@atlaskit/editor-plugin-autocomplete 8.0.0 → 9.0.0

package/src/pm-plugins/text-predictor.ts

@@ -3,7 +3,7 @@
  *
  * Two prediction modes:
  *   1. Word boundary → bigram-based next-word suggestion (grammar-filtered)
- *   2. Mid-word (≥3 chars) → trie prefix search → scoring pipeline → top result
+ *   2. Mid-word (≥1 char) → grouped prefetch; ≥3 chars + exact evidence → ghost result
  *
  * Scoring is delegated to scoring-pipeline.ts which handles:
  *   Stage 1 (semantic + frequency), grammar filter, Stage 2 (optional LM re-ranking).
@@ -17,14 +17,48 @@
 
 import { EXPERIENCE_NAME, failExp, startExp, succeedExp } from '../analytics/ufo';
 
-// The vocabulary, L3 and word-index payloads are fetched from the CDN in
-// loadDefaultVocabulary / loadVectorsAsync so their (large) contents never enter
-// the editor's bundle and only load when autocomplete is initialised.
-import { ARTIFACT_NAME, fetchArtifactJson, getArtifactUrl } from './artifacts-manifest';
-import { isAutocompleteDebugEnabled } from './debug-mode';
-import { rankCandidates, STAGE1_WEIGHT, STAGE2_WEIGHT, MIN_STAGE1_SCORE } from './scoring-pipeline';
-import type { ScoringCandidate } from './scoring-pipeline';
-import { getStoredContextVector, getStoredLmLogits } from './slow-lane-client';
+import { fetchAutocompleteArtifactBinary, fetchAutocompleteArtifactJson } from './artifact-loader';
+import { ARTIFACT_NAME, type ArtifactName } from './artifacts-manifest';
+import {
+	createCanonicalContextPositionCache,
+	deriveCanonicalCandidateContext,
+	deriveWhitespaceBoundaryContext,
+	logSoftmaxAt,
+	logSumExp,
+	selectBoundaryPrimeRequests,
+	type CanonicalCandidateContext,
+	type CanonicalLmEvidence,
+} from './canonical-lm-scoring';
+import {
+	CTC_STYLES,
+	ctcSection,
+	ctcTag,
+	isAutocompleteDebugEnabled,
+	isAutocompleteDebugVerbose,
+} from './debug-mode';
+import {
+	loadGrammarDataAsync,
+	rankCandidates,
+	STAGE1_WEIGHT,
+	STAGE2_WEIGHT,
+	MIN_STAGE1_SCORE,
+	MIN_WINNER_MARGIN,
+} from './scoring-pipeline';
+import type { ScoringCandidate, ScoredCandidate, TermType } from './scoring-pipeline';
+import {
+	getBoundaryLmState,
+	getCanonicalSurfaceCount,
+	getCanonicalSurfaceTokenIds,
+	getDefaultSlowLaneClientStatus,
+	getProgressiveSurfaceEvidence,
+	getStoredContextInput,
+	getStoredContextVector,
+	getStoredLmLogits,
+	getSurfaceScore,
+	isCanonicalSurfaceScoringSupported,
+	primeBoundaryLm,
+	requestProgressiveSurfaceScores,
+} from './slow-lane-client';
 
 // ─── Constants ───────────────────────────────────────────────────────────────
 
@@ -36,20 +70,178 @@ const WHITESPACE_SPLIT_REGEX = /\s+/;
 const SENTENCE_BOUNDARY_REGEX = /[\n.?!]+/;
 // eslint-disable-next-line require-unicode-regexp
 const TRAILING_WHITESPACE_REGEX = /\s$/;
+const TRAILING_SURFACE_TOKEN_REGEX = /[\p{L}\p{N}_'-]+$/u;
+const SURFACE_TOKEN_REGEX = /[\p{L}\p{N}_'-]+/gu;
+const ONLY_WHITESPACE_REGEX = /^\s+$/u;
 
-const MIN_PREFIX_LENGTH = 3;
+const UNIGRAM_PREFETCH_MIN_PREFIX_LENGTH = 1;
+const MULTIWORD_PREFETCH_MIN_PREFIX_LENGTH = 1;
+const DISPLAY_MIN_PREFIX_LENGTH = 3;
 const MAX_CANDIDATES = 200;
 const CONTEXT_WORDS = 10;
-const MIN_SCORE_THRESHOLD = 0.35;
 const L3_BASELINE_FREQ = 0.001;
+// Max tokens in a completion term (bigram=2, phrase=3). Mirrors the producer's
+// `phrase_max_words` default so the FE window never exceeds what was mined.
+const PHRASE_MAX_WORDS = 3;
+// Per-window cap on phrase-trie subtree collection (keeps the hot path bounded).
+const MAX_PHRASE_CANDIDATES = 50;
+/**
+ * Share of its context shortlist's probability mass the model must put on a
+ * surface before that surface may be shown.
+ *
+ * Candidates competing at the same boundary are normalised into a posterior
+ * over that shortlist, so this reads directly as "the model is at least this
+ * sure". Measuring distance from the best candidate instead cannot express
+ * being unsure: the leader scores full marks whether it won by a nose or a
+ * mile, which pushes the entire question of how contested a context is onto
+ * `MIN_WINNER_MARGIN` — and there the blended score's only remaining spread is
+ * the corpus prior, the weakest ranking signal we have.
+ *
+ * Each extra token raises the bar. A wrong multi-word ghost costs the reader
+ * more to notice and undo than a wrong single word, and it is offered from the
+ * same keystroke, so precision has to be bought per token of commitment.
+ */
+const MIN_LM_POSTERIOR: Record<TermType, number> = {
+	word: 0.5,
+	bigram: 0.6,
+	phrase: 0.7,
+};
+/**
+ * How many scored candidates a context must hold before its leader may show.
+ *
+ * A posterior is a share of a shortlist, so a shortlist of one hands its only
+ * member 1.0 whatever the model thinks of it — `logSumExp` of a single value
+ * returns that value. Such a ghost sits in the top bucket without having beaten
+ * anything, and it is the gate's blind spot rather than a case the gate ruled
+ * on: 1 in 14 of them were accepted against 34% for word ghosts overall, and
+ * every phrase shown through this hole was rejected.
+ *
+ * Counted over pool membership rather than over the normaliser, because those
+ * two deliberately differ. Where a pool holds a surface and its own extension
+ * the chain rule leaves one term in the denominator, and the resulting 1.0 is
+ * earned — whichever the user meant, the shorter form is a correct ghost.
+ *
+ * Waived where the context never held a second candidate at all. The 1-in-14
+ * acceptance above was measured over a population dominated by pools that read
+ * as one because rivals had not come back, and it is that absence of evidence
+ * the floor is for. A vocabulary offering a single continuation of a long
+ * distinctive prefix is the opposite situation, and it is also where a ghost
+ * saves the most keystrokes, so refusing it spends the most to learn the least.
+ */
+const MIN_SCORED_POOL_SIZE = 2;
+/**
+ * Mean per-token log-probability a surface must hold, in nats, independently of
+ * anything it was competing against.
+ *
+ * Every other display test is relative: a posterior is a share of a shortlist,
+ * and a winner margin is a distance from a rival. So a surface reaches the
+ * screen by being the best of what happened to be in its pool, and nothing ever
+ * asks whether the model finds it plausible at all. That gap was tolerable while
+ * little was being read, and stopped being tolerable once the scheduler started
+ * reading roughly twice as much per decision: shown-per-decision went from 3.2%
+ * to 7.4% with no threshold moved, and acceptance of what showed fell from 32%
+ * to 24.5%. More reading does not weaken a relative test — it hands it far more
+ * pools to be the winner of, and the marginal winner is worse than the average.
+ *
+ * Set at the point where a uniform distribution over the model's vocabulary
+ * sits: 49,152 tokens, so `ln(1/49152)` is about -10.8 nats. Below it the model
+ * assigns the surface less mass than a token drawn at random, which is the only
+ * statement about a surface that needs no reference to what it competed with.
+ *
+ * It was first tried at -8, chosen from the bucket where acceptance was measured
+ * to fall away, and that reasoning does not survive more data. Acceptance against
+ * this quantity turned out to be a smooth slope — about 19% below -8, 26% from -8
+ * to -6, then flat near 35% above - with no discontinuity to place a threshold
+ * at. A floor on a smooth slope is a coverage dial rather than a test: -8 removed
+ * a band converting at 26% and cost 3.7 points of show rate, for an acceptance
+ * movement no sample of this size can resolve. Anchoring to uniform instead makes
+ * the value an argument about what is indefensible rather than a point picked off
+ * a histogram, which is what stops it from being retuned every batch.
+ *
+ * Applied to the mean over the verified prefix, the same quantity
+ * `ghostsByMeanLogProb` buckets, and deliberately not to
+ * `meanTokenLogProbabilityUpperBound`. The bound is the right instrument for
+ * keeping a candidate in the race, where dropping something that could still
+ * clear the bar would be an error, but it is systematically higher than the
+ * quantity acceptance is measured against. Judging a surface on the prefix that
+ * was actually read is also what every other test here does.
+ *
+ * One value across all three shapes. Splitting it per shape needs an acceptance
+ * curve per shape, and 19 bigrams and 1 phrase across five sessions cannot
+ * support one.
+ */
+const MIN_MEAN_TOKEN_LOG_PROBABILITY = -10.8;
+// Tab replaces the whole remaining suffix, so an N-character completion saves
+// N-1 keystrokes: at one character it saves nothing at all and at two it saves
+// one, which does not repay noticing the ghost and reaching for Tab. Offering
+// them anyway spends the acceptance-rate denominator on completions nobody
+// wants and teaches the user to stop reading ghosts. `DISPLAY_MIN_PREFIX_LENGTH`
+// is the same guard on the side the user has already typed.
+const MIN_SUGGESTION_LENGTH = 3;
+/**
+ * How many leading tokens of a surface must be scored before it may be shown.
+ *
+ * Scoring a token costs one model round trip, and a round trip costs about the
+ * same no matter how much is in it, so the price of a candidate is its token
+ * count. Demanding every token therefore caps what can ever reach the screen at
+ * roughly two tokens inside the decision budget — which is why no phrase has
+ * ever been displayed. Verifying a fixed prefix makes the cost of a candidate
+ * independent of its length.
+ */
+const REQUIRED_VERIFIED_TOKENS = 2;
+/**
+ * Extra winner margin demanded of a surface judged on a verified prefix rather
+ * than on every one of its tokens, to buy back the precision given up by
+ * leaving the tail unread.
+ */
+const PARTIAL_EVIDENCE_MARGIN_PREMIUM = 0.08;
+const ARBITRATION_MODE = 'confidence-v2';
+const DEBUG_TEXT_TAIL_CHARS = 120;
 
 // ─── Types ───────────────────────────────────────────────────────────────────
 
+/**
+ * Which constraint stopped an evaluation from putting a ghost on screen.
+ *
+ * Each one implies different work: `below-posterior-gate` is a threshold to
+ * calibrate, `winner-margin` is two candidates the model cannot separate,
+ * `cold-competitor` and `unresolved-rival` are scheduling, and `no-candidate`
+ * is vocabulary coverage.
+ */
+type CtcAbstainReason =
+	| 'below-posterior-gate'
+	| 'cold-competitor'
+	| 'empty-completion'
+	| 'implausible-surface'
+	| 'lone-candidate'
+	| 'missing-artifact'
+	| 'no-candidate'
+	| 'no-evidence'
+	| 'prefetch'
+	| 'short-completion'
+	| 'unresolved-rival'
+	| 'winner-margin';
+
 interface Candidate {
 	node: TrieNode;
 	word: string;
 }
 
