@@ -7,10 +7,10 @@ The committed script is `website/audio/dialogue/<id>.json`; the audio lags it.
 
 The two hosts are the book's historic podcast personas.
 
-| Role | Persona | Voice | Job |
-| --- | --- | --- | --- |
-| `alex` | Kore — instructor | `Kore` (`config.speakers.alex`) | Clear, measured, pedagogical: guides the listener, explains what to do and why, and turns each figure or code block into the idea it carries. |
-| `sam` | Charon — senior engineer | `Charon` (`config.speakers.sam`) | Thoughtful: asks the clarifying question peers would ask, connects to production, challenges over-claims, **restates each beat's takeaway in their own words**. |
+| Role   | Persona                  | Voice                            | Job                                                                                                                                                             |
+| ------ | ------------------------ | -------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `alex` | Kore — instructor        | `Kore` (`config.speakers.alex`)  | Clear, measured, pedagogical: guides the listener, explains what to do and why, and turns each figure or code block into the idea it carries.                   |
+| `sam`  | Charon — senior engineer | `Charon` (`config.speakers.sam`) | Thoughtful: asks the clarifying question peers would ask, connects to production, challenges over-claims, **restates each beat's takeaway in their own words**. |
 
 Roles are epistemic, not a naive-questioner act: Sam never fakes ignorance.
 Voices are bound to roles in `config.ts`, never in the script, so a script is voice-agnostic.
@@ -24,15 +24,15 @@ criteria.
 
 Avoid: enthusiasm and exclamations, laundry lists, dumbing down, marketing hype, circular transitions.
 
-Delivery is *not* authored into the text. The model performs the turns with its natural, energetic
+Delivery is _not_ authored into the text. The model performs the turns with its natural, energetic
 delivery — an added "measured/unhurried" directive reads as sluggish, so none is used; the WER gate
 guards against any instruction leaking into the audio.
 
 ## Why two speakers (evidence, not vibes)
 
 Holding information equal, dialogue alone does not improve comprehension — **reformulation and
-challenging do**. Sam must rephrase and pressure-test, not "yes-and". Fun must be *about the
-topic*: off-topic banter impairs retention (seductive-details effect). See notebook
+challenging do**. Sam must rephrase and pressure-test, not "yes-and". Fun must be _about the
+topic_: off-topic banter impairs retention (seductive-details effect). See notebook
 `research-two-speaker-format`.
 
 ## Turn rules
@@ -48,8 +48,11 @@ topic*: off-topic banter impairs retention (seductive-details effect). See noteb
 ## Content rules
 
