealth, audio-only]
+    sections: [Medicare Telehealth Services]
+  - slug: care-management
+    definition: Monthly care management codes.
+retired: [old-theme]
+"""
+
+
+class TestParse:
+    def test_reads_themes_and_version(self):
+        v = parse(GOOD, source="t")
+        assert isinstance(v, Vocab) and v.version == 3 and len(v) == 2
+        assert v.slugs == ("telehealth", "care-management")
+        t = v.get("telehealth")
+        assert isinstance(t, Theme) and t.synonyms == ("telehealth", "audio-only")
+        assert (
+            v.get("care-management").label == "care-management"
+        )  # label defaults to slug
+        assert v.retired == ("old-theme",) and "telehealth" in v and "nope" not in v
+
+    def test_cards_and_tags(self):
+        v = parse(GOOD)
+        assert (
+            v.cards["telehealth"]
+            == "Telehealth. Medicare telehealth services. Also: telehealth, audio-only."
+        )
+        assert (
+            v.cards["care-management"]
+            == "care-management. Monthly care management codes."
+        )
+        assert (
+            v.tag("telehealth") == "llm:telehealth"
+            and v.get("telehealth").tag == "llm:telehealth"
+        )
+
+    @pytest.mark.parametrize(
+        "text, msg",
+        [
+            (
+                "version: 1\nthemes:\n  - slug: Bad Slug\n    definition: x\n",
+                "kebab-case",
+            ),
+            (
+                "version: 1\nthemes:\n  - slug: a\n    definition: ''\n",
+                "definition is empty",
+            ),
+            (
+                "version: 1\nthemes:\n  - slug: a\n    definition: x\n  - slug: a\n    definition: y\n",
+                "duplicate slug",
+            ),
+            (
+                "version: 0\nthemes:\n  - slug: a\n    definition: x\n",
+                "version must be >= 1",
+            ),
+            ("version: x\nthemes: []\n", "version must be an integer"),
+            ("version: 1\nthemes: []\n", "no themes"),
+            ("- a\n- b\n", "expected a mapping"),
+            (
+                "version: 1\nthemes:\n  - slug: a\n    definition: x\nretired: [a]\n",
+                "retired slugs still active",
+            ),
+        ],
+    )
+    def test_rejects_bad_files(self, text, msg):
+        with pytest.raises(VocabError, match=msg):
+            parse(text, source="t")
+
+
+class TestPackagedVocab:
+    def test_loads_and_is_sane(self):
+        v = load()
+        assert v.version >= 1 and 40 <= len(v) <= 150
+        assert all(t.definition.endswith(".") for t in v.themes)
+        assert all(len(t.synonyms) >= 1 for t in v.themes)
+        for must in (
+            "telehealth",
+            "care-management",
+            "conversion-factor",
+            "practice-expense",
+            "palliative-care",
+            "skin-substitutes",
+        ):
+            assert must in v
+
+    def test_load_from_path(self, tmp_path):
+        p = tmp_path / "v.yaml"
+        p.write_text(GOOD)
+        assert load(p).source == str(p)
+
+
+class TestTagHelpers:
+    def test_split(self):
+        ours, rest = split_tags(
+            ["llm:telehealth", "topic:palliative", "llm:care-management", "docket:x"]
+        )
+        assert ours == ["llm:telehealth", "llm:care-management"] and rest == [
+            "topic:palliative",
+            "docket:x",
+        ]
+        assert is_llm_tag("llm:x") and not is_llm_tag("llmx:y")