+/**
+ * A candidate paired with the length of the already-typed prefix it completes.
+ * For a single word this is the current partial token length; for a phrase it
+ * is the length of the matched multi-word window (previous words + partial),
+ * so the ghost tail is `word.slice(matchedPrefixLen)`.
+ */
+interface MatchedCandidate {
+	matchedPrefixLen: number;
+	node: TrieNode;
+	surfaceStart: number;
+	word: string;
+}
+
+interface CanonicalMatchedCandidate extends MatchedCandidate, CanonicalCandidateContext {}
+
 export interface WeightedTerm {
 	authorFreq: number;
 	docFreq: number;
@@ -61,6 +253,14 @@ export interface TenantVocabulary {
 	terms: WeightedTerm[];
 }
 
+/** Optional per-term metadata carried by bigram/phrase completion terms. */
+interface TermMeta {
+	association?: number | null;
+	headPos?: string | null;
+	posSeq?: string[] | null;
+	termType?: TermType;
+}
+
 interface VectorStore {
 	dim: number;
 	float32: Float32Array;
@@ -74,6 +274,12 @@ class TrieNode {
 	docFreq: number = 0;
 	authorFreq: number = 0;
 	sessionFreq: number = 0;
+	// Term-type + phrase metadata. Plain words keep `termType='word'` and null
+	// metadata; bigram/phrase nodes carry the producer-shipped POS + association.
+	termType: TermType = 'word';
+	posSeq: string[] | null = null;
+	headPos: string | null = null;
+	association: number | null = null;
 }
 
 class WeightedWordTrie {
@@ -81,7 +287,13 @@ class WeightedWordTrie {
 	/** Highest tenantFreq seen — used to normalize freq scores at query time */
 	maxTenantFreq: number = 1;
 
-	insert(word: string, tenantFreq: number, docFreq: number, authorFreq: number): void {
+	insert(
+		word: string,
+		tenantFreq: number,
+		docFreq: number,
+		authorFreq: number,
+		meta?: TermMeta,
+	): void {
 		let node = this.root;
 		for (const char of word.toLowerCase()) {
 			let next = node.children.get(char);
@@ -95,6 +307,10 @@ class WeightedWordTrie {
 		node.tenantFreq = tenantFreq;
 		node.docFreq = docFreq;
 		node.authorFreq = authorFreq;
+		node.termType = meta?.termType ?? 'word';
+		node.posSeq = meta?.posSeq ?? null;
+		node.headPos = meta?.headPos ?? null;
+		node.association = meta?.association ?? null;
 		if (tenantFreq > this.maxTenantFreq) {
 			this.maxTenantFreq = tenantFreq;
 		}
@@ -180,6 +396,15 @@ const wordTrie = new WeightedWordTrie();
 // L3 Trie (General English Fallback)
 const l3Trie = new WeightedWordTrie();
 
+/**
+ * L2 phrase/bigram trie. Kept SEPARATE from `wordTrie` so single-word
+ * completion + L3 gap-fill stay untouched; multi-word units are only surfaced
+ * by the (12b) multiword prefix-window matcher. Keys are the full normalized
+ * surface string (spaces included), e.g. `"return on investment"`, so a lookup
+ * is a plain char-trie walk over the whole phrase.
+ */
+const phraseTrie = new WeightedWordTrie();
+
 /**
  * Loads the General English vocabulary.
  * expects a simple array of strings: ["about", "above", "actually", ...]
@@ -190,10 +415,8 @@ export const initL3Vocabulary = (l3Words: string[]): void => {
 		// loses to any domain word in Stage 1, but still scores above 0.
 		l3Trie.insert(word, L3_BASELINE_FREQ, 0, 0);
 	}
-	if (isAutocompleteDebugEnabled()) {
-		// eslint-disable-next-line no-console
-		console.log(`[text-predictor] L3 General English loaded: ${l3Words.length} words`);
-	}
+	recallGeneration++;
+	ctcTag('init', `L3 general English loaded: ${l3Words.length} words`);
 };
 
 // const bigramMap: Map<string, Record<string, number>> = new Map(
@@ -206,9 +429,23 @@ let vectorStore: VectorStore | null = null;
 
 let vectorsLoadStarted = false;
 
+// Phrase/bigram artifacts load independently of the word vocabulary so a
+// missing/late `bigrams.json`/`phrases.json` never breaks word completion.
+let phrasesLoadStarted = false;
+let phrasesLoaded = false;
+let phraseTermCount = 0;
+// Per-type frequency maxima, tracked separately because bigrams and phrases
+// share one trie but have very different absolute frequency ranges (bigrams are
+// far more frequent than 3-word phrases). Normalizing each type by its OWN max
+// stops phrase freq scores from being crushed against the top bigram's count.
+let maxBigramFreq = 1;
+let maxPhraseFreq = 1;
+
 let lastPredictionDebug: {
+	awaitingAsyncEvidence: boolean;
 	contextWords: string[];
 	currentWord: string;
+	decision: string;
 	mode: 'cold' | 'warm';
 	suggestion: string | null;
 	textBefore: string;
@@ -221,6 +458,36 @@ let lastPredictionDebug: {
 	}>;
 } | null = null;
 
+// ── Stabilization (QI-2): post-accept cooldown + whole-surface repetition ────
+// Two guards that stop the accept→echo (`end to end` → `end to end to end`) and
+// generally keep a just-accepted unit from being re-offered:
+//  • repetition: never surface a candidate whose full surface already equals the
+//    run of tokens immediately before the cursor (handled inline in predict()).
+//  • cooldown: after an accept, briefly suppress re-offering that same unit. The
+//    cooldown lifts only once BOTH the keystroke budget and the time window have
+//    elapsed (either one still counting keeps it active).
+const COOLDOWN_KEYSTROKES = 5;
+const COOLDOWN_MS = 2000;
+
+// Active post-accept cooldown, or null when none.
+let acceptCooldown: { predictionsSince: number; surface: string; ts: number } | null = null;
+
+/**
+ * Start a short cooldown for the exact surface the editor inserted. The caller
+ * passes the committed snapshot's surface so background re-ranking can never
+ * move cooldown bookkeeping away from what the user actually accepted.
+ */
+export const noteSuggestionAccepted = (surface: string): void => {
+	const normalizedSurface = surface.trim().toLowerCase();
+	if (normalizedSurface) {
+		acceptCooldown = {
+			predictionsSince: 0,
+			surface: normalizedSurface,
+			ts: performance.now(),
+		};
+	}
+};
+
 /** Get vector for a word from the store. */
 const getWordVector = (word: string): Float32Array | null => {
 	if (!vectorStore) {
@@ -302,6 +569,118 @@ const extractPreviousWord = (text: string): string => {
 	return words.length >= 2 ? words[words.length - 2] : '';
 };
 
+/**
+ * Collect phrase/bigram completion candidates for the multi-word window ending
+ * at the current partial token.
+ *
+ * Builds windows of the last `2..PHRASE_MAX_WORDS` tokens of the current
+ * sentence (widest first) and prefix-searches the phrase trie for each, so
+ * `quarterly pla` matches `quarterly planning meeting`. Windows never cross a
+ * sentence boundary (mirrors `extractPreviousWord`). Each result records the
+ * matched window length so the caller can compute the ghost tail.
+ *
+ * :params:
+ *   trimmed: Text before the cursor, trailing whitespace already removed
+ * :returns:
+ *   Matched phrase candidates (deduped, widest-window match wins per term)
+ */
+const getPhraseCandidates = (trimmed: string): MatchedCandidate[] => {
+	// Records every prefix window we probe and its match count so the (verbose)
+	// debug line makes it obvious whether the phrase path found nothing because the
+	// trie is empty, or because no phrase begins with the typed window.
+	const windowsTried: Array<{ matches: number; window: string }> = [];
+	const logPhrasePath = (skipReason?: string): void => {
+		if (!isAutocompleteDebugVerbose()) {
+			return;
+		}
+		const detail = skipReason
+			? skipReason
+			: `windows: ${windowsTried.map((w) => `"${w.window}"→${w.matches}`).join(', ') || '(none)'}`;
+		ctcTag(
+			'phrase',
+			`loaded=${phraseTermCount} (${phrasesLoaded ? 'ready' : 'not loaded'}) · ${detail}`,
+			CTC_STYLES.section,
+		);
+	};
+
+	if (phraseTermCount === 0) {
+		logPhrasePath('phrase trie empty — bigrams.json/phrases.json not loaded (or 0 terms)');
+		return [];
+	}
+
+	const tokenMatches = Array.from(trimmed.matchAll(SURFACE_TOKEN_REGEX));
+	const lastMatch = tokenMatches[tokenMatches.length - 1];
+	const lastStart = lastMatch?.index;
+	if (!lastMatch || lastStart === undefined || lastStart + lastMatch[0].length !== trimmed.length) {
+		logPhrasePath('no tokens in current sentence');
+		return [];
+	}
+
+	const trailingTokens: Array<{ start: number; value: string }> = [
+		{ start: lastStart, value: lastMatch[0] },
+	];
+	for (let i = tokenMatches.length - 2; i >= 0; i--) {
+		const current = trailingTokens[0];
+		const match = tokenMatches[i];
+		const start = match.index;
+		if (start === undefined) {
+			break;
+		}
+		const end = start + match[0].length;
+		const gap = trimmed.slice(end, current.start);
+		if (!ONLY_WHITESPACE_REGEX.test(gap)) {
+			break;
+		}
+		trailingTokens.unshift({ start, value: match[0] });
+		if (trailingTokens.length >= PHRASE_MAX_WORDS) {
+			break;
+		}
+	}
+
+	const partial = trailingTokens[trailingTokens.length - 1].value;
+	const results: MatchedCandidate[] = [];
+	const seen = new Set<string>();
+	const maxPrev = trailingTokens.length - 1;
+
+	// Widest window first so the most specific (longest-context) match wins the
+	// dedup for a given phrase term. `prev` is the number of *preceding* complete
+	// words included ahead of the current partial:
+	//   prev ≥ 1 → "continue a phrase I've started" (e.g. "root c" → root cause).
+	//              Preceding-word context is strong, so a short partial is fine.
+	//   prev = 0 → "the partial is a phrase's FIRST word" (e.g. "root" → root
+	//              cause). No preceding context, so require the partial to be a
+	//              one character so grouped LM work can begin early. UI remains
+	//              gated at DISPLAY_MIN_PREFIX_LENGTH.
+	for (let prev = maxPrev; prev >= 0; prev--) {
+		if (prev === 0 && partial.length < MULTIWORD_PREFETCH_MIN_PREFIX_LENGTH) {
+			continue;
+		}
+		const windowTokens = trailingTokens.slice(-(prev + 1));
+		const windowPrefix = windowTokens
+			.map((token) => token.value)
+			.join(' ')
+			.toLowerCase();
+		const windowLen = windowPrefix.length;
+		const matches = phraseTrie.getCandidates(windowPrefix, MAX_PHRASE_CANDIDATES);
+		windowsTried.push({ window: windowPrefix, matches: matches.length });
+		for (const match of matches) {
+			if (seen.has(match.word)) {
+				continue;
+			}
+			seen.add(match.word);
+			results.push({
+				node: match.node,
+				word: match.word,
+				matchedPrefixLen: windowLen,
+				surfaceStart: windowTokens[0].start,
+			});
+		}
+	}
+
+	logPhrasePath();
+	return results;
+};
+
 // ─── Debug Helpers ───────────────────────────────────────────────────────────
 