 - Write for the ear: short clauses, no visual deixis ("this diagram", "below"); contractions per
-  "Authoring voice".
+  "Authoring voice". **Enforced (hard):** `lint` fails any turn that references an on-page visual —
+  the listener has no page, so the only fix is re-voicing the idea. Vocabulary: `AUDIO_CONFIG.deixis`.
 - Preserve precision: repeat exact identifiers and numbers; never paraphrase a critical term away.
+- **Declared critical terms (per chapter).** The term gate only auto-classifies numbers, all-caps acronyms and identifiers, so a plain word the argument turns on — a named example like _strawberry_ — is invisible to it and a paraphrase can silently drop it. Declare it as `declaredCriticalTerms: [{ "term": "<exact source word>", "reason": "<why it must be spoken>" }]`; the gate then requires it verbatim in the turns (and in the ASR transcript). Each entry must name a real source word, carry a reason, and not restate an auto-classified term, or lint fails instead of ignoring it. `audio:dialogue:source` prints the effective required set (auto + declared).
+- **Described literals (rare exemption).** An arbitrary example literal — e.g. a random hash inside a quoted benchmark prompt — may be described ("a short random prefix") instead of spelled aloud: the book text keeps the literal, only the ear is spared. Declare it in the script as `describedLiterals: [{ "term": "<exact literal>", "reason": "<why it is arbitrary>" }]`. Never for semantic numbers, identifiers or flags (scores, versions, model names, CLI flags): those stay spoken verbatim. Lint exempts a declared term from the spoken-critical-term requirement, but fails the chapter when a declared term is not a critical term of the source or its reason is empty — so the field cannot hide a term that matters.
 - **Coverage is enforced (hard fail).** `audio:dialogue:lint`, `audio:build` and `audio:verify` all
   refuse a chapter whose turns leave any source prose section, figure or code block uncovered, or any
   critical term unsaid. Author from `audio:dialogue:source` output, not from memory.
@@ -63,27 +66,56 @@ topic*: off-topic banter impairs retention (seductive-details effect). See noteb
   layout, color or position — "a gear labelled LLM", "a violet box", "a row of chips", "three panels in
   a row". A listener cannot see it, and it reads as a diagram walkthrough. Translate the figure into the
   one idea it exists to fix, in the intuitive, story language a listener can hold: why it matters and what
-  it changes. The same rule governs `alex`'s figure-anchored turns.
+  it changes. The same rule governs `alex`'s figure-anchored turns. **Enforced (hard):** extraction fails
+  a `narration` that references a visual (`deixis.ts`), so it can never leak into the draft. Two shapes of
+  the same defect are fatal, on `narration` and on the authored voice alike:
+  - **The page register.** `AUDIO_CONFIG.deixis` carries the layout nouns a walkthrough leans on (`panel`,
+    `tile`, `chip`, `ribbon`, `row`, `column`), a position word glued to one of them ("left panel", "top
+    tile"), and `labelled`/`labeled` ("a gear labelled LLM"). Bare positions stay legal ("the right
+    answer", "the middle ground"), and so does `labelled` in its machine-learning sense ("labeled
+    examples").
+  - **Text drawn inside the figure.** The words a diagram paints are declared once, in
+    `website/src/components/VisualElements/figureDrawnLabels.ts`, keyed by the component that draws them.
+    Naming one out loud ("the row that says trust earned") fails: extraction for the `narration`, `lint`
+    for a figure-anchored turn. Edit the artwork and that map moves with it, or its drift test fails.
+    When you edit figure copy, update the `figureDrawnLabels.ts` map in the same change.
 - **Narration adds, never echoes.** A figure or code narration must carry something the prose beside it does
   not already say; restating the neighbour's definition or list makes the audio say the same thing twice.
   Paraphrase is fine — `sam` restating `alex` in different words is the format — but the wording must change,
-  not only the sentence order. Lint fails any two turns in the chapter that share a ≥20-character run of content
-  words (function words ignored) — a repeat is a defect wherever it sits, so a closing recap must paraphrase too —
-  and `audio:dialogue:source` warns when a figure/code narration's content words are largely
-  the adjacent prose's.
+  not only the sentence order. One fact belongs on one surface: an `alex` figure-anchored turn voices the
+  consequence the figure exists to fix, never a restatement of the heading prose above it (that is the
+  heading→figure→recap doubling the repeat gate keeps catching). One fatal gate enforces it at text time,
+  before any audio is rendered:
+  - **Verbatim run.** Any two turns that share a ≥20-character run of content words (function words ignored)
+    fail: a repeat is a defect wherever it sits, so a closing recap must paraphrase too.
+  - **Vocabulary echo (advisory).** `lint` WARNs when same-speaker turns reach the IDF-cosine lexical
+    overlap threshold (`lexical.ts`, `AUDIO_CONFIG.redundancy.lexical.cosineMin`) — the same fact said
+    again in largely the same words, even reordered so no contiguous run survives. It is **advisory, not
+    fatal**: the bag-of-words metric measures lexical overlap, not meaning, and on the real corpus its
+    false positives outscore a genuine paraphrase, so no threshold separates them. Every cross-speaker
+    pair passes this cosine gate (`sam` restating `alex` is the format). Fired pairs are grouped into fact
+    clusters, so the fix is to consolidate the FACT to one surface and give the other turn a new claim —
+    not to reword each pair. The IDF weights are computed live over every chapter in the book
+    (`wholeBookCorpus`), so editing one chapter can shift another chapter's scores.
+- **Document-level lexical echoes are WARNs.** `docRedundancyProblems` surfaces repeated content across prose and
+  figure/code narration for editorial review; they do not block lint or rendering. Captions are page text and
+  are never spoken. Same-speaker dialogue echo is an advisory WARN too; only verbatim copied dialogue is fatal.
+- **The dialogue is not a mirror of the doc.** Coverage requires a turn per section, figure and code block,
+  not a restatement of each. When the reading text legitimately revisits a fact across sections (progressive
+  disclosure, standalone sections), voice it once in the dialogue and give the other surface a new angle.
 - Define a term on first use. Recap at the end of a section.
 - Keep the chapter's own structure: opener = the stakes; each heading = one beat; recap = Sam's
   restatement. Anchor every turn to the heading it explains so the player can highlight it.
 