 /**
@@ -310,16 +689,39 @@ const extractPreviousWord = (text: string): string => {
  * wordCount: number of words in vector store (0 if not loaded)
  */
 export const getPredictorStatus = (): {
+	canonicalScoringSupported: boolean;
+	canonicalSurfaceCount: number | null;
+	clientId: string | null;
+	clientKind: 'localLlm' | 'server' | null;
 	isInitialized: boolean;
+	localModelReady: boolean | null;
+	maxFreqByType: { bigram: number; phrase: number; word: number };
+	phraseCount: number;
+	phrasesLoaded: boolean;
+	slowLaneRegistered: boolean;
 	vectorsLoaded: boolean;
 	vectorsLoadStarted: boolean;
 	wordCount: number;
 } => {
+	const slowLaneStatus = getDefaultSlowLaneClientStatus();
 	return {
+		canonicalScoringSupported: slowLaneStatus.canonicalScoringSupported,
+		canonicalSurfaceCount: slowLaneStatus.canonicalSurfaceCount,
+		clientId: slowLaneStatus.clientId,
+		clientKind: slowLaneStatus.clientKind,
+		localModelReady: slowLaneStatus.localModelReady,
+		slowLaneRegistered: slowLaneStatus.registered,
 		vectorsLoaded: vectorStore !== null,
 		wordCount: vectorStore ? Object.keys(vectorStore.wordIndex).length : 0,
 		vectorsLoadStarted,
 		isInitialized,
+		phrasesLoaded,
+		phraseCount: phraseTermCount,
+		maxFreqByType: {
+			word: wordTrie.maxTenantFreq,
+			bigram: maxBigramFreq,
+			phrase: maxPhraseFreq,
+		},
 	};
 };
 
@@ -328,8 +730,10 @@ export const getPredictorStatus = (): {
  * Returns null if no prediction has run yet or debug was off.
  */
 export const getLastPredictionDebug = (): {
+	awaitingAsyncEvidence: boolean;
 	contextWords: string[];
 	currentWord: string;
+	decision: string;
 	mode: 'cold' | 'warm';
 	suggestion: string | null;
 	textBefore: string;
@@ -349,6 +753,40 @@ export const initVocabulary = (vocabulary: TenantVocabulary): void => {
 		wordTrie.insert(term.word, term.freq, term.docFreq, term.authorFreq);
 	}
 	isInitialized = true;
+	recallGeneration++;
+};
+
+/**
+ * Insert a producer-shipped bigram/phrase artifact into the phrase trie.
+ *
+ * :params:
+ *   artifact: Normalized `{term: {freq, doc_freq, author_freq, pos_seq, head_pos, association}}` map
+ *   termType: Whether these terms are `bigram` (2 tokens) or `phrase` (3 tokens)
+ * :returns:
+ *   The number of terms inserted
+ */
+export const initPhrases = (artifact: PhraseArtifactJson, termType: TermType): number => {
+	let count = 0;
+	for (const [term, stats] of Object.entries(artifact)) {
+		phraseTrie.insert(term, stats.freq, stats.doc_freq, stats.author_freq, {
+			termType,
+			posSeq: stats.pos_seq ?? null,
+			headPos: stats.head_pos ?? null,
+			association: stats.association ?? null,
+		});
+		// Track each type's own frequency ceiling for per-type normalization.
+		if (termType === 'bigram') {
+			if (stats.freq > maxBigramFreq) {
+				maxBigramFreq = stats.freq;
+			}
+		} else if (stats.freq > maxPhraseFreq) {
+			maxPhraseFreq = stats.freq;
+		}
+		count++;
+	}
+	phraseTermCount += count;
+	recallGeneration++;
+	return count;
 };
 
 /**
@@ -385,100 +823,253 @@ export const ingestDocumentPage = (pageContent: string | undefined): void => {
 	}
 
 	if (isAutocompleteDebugEnabled() && validBoostedWords.size > 0) {
-		// eslint-disable-next-line no-console
-		console.groupCollapsed(
-			`%c[L1 Session] %cPrimed ${validBoostedWords.size} valid dictionary words from page`,
-			'color: #00b8d9; font-weight: bold;',
-			'color: inherit; font-style: italic;',
-		);
-		// eslint-disable-next-line no-console
-		console.dir(Array.from(validBoostedWords).sort());
-		// eslint-disable-next-line no-console
-		console.groupEnd();
+		ctcTag('init', `L1 session primed ${validBoostedWords.size} words from page`, CTC_STYLES.brand);
+		if (isAutocompleteDebugVerbose()) {
+			// eslint-disable-next-line no-console
+			console.dir(Array.from(validBoostedWords).sort());
+		}
 	}
 };
 
-export const predict = (textBefore: string): string | null => {
-	if (!isInitialized) {
-		// Vocabulary JSON is code-split and loads asynchronously. Kick off the load
-		// and skip this keystroke; the plugin also primes it on focus, so the tries
-		// are usually ready before the user types.
-		void loadDefaultVocabulary().catch(() => {});
-		return null;
-	}
+/**
+ * Result of a prediction: the ghost tail to insert plus an immutable record of
+ * the evidence that authorized the UI commitment.
+ */
+export interface PredictionResult {
+	/** Canonical evidence depth for the selected full surface. */
+	evidenceDepth: {
+		totalChars: number;
+		totalTokens: number;
+		verifiedChars: number;
+		verifiedTokens: number;
+	};
+	/** Evidence tier that authorized display. Tier A is never display-eligible. */
+	evidenceTier: 'canonical-full-surface' | 'network-logit';
+	/**
+	 * Mean per-token log-probability of the verified prefix.
+	 *
+	 * Carried alongside the posterior because the two answer different
+	 * questions and a plausibility floor is placed on this one. A surface alone
+	 * in its normaliser holds the whole pool whatever the model thinks of it,
+	 * so its posterior cannot say whether it was worth showing and this can.
+	 */
+	meanTokenLogProbability: number;
+	/**
+	 * Whether a longer candidate in the same pool extends the selected surface.
+	 *
+	 * The chain rule leaves such a surface holding its pool's mass undivided, so
+	 * its posterior is high for a structural reason rather than a modelled one —
+	 * whichever continuation was meant, the prefix was right. Carried so the top
+	 * posterior bucket can be split on it, since a lifted nested leader and a
+	 * genuinely preferred surface are otherwise the same number.
+	 */
+	poolHeldExtension: boolean;
+	/** Share of its shortlist's mass the model put on the selected surface. */
+	posterior: number;
+	/** Final Stage-1 + LM ranking score. */
+	rankScore: number;
+	/** How many scored candidates the selected surface's normaliser divided between. */
+	shortlistSize: number;
+	/** Full surface of the selected candidate (identity for keep-vs-swap checks). */
+	surface: string;
+	/** Whether the selected surface is a single word, a bigram, or a phrase. */
+	termType: TermType;
+	/** Ghost tail actually shown/inserted (surface minus the already-typed prefix). */
+	text: string;
+	/** Final-score lead over the strongest evidence-backed runner-up. */
+	winnerMargin: number;
+}
 
-	const t0 = performance.now();
+// ─── Recall memo ─────────────────────────────────────────────────────────────
 
-	// ── Step 1: Bigram-based next-word suggestion at word boundary ───────────
-	// if (textBefore.length > 0 && /\s$/u.test(textBefore)) {
-	// 	const words = textBefore.toLowerCase().trimEnd().split(/\s+/u);
-	// 	const prevWord = words[words.length - 1];
-	// 	const nextWords = bigramMap.get(prevWord);
-	// 	if (nextWords) {
-	// 		const sorted = Object.entries(nextWords).sort((a, b) => b[1] - a[1]);
-	// 		let bestWord = '';
-	// 		for (const [word] of sorted) {
-	// 			if (!isGrammarAllowed(prevWord, word)) {
-	// 				continue;
-	// 			}
-	// 			bestWord = word;
-	// 			break;
-	// 		}
-	// 		if (bestWord) {
-	// 			if (debugMode) {
-	// 				const latencyMs = performance.now() - t0;
-	// 				console.log(
-	// 					'%c[autocomplete] BIGRAM',
-	// 					'color:cyan',
-	// 					'| "' + prevWord + '" -> "' + bestWord + '" | ' + latencyMs.toFixed(1) + 'ms',
-	// 				);
-	// 			}
-	// 			return bestWord;
-	// 		}
-	// 	}
-	// 	if (debugMode) {
-	// 		console.log(
-	// 			'%c[autocomplete] BIGRAM-MISS',
-	// 			'color:gray',
-	// 			'| no bigram for "' + words[words.length - 1] + '", skipping prefix completion',
-	// 		);
-	// 	}
-	// 	return null;
-	// }
-
-	// ── Step 2: Prefix completion (≥3 chars typed) ──────────────────────────
-	if (textBefore.length > 0 && TRAILING_WHITESPACE_REGEX.test(textBefore)) {
-		return null;
-	}
-
-	const trimmed = textBefore.trimEnd();
-	const lastSpaceIdx = trimmed.lastIndexOf(' ');
-	const currentWord = lastSpaceIdx === -1 ? trimmed : trimmed.slice(lastSpaceIdx + 1);
+/** Trie recall plus the canonical context derived for each matched term. */
+interface CanonicalRecall {
+	canonicalMatched: CanonicalMatchedCandidate[];
+	prefixLenByWord: Map<string, number>;
+}
 
-	if (currentWord.length < MIN_PREFIX_LENGTH) {
-		return null;
-	}
+interface RecallMemo extends CanonicalRecall {
+	generation: number;
+	surfaceCount: number;
+	trimmed: string;
+}
 
-	const candidates = wordTrie.getCandidates(currentWord, MAX_CANDIDATES);
+/**
+ * Bumped whenever an artifact load changes what the tries can return, which is
+ * the only way the candidate *set* for a given prefix can change.
+ *
+ * Session-frequency boosts are deliberately not counted: they mutate trie nodes
+ * in place, and the memo holds those nodes by reference, so a boost is already
+ * visible through a cached entry.
+ */
+let recallGeneration = 0;
+
+let recallMemo: RecallMemo | null = null;
+
+const computeCanonicalRecall = (
+	trimmed: string,
+	currentWord: string,
+	currentWordStart: number,
+): CanonicalRecall => {
+	// Start recall at one character so grouped causal work can overlap later
+	// keystrokes. Nothing is displayed until three characters.
+	const wordCandidates: Candidate[] =
+		currentWord.length >= UNIGRAM_PREFETCH_MIN_PREFIX_LENGTH
+			? wordTrie.getCandidates(currentWord, MAX_CANDIDATES)
+			: [];
 
 	// Gap-fill from the L3 general-English trie, requesting a full buffer so
 	// enough survive de-duplication against the L2 results.
-	if (candidates.length < MAX_CANDIDATES) {
+	if (
+		currentWord.length >= UNIGRAM_PREFETCH_MIN_PREFIX_LENGTH &&
+		wordCandidates.length < MAX_CANDIDATES
+	) {
 		const l3Candidates = l3Trie.getCandidates(currentWord, MAX_CANDIDATES);
 
-		const existingWords = new Set(candidates.map((c) => c.word));
+		const existingWords = new Set(wordCandidates.map((c) => c.word));
 
 		for (const l3c of l3Candidates) {
-			if (candidates.length >= MAX_CANDIDATES) break;
+			if (wordCandidates.length >= MAX_CANDIDATES) break;
 
 			if (!existingWords.has(l3c.word)) {
-				candidates.push(l3c);
+				wordCandidates.push(l3c);
 			}
 		}
 	}
 
-	// If both Tries are completely empty for this prefix
-	if (candidates.length === 0) {
+	// Multi-word completion terms (bigrams/phrases) matched on the trailing
+	// window ending at the current partial token. Runs even when the unigram path
+	// was skipped for a short partial, so first-word/early phrase matches surface.
+	const phraseCandidates = getPhraseCandidates(trimmed);
+
+	// Unify: a word completes the current partial token; a phrase completes its
+	// matched multi-word window. Track the prefix length per term so the ghost
+	// tail is sliced correctly regardless of term type.
+	const matched: MatchedCandidate[] = [
+		...wordCandidates.map(({ word, node }) => ({
+			word,
+			node,
+			matchedPrefixLen: currentWord.length,
+			surfaceStart: currentWordStart,
+		})),
+		...phraseCandidates,
+	];
+
+	const positionCache = createCanonicalContextPositionCache();
+	const canonicalMatched: CanonicalMatchedCandidate[] = matched.map((candidate) => ({
+		...candidate,
+		...deriveCanonicalCandidateContext(
+			trimmed,
+			candidate.matchedPrefixLen,
+			candidate.word,
+			getCanonicalSurfaceTokenIds,
+			candidate.surfaceStart,
+			positionCache,
+		),
+	}));
+
+	const prefixLenByWord = new Map<string, number>();
+	for (const m of canonicalMatched) {
+		if (!prefixLenByWord.has(m.word)) {
+			prefixLenByWord.set(m.word, m.matchedPrefixLen);
+		}
+	}
+
+	return { canonicalMatched, prefixLenByWord };
+};
+
+/**
+ * Recall and canonical derivation for `trimmed`, reusing the previous result
+ * when nothing they depend on has changed.
+ *
+ * `predict()` runs several times for one keystroke — once when the decision
+ * opens and again for each async evidence signal that lands inside the budget —
+ * and only the evidence differs between those runs. Without this, every run
+ * re-walks the tries and re-derives a canonical context per candidate.
+ */
+const getCanonicalRecall = (
+	trimmed: string,
+	currentWord: string,
+	currentWordStart: number,
+): CanonicalRecall => {
+	const surfaceCount = getCanonicalSurfaceCount();
+	if (
+		recallMemo !== null &&
+		recallMemo.trimmed === trimmed &&
+		recallMemo.generation === recallGeneration &&
+		recallMemo.surfaceCount === surfaceCount
+	) {
+		return recallMemo;
+	}
+
+	const recall = computeCanonicalRecall(trimmed, currentWord, currentWordStart);
+	recallMemo = { ...recall, generation: recallGeneration, surfaceCount, trimmed };
+	return recall;
+};
+
+export const predict = (textBefore: string): PredictionResult | null => {
+	if (!isInitialized) {
+		// Vocabulary artifacts load asynchronously over the autocomplete gateway.
+		// Kick off the load and skip this keystroke; the plugin also primes it on
+		// focus, so the tries are usually ready before the user types.
+		void loadDefaultVocabulary({ source: 'predict' }).catch(() => {});
+		return null;
+	}
+
+	const t0 = performance.now();
+
+	// Advance the post-accept cooldown once per prediction. It stays active until
+	// BOTH the keystroke budget and the time window have elapsed; while active it
+	// suppresses re-offering the just-accepted surface (see arbitration below).
+	let cooledSurface: string | null = null;
+	if (acceptCooldown) {
+		acceptCooldown.predictionsSince += 1;
+		const stillByKeys = acceptCooldown.predictionsSince <= COOLDOWN_KEYSTROKES;
+		const stillByTime = performance.now() - acceptCooldown.ts < COOLDOWN_MS;
+		if (stillByKeys || stillByTime) {
+			cooledSurface = acceptCooldown.surface;
+		} else {
+			acceptCooldown = null;
+		}
+	}
+
+	// ── Boundary/early-prefix prefetch; display remains gated at ≥3 ─────────
+	if (textBefore.length > 0 && TRAILING_WHITESPACE_REGEX.test(textBefore)) {
+		const boundaryContext = deriveWhitespaceBoundaryContext(textBefore);
+		if (boundaryContext) {
+			primeBoundaryLm({
+				...boundaryContext,
+				familyKey: boundaryContext.contextKey,
+				priority: 0,
+			});
+		}
+		return null;
+	}
+
+	const trimmed = textBefore.trimEnd();
+	const trailingSurfaceToken = trimmed.match(TRAILING_SURFACE_TOKEN_REGEX)?.[0] ?? '';
+	if (trailingSurfaceToken.length === 0) {
+		return null;
+	}
+	const currentWord = trailingSurfaceToken;
+	const currentWordStart = trimmed.length - currentWord.length;
+
+	const { canonicalMatched, prefixLenByWord } = getCanonicalRecall(
+		trimmed,
+		currentWord,
+		currentWordStart,
+	);
+
+	// If every trie was empty for this prefix
+	if (canonicalMatched.length === 0) {
+		if (isAutocompleteDebugEnabled()) {
+			// eslint-disable-next-line no-console
+			console.log(
+				`%c[CTC]%c — abstain: no matches for "${currentWord}"`,
+				CTC_STYLES.brand,
+				CTC_STYLES.body,
+			);
+		}
 		return null;
 	}
 
@@ -487,36 +1078,135 @@ export const predict = (textBefore: string): string | null => {
 	const contextVector = getContextVectorForScoring(trimmed);
 	const lmLogits = getStoredLmLogits();
 
-	// Raw LM Output Logger
-	if (isAutocompleteDebugEnabled() && lmLogits && Object.keys(lmLogits).length > 0) {
-		const rawLmTop = Object.entries(lmLogits)
-			.sort((a, b) => b[1] - a[1])
-			.slice(0, 5)
-			.map(([word, score]) => ({
-				Word: word,
-				Prob: Number(score.toFixed(5)),
-			}));
-		// eslint-disable-next-line no-console
-		console.log('%c[Raw LM Prediction]🧠', 'color: #e83e8c; font-weight: bold;', rawLmTop);
-	}
-
 	const mode: 'cold' | 'warm' = contextVector ? 'warm' : 'cold';
 
-	// Build ScoringCandidate array from TrieNodes
-	const scoringCandidates: ScoringCandidate[] = candidates.map(({ word, node }) => ({
+	// Build ScoringCandidate array from matched terms (words + phrases)
+	const scoringCandidates: ScoringCandidate[] = canonicalMatched.map(({ word, node }) => ({
 		word,
 		tenantFreq: node.tenantFreq,
 		docFreq: node.docFreq,
 		authorFreq: node.authorFreq,
 		sessionFreq: node.sessionFreq,
+		termType: node.termType,
+		posSeq: node.posSeq,
+		headPos: node.headPos,
+		association: node.association,
 	}));
 