 ## Structure mapping
 
-| Source | Turns |
-| --- | --- |
-| chapter title | one `alex` turn opening the episode |
-| `opener` | `alex` states the stakes / the question |
-| heading section | `alex` delivers; `sam` challenges + restates |
-| figure / code | `alex` delivers the meaning the figure carries; `sam` asks what it looks like in practice |
-| `recap` | `sam` restates the whole takeaway |
+| Source          | Turns                                                                                     |
+| --------------- | ----------------------------------------------------------------------------------------- |
+| chapter title   | one `alex` turn opening the episode                                                       |
+| `opener`        | `alex` states the stakes / the question                                                   |
+| heading section | `alex` delivers; `sam` challenges + restates                                              |
+| figure / code   | `alex` delivers the meaning the figure carries; `sam` asks what it looks like in practice |
+| `recap`         | `sam` restates the whole takeaway                                                         |
 
 Heading rows exist only for markdown headings: `doc.toc` and the extractor read the markdown AST,
 so a component-rendered `<h1..6>` is invisible to both. Never render heading tags from components —
@@ -95,9 +127,18 @@ see `README.md` "Sources of truth".
 {
   "chapterId": "<doc id>",
   "sourceHash": "<sha256 of the doc source this was authored from>",
+  // Plain source words the pattern gate cannot see and the voice must speak verbatim.
+  // "declaredCriticalTerms": [{ "term": "<exact source word>", "reason": "<why it must be spoken>" }],
+  // Rare: arbitrary example literal only, never a semantic number/identifier/flag.
+  // "describedLiterals": [{ "term": "<exact literal>", "reason": "<why it is arbitrary>" }],
   "turns": [
-    { "id": "t1", "speaker": "alex", "anchor": { "kind": "heading", "id": null }, "text": "..." }
-  ]
+    {
+      "id": "t1",
+      "speaker": "alex",
+      "anchor": { "kind": "heading", "id": null },
+      "text": "...",
+    },
+  ],
 }
 ```
 
@@ -112,10 +153,11 @@ see `README.md` "Sources of truth".
 `sourceHash` is the hash of the chapter content the script was authored from. The build recomputes it
 from the current document and **warns** (and marks the chapter `DRIFT` in `audio:list`) when they
 differ — the book moved on, so re-author before publishing. Drift is a **warning** for
-`audio:build`/`audio:verify` (the audio legitimately lags the text). Every other violation — missing
+`audio:build`/`audio:verify` (the audio legitimately lags the text) but **fatal** for
+`audio:dialogue:lint` (lint runs on authored text, never on lagging audio, so a drifted script means
+the author has not re-based). Every other violation — missing
 narration, uncovered section/figure/code, missing critical term, bad turn ids — is **fatal** to
-`audio:build` and `audio:verify`; `audio:dialogue:lint` treats drift as fatal too, so it stays the
-strict pre-publish gate.
+`audio:build`, `audio:verify` and `audio:dialogue:lint` alike.
 
 ## Chunking (why turns are grouped during render)
 
 
@@ -11,15 +11,23 @@
 import { AUDIO_CONFIG, assertRenderConfig } from '../../website/src/audiobook/config.ts';
 import { AUDIO_DOC_IDS, chapterMeta } from '../../website/src/audiobook/docs.ts';
 import {
-  chunkTurns,
   dialogueBloatWarnings,
+} from '../../website/src/audiobook/dialogueCoverage.ts';
+import {
+  dialogueLexicalHits,
+  docRedundancyProblems,
+} from '../../website/src/audiobook/dialogueLexical.ts';
+import {
+  chunkTurns,
   hardViolations,
-  isDialogueChapter,
   isDriftViolation,
   loadDialogue,
-  loadDialogueFile,
   narrationOverlapWarnings,
-} from '../../website/src/audiobook/dialogue.ts';
+} from '../../website/src/audiobook/dialogueScript.ts';
+import {
+  isDialogueChapter,
+  loadDialogueFile,
+} from '../../website/src/audiobook/dialogueTypes.ts';
 import { criticalTerms } from '../../website/src/audiobook/wer.ts';
 import { dialogueStream, chunkState, discardChunk, printDialoguePlan } from './dialogue.mjs';
 import { encodeChapter } from './encode.mjs';
@@ -185,11 +193,13 @@ const withId = (id, problem) => (problem.startsWith(`${id}:`) ? problem : `${id}
 /**
  * Dialogue lint: extraction + structural/term coverage + drift + chunk cap are
  * hard failures (they mean the script is not the source); section bloat is a
- * WARN line that does not fail.
+ * WARN line that does not fail. Drift is fatal here — lint runs on authored
+ * text, never on lagging audio — while build/verify only warn on it.
  */
 function lintDialogue(id, loaded) {
   const chunks = chunkTurns(loaded.script.segments, AUDIO_CONFIG.dialogue.maxChunkChars);
   const problems = [...loaded.violations, ...chunkProblems(chunks)];
+  loaded.warnings.forEach((warning) => warn('lint', `WARN  ${warning}`));
   dialogueBloatWarnings(chapterMeta(id), loadDialogueFile(id)).forEach((warning) => warn('lint', `WARN  ${warning}`));
   log('lint', `${id}: ${problems.length} problem(s) across ${loaded.script.segments.length} turns in ${chunks.length} chunk(s)`);
   // Problems are not echoed here: `main` prints every chapter problem once as `FAIL` lines.
@@ -202,8 +212,12 @@ function printSource(id) {
   console.log(`chapter ${id}: ${meta.title}`);
   meta.headings.forEach((heading) => console.log(`  heading ${heading.id} → ${heading.title}`));
   meta.blocks.forEach((block, index) => console.log(`  block ${index} ${block.kind} ${JSON.stringify(block.anchor)} ${block.text}`));
-  console.log(`  criticalTerms ${[...new Set(criticalTerms(meta.blocks.map((block) => block.text).join(' ')))].join(', ')}`);
+  const data = loadDialogueFile(id);
+  const declared = (data?.declaredCriticalTerms ?? []).map((entry) => entry.term);
+  console.log(`  criticalTerms ${[...new Set([...criticalTerms(meta.blocks.map((block) => block.text).join(' ')), ...declared])].join(', ')}`);
   narrationOverlapWarnings(meta).forEach((warning) => warn('source', `WARN  ${warning}`));
+  docRedundancyProblems(id, meta).forEach((warning) => warn('source', `WARN  ${warning}`));
+  if (data) dialogueLexicalHits(id, data).filter((hit) => hit.score > 0).slice(0, 5).forEach((hit) => warn('source', `WARN  lexical ${hit.a} and ${hit.b} (${hit.distance} turn(s) apart): ${Math.round(hit.score * 100)}% shared content${hit.phrase ? ' [shared phrase]' : ''} (${hit.shared.slice(0, 8).join(', ')})`));
 }
 
 async function runChapter(id, options) {
@@ -274,13 +288,11 @@ function preflightOrExit(options) {
 }
 
 /**
- * `--all` build/verify skips chapters without a committed dialogue script instead
- * of failing them one by one: audio legitimately lags the text, so an unauthored
- * chapter is not an error. An explicitly named unauthored chapter still fails loudly
- * in `loadChapter`. The skip is announced, never silent.
+ * `--all` skips chapters without committed dialogue; explicit chapter selection
+ * still reaches `loadChapter` and fails loudly when unauthored.
  */
 function authoredScope(run) {
-  if (run.options.command !== 'build' && run.options.command !== 'verify') return run.ids;
+  if (!run.options.all || !['build', 'verify', 'lint'].includes(run.options.command)) return run.ids;
   const authored = run.ids.filter(isDialogueChapter);
   const skipped = run.ids.length - authored.length;
   if (skipped) log(run.options.command, `skipping ${skipped} unauthored chapter(s)`);
 
@@ -10,7 +10,7 @@
  */
 import { rmSync } from 'node:fs';
 import { AUDIO_CONFIG } from '../../website/src/audiobook/config.ts';
-import { chunkTurns, turnSliceBounds, turnStartMs } from '../../website/src/audiobook/dialogue.ts';
+import { chunkTurns, turnSliceBounds, turnStartMs } from '../../website/src/audiobook/dialogueScript.ts';
 import { dialogueChunkCacheKey } from '../../website/src/audiobook/hash.ts';
 import { renderDialogueChunk } from './render.mjs';
 import { cacheFile, exists, log, pcmBytesForMs, pcmDurationMs, readBytes, readJsonIfExists, table, transcribe, warn, writeFile, writeJson } from './report.mjs';
 
@@ -12,7 +12,7 @@
  * in `inlineData` — no server-side WAV/MP3.
  */
 import { AUDIO_CONFIG } from '../../website/src/audiobook/config.ts';
-import { dialogueTranscript, roleLabel } from '../../website/src/audiobook/dialogue.ts';
+import { dialogueTranscript, roleLabel } from '../../website/src/audiobook/dialogueScript.ts';
 import { gemini, withRetry } from './report.mjs';
 
 const MIME_RATE = /rate=(\d+)/;