 	// Filter the LM payload to only words matching the current prefix so that
 	// lmMax in rankCandidates reflects prefix-relevant signal, not the global distribution.
 	const prefix = currentWord.toLowerCase();
-	const prefixLmLogits = lmLogits
-		? Object.fromEntries(Object.entries(lmLogits).filter(([word]) => word.startsWith(prefix)))
-		: null;
+	const currentWordSeparator = canonicalMatched.find(
+		(candidate) => candidate.node.termType === 'word',
+	)?.separatorKind;
+	const prefixLmLogits =
+		lmLogits && currentWordSeparator === 'whitespace'
+			? Object.fromEntries(Object.entries(lmLogits).filter(([word]) => word.startsWith(prefix)))
+			: null;
+
+	const canonicalScoringSupported = isCanonicalSurfaceScoringSupported();
+	const eligibleContextKeys = Array.from(
+		new Set(
+			canonicalMatched
+				.filter((candidate) => candidate.canonicalTokenIds !== null)
+				.map((candidate) => candidate.contextKey),
+		),
+	).sort();
+	const familyKey = eligibleContextKeys.join('\u0001');
+	const primeRequests = selectBoundaryPrimeRequests(familyKey, canonicalMatched, PHRASE_MAX_WORDS);
+	for (const request of primeRequests) {
+		primeBoundaryLm(request);
+	}
+
+	const runtimeBySurface = new Map(
+		canonicalMatched.map((candidate) => [candidate.word, candidate]),
+	);
+	const canonicalEvidence = new Map<string, CanonicalLmEvidence>();
+	const firstTokenGroups = new Map<string, CanonicalMatchedCandidate[]>();
+	for (const candidate of canonicalMatched) {
+		if (candidate.canonicalTokenIds === null) {
+			continue;
+		}
+		// This accessor also materializes an exact cache entry when all tokens
+		// have already been covered by shared-prefix work.
+		getProgressiveSurfaceEvidence(candidate.contextKey, candidate.word);
+		const exact = getSurfaceScore(candidate.contextKey, candidate.word);
+		if (exact) {
+			canonicalEvidence.set(candidate.word, {
+				contextKey: candidate.contextKey,
+				meanTokenLogProbability: exact.meanTokenLogProbability,
+				score: 0,
+				separatorKind: 'whitespace',
+				source: 'canonical-full-surface',
+				totalLogProbability: exact.totalLogProbability,
+				totalSurfaceCharCount: candidate.word.length,
+				totalTokenCount: exact.tokenCount,
+				verifiedCharCount: candidate.word.length,
+				verifiedTokenCount: exact.tokenCount,
+			});
+			continue;
+		}
+		const group = firstTokenGroups.get(candidate.contextKey) ?? [];
+		group.push(candidate);
+		firstTokenGroups.set(candidate.contextKey, group);
+	}
+
+	for (const [contextKey, group] of firstTokenGroups) {
+		const boundary = getBoundaryLmState(contextKey);
+		if (!boundary) {
+			continue;
+		}
+		let maxLogit = -Infinity;
+		for (const candidate of group) {
+			const tokenId = candidate.canonicalTokenIds?.[0];
+			const rawLogit = tokenId === undefined ? undefined : boundary.rawLogits[tokenId];
+			if (rawLogit !== undefined && Number.isFinite(rawLogit) && rawLogit > maxLogit) {
+				maxLogit = rawLogit;
+			}
+		}
+		if (!Number.isFinite(maxLogit)) {
+			continue;
+		}
+		for (const candidate of group) {
+			const tokenId = candidate.canonicalTokenIds?.[0];
+			const rawLogit = tokenId === undefined ? undefined : boundary.rawLogits[tokenId];
+			if (rawLogit === undefined || !Number.isFinite(rawLogit)) {
+				continue;
+			}
+			const tokenCount = candidate.canonicalTokenIds?.length ?? 0;
+			if (tokenId !== undefined && tokenCount === 1) {
+				const totalLogProbability = logSoftmaxAt(boundary.rawLogits, tokenId);
+				if (!Number.isFinite(totalLogProbability)) {
+					continue;
+				}
+				canonicalEvidence.set(candidate.word, {
+					contextKey,
+					meanTokenLogProbability: totalLogProbability,
+					score: 0,
+					separatorKind: 'whitespace',
+					source: 'canonical-full-surface',
+					totalLogProbability,
+					totalSurfaceCharCount: candidate.word.length,
+					totalTokenCount: 1,
+					verifiedCharCount: candidate.word.length,
+					verifiedTokenCount: 1,
+				});
+				continue;
+			}
+			canonicalEvidence.set(candidate.word, {
+				contextKey,
+				rawLogit,
+				score: 0,
+				separatorKind: 'whitespace',
+				source: 'canonical-first-token',
+				totalSurfaceCharCount: candidate.word.length,
+				totalTokenCount: tokenCount,
+				verifiedCharCount: 0,
+				verifiedTokenCount: 1,
+			});
+		}
+	}
+
 	const {
 		candidates: ranked,
 		grammarMeta,
@@ -526,29 +1216,681 @@ export const predict = (textBefore: string): string | null => {
 		contextVector,
 		(w: string) => getWordVector(w),
 		prefixLmLogits,
-		wordTrie.maxTenantFreq,
+		{
+			word: wordTrie.maxTenantFreq,
+			bigram: maxBigramFreq,
+			phrase: maxPhraseFreq,
+		},
 		previousWord,
+		(surface) => canonicalEvidence.get(surface) ?? null,
+	);
+
+	const progressiveEligible = ranked.filter((candidate) => {
+		const runtime = runtimeBySurface.get(candidate.word);
+		return runtime?.canonicalTokenIds !== null && runtime?.canonicalTokenIds !== undefined;
+	});
+	if (canonicalScoringSupported && progressiveEligible.length > 0) {
+		const byContext = new Map<
+			string,
+			Array<{ rankHint: number; runtime: CanonicalMatchedCandidate }>
+		>();
+		for (const candidate of progressiveEligible) {
+			const runtime = runtimeBySurface.get(candidate.word);
+			if (!runtime || runtime.canonicalTokenIds === null) {
+				continue;
+			}
+			const group = byContext.get(runtime.contextKey) ?? [];
+			group.push({ runtime, rankHint: candidate.stage1Score });
+			byContext.set(runtime.contextKey, group);
+		}
+		for (const [contextKey, group] of byContext) {
+			requestProgressiveSurfaceScores({
+				familyKey,
+				contextKey,
+				prompt: group[0].runtime.contextBeforeSurface,
+				candidates: group.map(({ runtime, rankHint }) => ({
+					surface: runtime.word,
+					tokenIds: runtime.canonicalTokenIds ?? [],
+					rankHint,
+				})),
+			});
+		}
+	}
+
+	// The one/two-character passes exist only to hide model latency. They may
+	// prime and expand token groups above, but cannot commit UI.
+	const displayPrefixReady = currentWord.length >= DISPLAY_MIN_PREFIX_LENGTH;
+
+	// ── Arbitration: full-surface evidence → plausibility → winner margin ──
+	const canonicalLmSupported = canonicalScoringSupported;
+	const hasExactEvidence = (candidate: ScoredCandidate): boolean =>
+		candidate.lmEvidence?.source === 'canonical-full-surface' &&
+		candidate.lmEvidence.meanTokenLogProbability !== undefined &&
+		Number.isFinite(candidate.lmEvidence.meanTokenLogProbability);
+
+	/**
+	 * The evidence this candidate is judged on: the summed log-probability of
+	 * the tokens that have been read, their per-token mean, and whether they
+	 * cover the whole surface.
+	 *
+	 * `total` is the quantity the posterior is built from — it is a sequence
+	 * log-likelihood, so it is comparable across surfaces of different lengths
+	 * in a way the mean is not. `mean` is retained for reporting and for the
+	 * oracle arms, which compare aggregation rules against each other.
+	 *
+	 * Every token read costs a round trip, so a surface longer than
+	 * `REQUIRED_VERIFIED_TOKENS` is judged on its leading tokens and its tail is
+	 * left unread. Callers compensate with a wider winner margin rather than
+	 * treating the two kinds of evidence as equivalent.
+	 *
+	 * This is the depth-unaware read, and it is what the pool below is built
+	 * from. Arbitration goes through `judgedEvidenceFor`, which adds the one
+	 * requirement that can only be stated once the pool is known: a surface
+	 * cannot be judged against members it has not yet been read level with.
+	 */
+	const readJudgedEvidence = (
+		candidate: ScoredCandidate,
+	): { complete: boolean; mean: number; total: number } | null => {
+		const evidence = candidate.lmEvidence;
+		const exactMean = evidence?.meanTokenLogProbability;
+		if (hasExactEvidence(candidate) && evidence && exactMean !== undefined) {
+			const total = evidence.totalLogProbability ?? exactMean * evidence.totalTokenCount;
+			return Number.isFinite(total) ? { complete: true, mean: exactMean, total } : null;
+		}
+		const runtime = runtimeBySurface.get(candidate.word);
+		if (!runtime || runtime.canonicalTokenIds === null) {
+			return null;
+		}
+		const progressive = getProgressiveSurfaceEvidence(runtime.contextKey, candidate.word);
+		if (!progressive || progressive.scoredTokenCount < REQUIRED_VERIFIED_TOKENS) {
+			return null;
+		}
+		const total = progressive.totalLogProbability;
+		const mean = total / progressive.scoredTokenCount;
+		return Number.isFinite(mean) && Number.isFinite(total)
+			? { complete: false, mean, total }
+			: null;
+	};
+
+	// Normalise each context's shortlist into a posterior. `total` is a sequence
+	// log-likelihood, so `exp(total - logSumExp(totals))` is the share of the
+	// shortlist's probability mass the model puts on that surface.
+	//
+	// Unknown future token log-probabilities cannot exceed zero, so a partially
+	// read candidate's total is an upper bound on its eventual total, and its
+	// posterior an upper bound on its eventual posterior. That is what lets the
+	// plausibility floor and the display gate be the same threshold — one
+	// applied to the bound, one to the verified value — instead of two numbers
+	// that have to be kept consistent by hand.
+	//
+	// Candidates with no evidence at all are left out of the sum rather than
+	// defaulted. Any default would have to stand in for a log-probability, and a
+	// value low enough not to distort the normaliser is indistinguishable from
+	// omitting the candidate.
+	const optimisticTotalByWord = new Map<string, number>();
+	const contextTotals = new Map<string, number[]>();
+	const contextSurfaces = new Map<string, string[]>();
+	// Counted before the total is looked up, so this is every candidate the
+	// context could have scored rather than every candidate it did. The gap
+	// between the two is what separates a context holding one match from a
+	// context whose rivals have not come back yet.
+	const contextRequestedCount = new Map<string, number>();
+	const bestTotalByContext = new Map<string, number>();
+	const bestCandidateByContext = new Map<string, ScoredCandidate>();
+	for (const candidate of ranked) {
+		const runtime = runtimeBySurface.get(candidate.word);
+		if (!runtime || runtime.canonicalTokenIds === null) {
+			continue;
+		}
+		contextRequestedCount.set(
+			runtime.contextKey,
+			(contextRequestedCount.get(runtime.contextKey) ?? 0) + 1,
+		);
+		const judged = readJudgedEvidence(candidate);
+		const progressive = getProgressiveSurfaceEvidence(runtime.contextKey, candidate.word);
+		// Once a candidate has been read far enough to judge, its verified total
+		// replaces the bound.
+		const optimisticTotal = judged?.total ?? progressive?.totalLogProbability;
+		if (optimisticTotal === undefined || !Number.isFinite(optimisticTotal)) {
+			continue;
+		}
+		optimisticTotalByWord.set(candidate.word, optimisticTotal);
+		const totals = contextTotals.get(runtime.contextKey);
+		const surfaces = contextSurfaces.get(runtime.contextKey);
+		if (totals && surfaces) {
+			totals.push(optimisticTotal);
+			surfaces.push(candidate.word);
+		} else {
+			contextTotals.set(runtime.contextKey, [optimisticTotal]);
+			contextSurfaces.set(runtime.contextKey, [candidate.word]);
+		}
+		const best = bestTotalByContext.get(runtime.contextKey);
+		if (best === undefined || optimisticTotal > best) {
+			bestTotalByContext.set(runtime.contextKey, optimisticTotal);
+			bestCandidateByContext.set(runtime.contextKey, candidate);
+		}
+	}
+
+	/**
+	 * Whether a shorter member of the same pool is this surface's word-boundary
+	 * prefix — `happy path` where `happy` is also present.
+	 *
+	 * Only a multi-word surface can extend another, and a pool is a set, so this
+	 * costs one lookup per space rather than a comparison against every member.
+	 */
+	const extendsAPoolMember = (surface: string, pool: ReadonlySet<string>): boolean => {
+		for (let space = surface.indexOf(' '); space !== -1; space = surface.indexOf(' ', space + 1)) {
+			if (pool.has(surface.slice(0, space))) {
+				return true;
+			}
+		}
+		return false;
+	};
+
+	// A sequence total only falls as more of the sequence is read, so a partial
+	// sum over `k` tokens sits above the same surface's total over `m > k`. An
+	// extension read less far than the member it continues therefore carries a
+	// total above that member's — and because the chain rule narrows the
+	// denominator to the pool's minimal members, `exp(total - normaliser)` comes
+	// out above 1 for a surface whose own continuation has not been looked at.
+	// It needs a three-token first word to happen, which 2.6% of served bigrams
+	// have, and it inflates both the posterior the gate reads and the
+	// `confidenceScore` every other candidate's margin is measured against.
+	//
+	// A surface's canonical tokenization begins with the tokenization of its own
+	// leading words — the producer encodes `" " + surface` in one call and the
+	// ByteLevel pre-tokenizer splits on word boundaries, so no merge crosses a
+	// space — which is what makes both corrections below exact rather than
+	// approximate.
+	//
+	// The total is capped at what it continues: `P(happy path)` cannot exceed
+	// `P(happy)`. Shortest-first, so a three-word phrase is capped against a
+	// two-word prefix that has already been capped itself.
+	//
+	// And the surface is held unjudgeable until it has been read level with what
+	// it continues, because the cap alone would leave it looking exactly as
+	// certain as its prefix while saying nothing about its own tail.
+	const underReadExtensions = new Set<string>();
+	for (const [contextKey, surfaces] of contextSurfaces) {
+		const pool = new Set(surfaces);
+		for (const surface of [...surfaces].sort((a, b) => a.length - b.length)) {
+			const own = optimisticTotalByWord.get(surface);
+			if (own === undefined) {
+				continue;
+			}
+			let cap = Infinity;
+			let continuedDepth = 0;
+			for (
+				let space = surface.indexOf(' ');
+				space !== -1;
+				space = surface.indexOf(' ', space + 1)
+			) {
+				const prefix = surface.slice(0, space);
+				if (!pool.has(prefix)) {
+					continue;
+				}
+				const prefixTotal = optimisticTotalByWord.get(prefix);
+				if (prefixTotal !== undefined && prefixTotal < cap) {
+					cap = prefixTotal;
+				}
+				const prefixDepth = runtimeBySurface.get(prefix)?.canonicalTokenIds?.length ?? 0;
+				if (prefixDepth > continuedDepth) {
+					continuedDepth = prefixDepth;
+				}
+			}
+			if (own > cap) {
+				optimisticTotalByWord.set(surface, cap);
+			}
+			const scoredDepth = getProgressiveSurfaceEvidence(contextKey, surface)?.scoredTokenCount ?? 0;
+			const surfaceDepth = runtimeBySurface.get(surface)?.canonicalTokenIds?.length ?? 0;
+			if (scoredDepth < continuedDepth && scoredDepth < surfaceDepth) {
+				underReadExtensions.add(surface);
+			}
+		}
+	}
+
+	/**
+	 * The evidence arbitration judges a candidate on.
+	 *
+	 * Wraps the depth-unaware read with the pool-relative requirement: a surface
+	 * that continues another pool member is not judgeable until it has been read
+	 * level with it, so the two totals being compared are taken at comparable
+	 * depths.
+	 */
+	const judgedEvidenceFor = (
+		candidate: ScoredCandidate,
+	): { complete: boolean; mean: number; total: number } | null =>
+		underReadExtensions.has(candidate.word) ? null : readJudgedEvidence(candidate);
+	const hasJudgeableEvidence = (candidate: ScoredCandidate): boolean =>
+		judgedEvidenceFor(candidate) !== null;
+
+	// Normalise over the pool's minimal members only, and score everyone —
+	// extensions included — against that.
+	//
+	// A sequence total is its prefix's total plus a log-probability that cannot
+	// be positive, so summing a surface and its own extension double-counts:
+	// `exp(total)` for `happy` already contains every continuation of `happy`,
+	// `happy path` among them. Dividing one pool's mass between the two capped
+	// the shorter form at half however certain it was, and the longer form below
+	// that — both under gates neither could then clear, for a reason that was
+	// arithmetic rather than evidence.
+	//
+	// Excluding extensions from the denominator is the chain rule: the minimal
+	// members partition the mass, and an extension's share of that same
+	// denominator is `P(prefix) * P(continuation | prefix)`, which is exactly
+	// what a multi-word suggestion should be held to. Extensions stay in
+	// `contextTotals`, so how contested a context is still counts every scored
+	// candidate.
+	const logSumExpByContext = new Map<string, number>();
+	for (const [contextKey, surfaces] of contextSurfaces) {
+		const pool = new Set(surfaces);
+		const minimalTotals: number[] = [];
+		for (const surface of surfaces) {
+			const total = optimisticTotalByWord.get(surface);
+			if (total !== undefined && !extendsAPoolMember(surface, pool)) {
+				minimalTotals.push(total);
+			}
+		}
+		// An extension is strictly longer than what it extends, so the shortest
+		// member of any non-empty pool is always minimal and this is never empty.
+		logSumExpByContext.set(contextKey, logSumExp(minimalTotals));
+	}
+
+	/**
+	 * A candidate's share of its context shortlist's probability mass.
+	 *
+	 * Pass an optimistic (partially read) total for an upper bound on the
+	 * eventual posterior, or a verified total for the posterior itself.
+	 */
+	const posteriorFor = (candidate: ScoredCandidate, total: number | undefined): number => {
+		const runtime = runtimeBySurface.get(candidate.word);
+		const normalizer = runtime ? logSumExpByContext.get(runtime.contextKey) : undefined;
+		if (total === undefined || normalizer === undefined || !Number.isFinite(normalizer)) {
+			return 0;
+		}
+		return Math.exp(total - normalizer);
+	};
+	const optimisticPosterior = (candidate: ScoredCandidate): number =>
+		posteriorFor(candidate, optimisticTotalByWord.get(candidate.word));
+	const judgedPosterior = (candidate: ScoredCandidate): number =>
+		posteriorFor(candidate, judgedEvidenceFor(candidate)?.total);
+
+	// A candidate stays in the race for as long as its optimistic posterior
+	// could still clear the gate it will be held to. The bound only falls as
+	// more tokens are read, so nothing dropped here could have gone on to win.
+	const couldClearGate = (candidate: ScoredCandidate): boolean =>
+		optimisticPosterior(candidate) >= MIN_LM_POSTERIOR[candidate.termType];
+	const passesLmFloor = (candidate: ScoredCandidate): boolean =>
+		judgedEvidenceFor(candidate) !== null && couldClearGate(candidate);
+	const hasRequiredEvidence = (
+		candidate: ScoredCandidate,
+	): candidate is ScoredCandidate & {
+		lmEvidence: CanonicalLmEvidence;
+	} => {
+		if (canonicalLmSupported) {
+			return passesLmFloor(candidate);
+		}
+		return candidate.termType === 'word' && candidate.lmEvidence?.source === 'network-logit';
+	};
+
+	// ── Stabilization guards (QI-2) ──────────────────────────────────────────
+	// Whole-surface repetition: drop a candidate whose full surface already equals
+	// the run of tokens immediately before the cursor — accepting it would
+	// duplicate what was just typed (`end to end` → `end to end to end`).
+	const beforeTokens = tokenize(trimmed);
+	const duplicatesPreceding = (c: ScoredCandidate): boolean => {
+		const surfaceTokens = tokenize(c.word);
+		const n = surfaceTokens.length;
+		if (n === 0 || beforeTokens.length < n) {
+			return false;
+		}
+		for (let i = 0; i < n; i++) {
+			if (beforeTokens[beforeTokens.length - n + i] !== surfaceTokens[i]) {
+				return false;
+			}
+		}
+		return true;
+	};
+	// Post-accept cooldown: don't re-offer the surface the user just accepted while
+	// its cooldown is still active (see cooldown advance near t0).
+	const isCooledDown = (c: ScoredCandidate): boolean =>
+		cooledSurface !== null && c.word.toLowerCase() === cooledSurface;
+
+	const isStabilized = (candidate: ScoredCandidate): boolean =>
+		!duplicatesPreceding(candidate) && !isCooledDown(candidate);
+	// A candidate the user has all but finished typing is not worth a ghost, and
+	// a candidate we would never show is not worth waiting for either.
+	const suggestionLengthFor = (candidate: ScoredCandidate): number =>
+		candidate.word.length - (prefixLenByWord.get(candidate.word) ?? currentWord.length);
+	const isWorthShowing = (candidate: ScoredCandidate): boolean =>
+		suggestionLengthFor(candidate) >= MIN_SUGGESTION_LENGTH;
+	const evidenceBacked = ranked.filter(
+		(candidate) => hasRequiredEvidence(candidate) && isStabilized(candidate),
 	);
 
-	const best = ranked[0];
+	const confidenceScore = (candidate: ScoredCandidate): number => {
+		if (!canonicalLmSupported) {
+			return candidate.finalScore;
+		}
+		const runtime = runtimeBySurface.get(candidate.word);
+		if (runtime?.canonicalTokenIds === null || !optimisticTotalByWord.has(candidate.word)) {
+			return candidate.stage1Score;
+		}
+		return STAGE1_WEIGHT * candidate.stage1Score + STAGE2_WEIGHT * optimisticPosterior(candidate);
+	};
+	const exactConfidenceScore = (candidate: ScoredCandidate): number => {
+		if (!canonicalLmSupported || judgedEvidenceFor(candidate) === null) {
+			return candidate.finalScore;
+		}
+		return STAGE1_WEIGHT * candidate.stage1Score + STAGE2_WEIGHT * judgedPosterior(candidate);
+	};
+
+	/**
+	 * How many candidates in this one's context were scored at all.
+	 *
+	 * Read off pool membership, which counts every candidate that contributed a
+	 * total, rather than off the normaliser, which the chain rule narrows to the
+	 * pool's minimal members.
+	 */
+	const scoredPoolSize = (candidate: ScoredCandidate): number => {
+		const contextKey = runtimeBySurface.get(candidate.word)?.contextKey;
+		return contextKey === undefined ? 0 : (contextTotals.get(contextKey)?.length ?? 0);
+	};
+
+	/**
+	 * How many candidates the context held before any of them were scored.
+	 *
+	 * A pool of one means two opposite things. Either the vocabulary offered a
+	 * single continuation of what was typed, which is the least contested case
+	 * there is; or rivals were offered and have not been read yet, which is the
+	 * least informed. Only the second deserves refusing, and the two are
+	 * distinguishable exactly here.
+	 */
+	const requestedPoolSize = (candidate: ScoredCandidate): number => {
+		const contextKey = runtimeBySurface.get(candidate.word)?.contextKey;
+		return contextKey === undefined ? 0 : (contextRequestedCount.get(contextKey) ?? 0);
+	};
+
+	// A suppressed short completion still counts as competition below, so losing
+	// to one abstains rather than promoting the runner-up in its place.
+	const gateCleared = evidenceBacked
+		.filter(isWorthShowing)
+		.map((candidate) => ({
+			candidate,
+			posterior: judgedPosterior(candidate),
+			score: exactConfidenceScore(candidate),
+		}))
+		// The gate is the model's own confidence in the surface; the blended score
+		// only orders what has already cleared it, so a strong corpus prior can no
+		// longer carry a surface the model is unsure of onto the screen.
+		.filter(({ candidate, posterior }) => posterior >= MIN_LM_POSTERIOR[candidate.termType]);
+	/**
+	 * Whether the model finds the surface plausible on its own terms.
+	 *
+	 * The gate above asks which candidate won its pool; this asks whether
+	 * winning it meant anything. The two come apart exactly where a pool is
+	 * thin — a shortlist of one hands its only member 1.0 by construction — and
+	 * a relative test can never see through that however it is tuned.
+	 *
+	 * Waived without canonical scoring, where there is no judged mean to test
+	 * and the network-logit path would otherwise refuse everything.
+	 */
+	const isPlausibleSurface = (candidate: ScoredCandidate): boolean => {
+		if (!canonicalLmSupported) {
+			return true;
+		}
+		const mean = judgedEvidenceFor(candidate)?.mean;
+		return mean !== undefined && mean >= MIN_MEAN_TOKEN_LOG_PROBABILITY;
+	};
+	// Kept as its own stage rather than folded into the gate so the three
+	// populations stay separable in the abstain cascade: cleared its pool and was
+	// implausible, was plausible and had nothing to clear the pool against, or
+	// cleared both. A candidate refused here stays in `ranked` and so still
+	// counts as competition below — promoting the runner-up in place of an
+	// implausible leader would show something worse, not something better.
+	const plausible = gateCleared.filter(({ candidate }) => isPlausibleSurface(candidate));
+	// Applied after the gate rather than folded into it, so the two populations
+	// stay separable: a surface refused here cleared its threshold and was
+	// refused for having had nothing to clear it against.
+	const eligible = plausible
+		.filter(
+			({ candidate }) =>
+				!canonicalLmSupported ||
+				scoredPoolSize(candidate) >= MIN_SCORED_POOL_SIZE ||
+				requestedPoolSize(candidate) <= 1,
+		)
+		.sort((a, b) => b.score - a.score);
+	const selected = eligible[0]?.candidate;
+	const selectedDisplayScore = eligible[0]?.score ?? 0;
+	const selectedEvidence = selected && hasRequiredEvidence(selected) ? selected.lmEvidence : null;
+	/**
+	 * Whether `longer` is `shorter` continued past a word boundary, in the same
+	 * normaliser.
+	 *
+	 * Shared by the two places that have to agree on what an extension is: the
+	 * chain rule, which stops dividing a pool's mass between a sequence and its
+	 * own prefix, and the winner margin below, which stops treating one as the
+	 * other's rival. They were the same category error and are now the same test.
+	 */
+	const extendsInSamePool = (shorter: ScoredCandidate, longer: ScoredCandidate): boolean => {
+		if (longer.word === shorter.word || !optimisticTotalByWord.has(longer.word)) {
+			return false;
+		}
+		const pool = runtimeBySurface.get(shorter.word)?.contextKey;
+		if (pool === undefined || runtimeBySurface.get(longer.word)?.contextKey !== pool) {
+			return false;
+		}
+		return (
+			longer.word.length > shorter.word.length &&
+			longer.word.startsWith(shorter.word) &&
+			longer.word[shorter.word.length] === ' '
+		);
+	};
+	// A candidate that extends the selection is not competing with it. Whichever
+	// continuation the user meant, the shorter surface is a correct ghost — it is
+	// the prefix of both — so the margin has nothing to arbitrate and charging the
+	// selection for its own extension abstains on a case that could not be wrong.
+	// The chain rule already removed this from the denominator; leaving it in the
+	// margin meant a nested leader cleared its gate and then lost to the candidate
+	// that had just been freed from dividing its mass. It penalised bigrams and
+	// phrases specifically, because a nested pool is where they mostly appear.
+	const runnerUpScore = selected
+		? Math.max(
+				...ranked
+					.filter(
+						(candidate) =>
+							candidate.word !== selected.word &&
+							isStabilized(candidate) &&
+							!extendsInSamePool(selected, candidate),
+					)
+					.map(confidenceScore),
+				0,
+			)
+		: 0;
+	const winnerMargin = selected ? selectedDisplayScore - runnerUpScore : 0;
+	const hasMissingArtifactCompetitor =
+		canonicalLmSupported &&
+		ranked.some((candidate) => {
+			const runtime = runtimeBySurface.get(candidate.word);
+			return runtime?.separatorKind === 'whitespace' && runtime.canonicalTokenIds === null;
+		});
+	// A candidate with canonical tokens but no evidence at all cannot join the
+	// posterior: it has no likelihood to contribute, and any stand-in value would
+	// distort the normaliser for every other candidate in its context. Absent
+	// evidence is not evidence of absence though — the context's boundary may
+	// simply not be primed yet — so it holds the ghost back instead of being
+	// silently ignored. In a primed context every candidate has at least
+	// first-token evidence, so this only fires while a context is still cold.
+	const hasUnscoredCompetitor =
+		canonicalLmSupported &&
+		ranked.some((candidate) => {
+			const runtime = runtimeBySurface.get(candidate.word);
+			return (
+				candidate.word !== selected?.word &&
+				runtime?.canonicalTokenIds !== null &&
+				runtime?.canonicalTokenIds !== undefined &&
+				!optimisticTotalByWord.has(candidate.word) &&
+				isStabilized(candidate) &&
+				isWorthShowing(candidate)
+			);
+		});
+	// A rival only holds the ghost back while it is still too unread to judge.
+	// Under full-surface evidence that meant any long surface blocked everything
+	// until every one of its tokens had been paid for, which abstained far more
+	// often than it ever changed the winner. A rival now becomes judgeable after
+	// `REQUIRED_VERIFIED_TOKENS`, so this settles within the budget.
+	const hasUnresolvedPotential =
+		canonicalLmSupported &&
+		ranked.some((candidate) => {
+			const runtime = runtimeBySurface.get(candidate.word);
+			return (
+				candidate.word !== selected?.word &&
+				runtime?.canonicalTokenIds !== null &&
+				runtime?.canonicalTokenIds !== undefined &&
+				!hasJudgeableEvidence(candidate) &&
+				couldClearGate(candidate) &&
+				isStabilized(candidate) &&
+				isWorthShowing(candidate)
+			);
+		});
+	const requiredWinnerMargin =
+		MIN_WINNER_MARGIN +
+		(selected && judgedEvidenceFor(selected)?.complete === false
+			? PARTIAL_EVIDENCE_MARGIN_PREMIUM
+			: 0);
+	// Everything the display decision needs except the veto itself. Splitting it
+	// out is what lets the veto be priced: the value here at the moment the veto
+	// fires is exactly the ghost the veto cost us.
+	const clearsMarginBeforeColdVeto =
+		selected !== undefined && !hasMissingArtifactCompetitor && winnerMargin >= requiredWinnerMargin;
+	const clearsWinnerMargin = clearsMarginBeforeColdVeto && !hasUnscoredCompetitor;
+
 	const suggestion =
-		best && best.finalScore >= MIN_SCORE_THRESHOLD ? best.word.slice(currentWord.length) : null;
+		displayPrefixReady && selected && clearsWinnerMargin
+			? selected.word.slice(prefixLenByWord.get(selected.word) ?? currentWord.length)
+			: null;
+
+	// Which constraint bound. Computed on every evaluation rather than only under
+	// debug, because it is counted for the session as well as printed, and
+	// deriving the debug header from it keeps the two from disagreeing.
+	const resolveAbstainReason = (): CtcAbstainReason | null => {
+		if (suggestion !== null && suggestion.length > 0) {
+			return null;
+		}
+		if (!displayPrefixReady) {
+			return 'prefetch';
+		}
+		if (ranked.length === 0) {
+			return 'no-candidate';
+		}
+		if (hasMissingArtifactCompetitor) {
+			return 'missing-artifact';
+		}
+		if (hasUnscoredCompetitor) {
+			return 'cold-competitor';
+		}
+		if (hasUnresolvedPotential && !clearsWinnerMargin) {
+			return 'unresolved-rival';
+		}
+		if (eligible.length > 0) {
+			return selected && !clearsWinnerMargin ? 'winner-margin' : 'empty-completion';
+		}
+		if (plausible.length > 0) {
+			// Cleared its gate, was plausible on its own terms, and was refused for
+			// the shortlist it cleared that gate against holding nobody else.
+			return 'lone-candidate';
+		}
+		if (gateCleared.length > 0) {
+			// Held the largest share of its shortlist and still was not a surface
+			// the model found plausible — which is only possible because a share is
+			// relative and this is not.
+			return 'implausible-surface';
+		}
+		if (
+			evidenceBacked.length > 0 &&
+			evidenceBacked.every((candidate) => !isWorthShowing(candidate))
+		) {
+			return 'short-completion';
+		}
+		// The floor and the gate are one threshold, applied to the bound and to
+		// the verified value, so a candidate short of it is dropped before
+		// `evidenceBacked` is built. Asking whether anything was judgeable
+		// separates a gate that rejected read candidates from a context where
+		// nothing was read at all — the first is a threshold to calibrate, the
+		// second is scheduling.
+		return ranked.some(hasJudgeableEvidence) ? 'below-posterior-gate' : 'no-evidence';
+	};
+	const abstainReason: CtcAbstainReason | null = resolveAbstainReason();
+
+	// The leader is the best-supported candidate, not the selected one: an
+	// evaluation that showed nothing is exactly the one whose posterior needs
+	// recording, and it has no selection to report.
+	const posteriorLeader = ranked.reduce<{
+		candidate: ScoredCandidate;
+		posterior: number;
+	} | null>((best, candidate) => {
+		const posterior = optimisticPosterior(candidate);
+		return posterior > 0 && (best === null || posterior > best.posterior)
+			? { candidate, posterior }
+			: best;
+	}, null);
+	/**
+	 * The longer surface in the leader's normaliser that extends the given
+	 * candidate, if the pool holds one.
+	 *
+	 * `logSumExp` divides one pool's mass between a sequence and its own prefix,
+	 * and the longer total is the shorter one plus a log-probability that cannot
+	 * be positive — so the longer form holds at most half, or at most a third
+	 * where both shorter forms are present. Both sit under the gate it is held
+	 * to, so it cannot display however certain its continuation is.
+	 *
+	 * Asked from the shorter form because that is the one that leads: the
+	 * extending candidate is strictly lower in the same pool and so can never be
+	 * the leader.
+	 */
+	const poolExtensionOf = (shorter: ScoredCandidate): ScoredCandidate | undefined =>
+		ranked.find((candidate) => extendsInSamePool(shorter, candidate));
 
 	if (isAutocompleteDebugEnabled()) {
+		const verbose = isAutocompleteDebugVerbose();
 		const latencyMs = performance.now() - t0;
 		const tokens = tokenize(trimmed);
 		const contextWords = tokens.slice(-CONTEXT_WORDS);
-		const belowThreshold = best && best.finalScore < MIN_SCORE_THRESHOLD;
 
-		const suggestionLabel = belowThreshold
-			? '🚫 (below threshold)'
-			: suggestion && suggestion.length > 0
-				? `✨ "${suggestion}"`
-				: '🚫 (no match)';
+		// Decision drives the (collapsed) group header. Abstains carry the reason
+		// the session counters recorded, plus the numbers behind it, so it is clear
+		// *why* nothing showed without expanding the group.
+		const abstainDetail: Record<CtcAbstainReason, string> = {
+			'below-posterior-gate': `posterior below ${
+				MIN_LM_POSTERIOR[(posteriorLeader?.candidate ?? ranked[0]).termType]
+			} (best ${(posteriorLeader?.posterior ?? 0).toFixed(2)})`,
+			'cold-competitor': 'competitor has no LM evidence yet',
+			'empty-completion': 'empty completion',
+			'implausible-surface': `mean per-token log-probability below ${MIN_MEAN_TOKEN_LOG_PROBABILITY} (best ${gateCleared
+				.map(({ candidate }) => judgedEvidenceFor(candidate)?.mean ?? -Infinity)
+				.reduce((best, mean) => Math.max(best, mean), -Infinity)
+				.toFixed(2)})`,
+			'lone-candidate': `cleared its gate against fewer than ${MIN_SCORED_POOL_SIZE} scored candidates, rivals unread`,
+			'missing-artifact': 'canonical artifact coverage missing',
+			'no-candidate': 'nothing cleared scoring',
+			'no-evidence': 'full-surface evidence absent',
+			prefetch: `prefetch: ${currentWord.length}/${DISPLAY_MIN_PREFIX_LENGTH} chars`,
+			'short-completion': `completion shorter than ${MIN_SUGGESTION_LENGTH} chars`,
+			'unresolved-rival': 'expanding plausible token-prefix groups',
+			'winner-margin': `winner margin ${winnerMargin.toFixed(2)} < ${requiredWinnerMargin.toFixed(2)}`,
+		};
+		const decision =
+			abstainReason === null
+				? `✨ "${suggestion ?? ''}"`
+				: abstainReason === 'prefetch'
+					? `⏳ ${abstainDetail.prefetch}`
+					: `— abstain: ${abstainDetail[abstainReason]}`;
 
 		lastPredictionDebug = {
+			awaitingAsyncEvidence: hasUnresolvedPotential && !clearsWinnerMargin,
 			textBefore: trimmed,
 			currentWord,
+			decision,
 			mode,
 			contextWords,
 			topCandidates: ranked.slice(0, 5).map((r) => ({
@@ -561,98 +1903,239 @@ export const predict = (textBefore: string): string | null => {
 			suggestion: suggestion && suggestion.length > 0 ? suggestion : null,
 		};
 
-		// ── Mode label: COLD / WARM(local) / WARM(BE) ───────────────────────
+		// Mode describes the semantic-vector source without implying that a
+		// network backend ran; the snapshot may come from the local MLC embedder.
 		const slowLaneVec = getStoredContextVector();
 		const isUsingSlowLaneVector =
 			slowLaneVec !== null && vectorStore !== null && slowLaneVec.length === vectorStore.dim;
-		const modeLabel = !contextVector ? 'COLD' : isUsingSlowLaneVector ? 'WARM(BE)' : 'WARM(local)';
+		const modeLabel = !contextVector
+			? 'COLD'
+			: isUsingSlowLaneVector
+				? 'WARM(semantic)'
+				: 'WARM(local-avg)';
 		const modeColor =
-			modeLabel === 'WARM(BE)'
-				? 'color: #ff9800; font-weight: bold;'
-				: modeLabel === 'WARM(local)'
-					? 'color: #4caf50; font-weight: bold;'
-					: 'color: #9e9e9e; font-weight: bold;';
-
-		// 1. Collapsible group header
+			modeLabel === 'WARM(semantic)'
+				? CTC_STYLES.good
+				: modeLabel === 'WARM(local-avg)'
+					? CTC_STYLES.warn
+					: CTC_STYLES.cold;
+
+		// Group header: [CTC:<mode>] <decision> · <mode> · <type> · <latency>
+		// arbitrationMode ('v1') tags the line so this deploy's logs diff cleanly
+		// against the baseline deploy's.
 		// eslint-disable-next-line no-console
 		console.groupCollapsed(
-			`%c[Autocomplete] %c${modeLabel} %c| "${currentWord}" ➔ ${suggestionLabel} | ⏱ ${latencyMs.toFixed(1)}ms`,
-			'color: #00b8d9; font-weight: bold;',
+			`%c[CTC:${ARBITRATION_MODE}]%c ${decision} %c· ${modeLabel}${selected ? ` · ${selected.termType}` : ''} · ⏱ ${latencyMs.toFixed(1)}ms`,
+			CTC_STYLES.brand,
+			CTC_STYLES.body,
 			modeColor,
-			'color: inherit; font-weight: normal;',
 		);
 
-		// 2. Context window (what local vector averaging sees)
-		// eslint-disable-next-line no-console
-		console.log(
-			'%cContext Window:',
-			'color: #888; font-style: italic;',
-			contextWords.length ? contextWords.join(' ') : '(none)',
+		const semanticInput = getStoredContextInput();
+		const causalPrompts = primeRequests.map((request, index) => {
+			const tail = request.prompt.slice(-DEBUG_TEXT_TAIL_CHARS);
+			return `#${index + 1}:${JSON.stringify(tail)}`;
+		});
+
+		// Keep the exact model inputs separate from normalized lookup tokens.
+		ctcSection(
+			'INPUT',
+			`raw: ${JSON.stringify(trimmed.slice(-DEBUG_TEXT_TAIL_CHARS))} · prev: ${previousWord ? JSON.stringify(previousWord) : '—'}`,
+		);
+		ctcSection(
+			'LOOKUP',
+			`normalized(last ${CONTEXT_WORDS}): ${contextWords.length ? JSON.stringify(contextWords.join(' ')) : '(none)'}`,
+		);
+		ctcSection(
+			'SEMANTIC',
+			`snapshot: ${semanticInput ? JSON.stringify(semanticInput.slice(-DEBUG_TEXT_TAIL_CHARS)) : '(unavailable)'}`,
+		);
+		ctcSection(
+			'CAUSAL',
+			`prompt${causalPrompts.length === 1 ? '' : 's'}: ${causalPrompts.join(' · ') || '(not ready)'}`,
 		);
 
-		// 3. Slow-lane status
-		const vectorStatus = isUsingSlowLaneVector
-			? `✅ BE semantic vector (dim=${slowLaneVec?.length})`
+		// SIGNALS — semantic state plus context-keyed canonical LM coverage.
+		const logitCount = lmLogits ? Object.keys(lmLogits).length : 0;
+		const semanticStatus = isUsingSlowLaneVector
+			? `semantic ✅ snapshot(${slowLaneVec?.length}d)`
 			: vectorStore
-				? '⚠️ Local vector average (slow-lane not yet returned)'
-				: '❌ No vectors (cold)';
+				? 'semantic ⚠️ local-avg'
+				: 'semantic ❌ cold';
+		const tierACount = ranked.filter(
+			(candidate) => candidate.lmEvidence?.source === 'canonical-first-token',
+		).length;
+		const exactCount = ranked.filter(
+			(candidate) => candidate.lmEvidence?.source === 'canonical-full-surface',
+		).length;
+		const boundaryReady = primeRequests.filter(
+			(request) => getBoundaryLmState(request.contextKey) !== null,
+		).length;
+		const artifactEligible = canonicalMatched.filter(
+			(candidate) => candidate.canonicalTokenIds !== null,
+		).length;
+		const documentStartFallbacks = canonicalMatched.filter(
+			(candidate) => candidate.separatorKind === 'document-start',
+		).length;
+		const punctuationFallbacks = canonicalMatched.filter(
+			(candidate) => candidate.separatorKind === 'non-space',
+		).length;
 		const logitsStatus =
-			lmLogits && Object.keys(lmLogits).length > 0
-				? `✅ LM logits active (${Object.keys(lmLogits).length} tokens)`
-				: '⏳ No LM logits (slow-lane pending or failed)';
-		// eslint-disable-next-line no-console
-		console.log(
-			'%cSlow Lane:',
-			'color: #888; font-style: italic;',
-			vectorStatus,
-			'|',
-			logitsStatus,
+			tierACount + exactCount > 0
+				? `canonical ✅ first-token:${tierACount} exact:${exactCount}`
+				: logitCount > 0
+					? `network logits ✅ ${logitCount}`
+					: 'canonical ⏳/absent';
+		const formula =
+			tierACount + exactCount + logitCount > 0
+				? `Stage1×${STAGE1_WEIGHT}+LM×${STAGE2_WEIGHT}`
+				: 'Stage1-only';
+		ctcSection('SIGNALS', `${semanticStatus} · ${logitsStatus} · ${formula}`);
+		ctcSection(
+			'CANONICAL',
+			`eligible ${artifactEligible}/${canonicalMatched.length} · separator fallback doc:${documentStartFallbacks} punct:${punctuationFallbacks} · contexts ${boundaryReady}/${primeRequests.length} ready (≤${PHRASE_MAX_WORDS}) · family=${familyKey.slice(0, 72) || 'none'}`,
 		);
+		if (verbose && logitCount > 0 && lmLogits) {
+			const rawLmTop = Object.entries(lmLogits)
+				.sort((a, b) => b[1] - a[1])
+				.slice(0, 5)
+				.map(([word, score]) => `${word}:${score.toFixed(3)}`)
+				.join(', ');
+			ctcSection('  rawLM', `🧠 ${rawLmTop}`);
+		}
 
-		// 4. Scoring formula active this prediction
-		const formulaLabel =
-			lmLogits && Object.keys(lmLogits).length > 0
-				? `Stage1(×${STAGE1_WEIGHT}) + LM(×${STAGE2_WEIGHT})`
-				: 'Stage1 only (no LM logits)';
-		// eslint-disable-next-line no-console
-		console.log('%cFormula:', 'color: #888; font-style: italic;', formulaLabel);
+		// GENERATE — how many candidates matched, split by term type
+		const genByType: Record<TermType, number> = { word: 0, bigram: 0, phrase: 0 };
+		for (const m of canonicalMatched) {
+			genByType[m.node.termType] += 1;
+		}
+		ctcSection(
+			'GENERATE',
+			`matched ${canonicalMatched.length} → word:${genByType.word} bigram:${genByType.bigram} phrase:${genByType.phrase}${canonicalLmSupported ? ' · display needs exact surface evidence' : ''}`,
+		);
 
-		// 5. Grammar filter result (collected inside rankCandidates, logged here)
+		// SCORE — Stage-1 + grammar funnel and canonical evidence depth.
+		const lmCoverage = ranked.slice(0, 10).filter((r) => r.lmScore > 0).length;
+		ctcSection(
+			'SCORE',
+			`in ${pipelineDebug.initial} → stage1(<${MIN_STAGE1_SCORE}) −${pipelineDebug.stage1Rejected.length} → grammar −${pipelineDebug.grammarRejected.length} → final ${pipelineDebug.final} · LM cov ${lmCoverage}/${Math.min(ranked.length, 10)} · first-token ${tierACount} · exact ${exactCount}`,
+		);
 		if (grammarMeta) {
-			// eslint-disable-next-line no-console
-			console.log(
-				`%c[Grammar] "${grammarMeta.prevWord}" [${grammarMeta.prevTags.join('|')}] → ${grammarMeta.before} candidates → ${grammarMeta.after} after filter`,
-				'color: #4caf50; font-weight: bold;',
+			ctcSection(
+				'  grammar',
+				`"${grammarMeta.prevWord}" [${grammarMeta.prevTags.join('|')}] ${grammarMeta.before}→${grammarMeta.after}${verbose && grammarMeta.dropped.length > 0 ? ` · dropped: ${grammarMeta.dropped.join(', ')}` : ''}`,
 			);
-			if (grammarMeta.dropped.length > 0) {
-				// eslint-disable-next-line no-console
-				console.log(
-					`%c🚫 Dropped: ${grammarMeta.dropped.join(', ')}`,
-					'color: #f44336; font-style: italic;',
-				);
-			}
 		}
 
-		// 6. Pipeline funnel
-		// eslint-disable-next-line no-console
-		console.log(
-			`%c[Pipeline Funnel] %c📥 In: ${pipelineDebug.initial} | ❌ Stage 1 (< ${MIN_STAGE1_SCORE}): -${pipelineDebug.stage1Rejected.length} | ❌ Grammar: -${pipelineDebug.grammarRejected.length} | ✅ Final: ${pipelineDebug.final}`,
-			'color: #9c27b0; font-weight: bold;',
-			'color: inherit;',
+		const gateReasonFor = (candidate: ScoredCandidate): string => {
+			const runtime = runtimeBySurface.get(candidate.word);
+			if (runtime?.separatorKind !== 'whitespace') {
+				return `separator:${runtime?.separatorKind ?? 'unknown'}`;
+			}
+			if (runtime.canonicalTokenIds === null) {
+				return 'missing-artifact';
+			}
+			if (!hasJudgeableEvidence(candidate)) {
+				return `awaiting-tokens:${REQUIRED_VERIFIED_TOKENS}`;
+			}
+			if (!passesLmFloor(candidate)) {
+				return `bound:${optimisticPosterior(candidate).toFixed(2)}<${MIN_LM_POSTERIOR[candidate.termType]}`;
+			}
+			if (duplicatesPreceding(candidate)) {
+				return 'repetition';
+			}
+			if (isCooledDown(candidate)) {
+				return 'cooldown';
+			}
+			if (!isWorthShowing(candidate)) {
+				return `suffix:${suggestionLengthFor(candidate)}<${MIN_SUGGESTION_LENGTH}`;
+			}
+			const floor = MIN_LM_POSTERIOR[candidate.termType];
+			const posterior = judgedPosterior(candidate);
+			if (posterior < floor) {
+				return `posterior:${posterior.toFixed(2)}<${floor}`;
+			}
+			if (!isPlausibleSurface(candidate)) {
+				return `implausible:${(judgedEvidenceFor(candidate)?.mean ?? NaN).toFixed(2)}<${MIN_MEAN_TOKEN_LOG_PROBABILITY}`;
+			}
+			return 'eligible';
+		};
+		const gateFunnel = (termType: TermType): string => {
+			const typeRanked = ranked.filter((candidate) => candidate.termType === termType);
+			const exact = typeRanked.filter(hasJudgeableEvidence);
+			const absolute = exact.filter(passesLmFloor);
+			const stabilized = absolute.filter(isStabilized);
+			const longEnough = stabilized.filter(isWorthShowing);
+			const floorPassed = longEnough.filter(
+				(candidate) => judgedPosterior(candidate) >= MIN_LM_POSTERIOR[termType],
+			);
+			const plausiblePassed = floorPassed.filter(isPlausibleSurface);
+			return `${termType} m:${genByType[termType]} r:${typeRanked.length} exact:${exact.length} abs:${absolute.length} stable:${stabilized.length} suffix:${longEnough.length} floor:${floorPassed.length} plausible:${plausiblePassed.length} eligible:${eligible.filter(({ candidate }) => candidate.termType === termType).length}`;
+		};
+		ctcSection(
+			'GATES',
+			`${gateFunnel('word')}  |  ${gateFunnel('bigram')}  |  ${gateFunnel('phrase')}`,
 		);
 
-		// 7. Candidate table
-		if (ranked.length > 0) {
-			const lmCoverage = ranked.slice(0, 10).filter((r) => r.lmScore > 0).length;
-			// eslint-disable-next-line no-console
-			console.log(
-				`%cLM coverage: ${lmCoverage}/${Math.min(ranked.length, 10)} candidates had real logit scores`,
-				'color: #888; font-style: italic;',
-			);
+		// One line per context: how many candidates share the normaliser, and how
+		// much of the mass the leader holds. A leader well under its threshold
+		// means the context is contested, which is the abstention we want.
+		const contextLeaders = Array.from(bestCandidateByContext.entries())
+			.slice(0, PHRASE_MAX_WORDS)
+			.map(([contextKey, leader]) => {
+				const shortlistSize = contextTotals.get(contextKey)?.length ?? 0;
+				const evidenceKind = hasExactEvidence(leader) ? 'exact' : 'upper';
+				return `${contextKey.slice(0, 32)} → n=${shortlistSize} ${leader.termType}:"${leader.word}" ${evidenceKind} p=${optimisticPosterior(leader).toFixed(3)}/${MIN_LM_POSTERIOR[leader.termType]}`;
+			})
+			.join(' | ');
+		ctcSection(
+			'NORMALIZE',
+			contextLeaders.length > 0 ? contextLeaders : 'no scored candidate in any context',
+		);
+		ctcSection(
+			'READINESS',
+			`prefix ${currentWord.length}/${DISPLAY_MIN_PREFIX_LENGTH} · boundary ${boundaryReady}/${primeRequests.length} · exact word:${ranked.filter((candidate) => candidate.termType === 'word' && hasExactEvidence(candidate)).length} bigram:${ranked.filter((candidate) => candidate.termType === 'bigram' && hasExactEvidence(candidate)).length} phrase:${ranked.filter((candidate) => candidate.termType === 'phrase' && hasExactEvidence(candidate)).length} · ${suggestion ? 'display-ready' : decision}`,
+		);
+
+		// ARBITRATE — expose every condition that can authorize or suppress ghost.
+		const topOfType = (t: TermType): ScoredCandidate | undefined =>
+			ranked.find((c) => c.termType === t);
+		const fmtType = (t: TermType): string => {
+			const top = topOfType(t);
+			if (!top) {
+				return `${t}:—`;
+			}
+			const floor = MIN_LM_POSTERIOR[t];
+			const bound = optimisticPosterior(top);
+			const evidence = hasRequiredEvidence(top)
+				? 'TierB✅'
+				: `${top.lmEvidence?.source ?? 'absent'}❌`;
+			return `${t}:"${top.word}" p≤${bound.toFixed(3)}${bound >= floor ? '≥' : '<'}${floor} fs≤${confidenceScore(top).toFixed(2)} ${evidence}`;
+		};
+		ctcSection(
+			'ARBITRATE',
+			`${selected ? `▸ ${selected.termType} "${selected.word}" fs=${selectedDisplayScore.toFixed(2)} mean=${selectedEvidence?.meanTokenLogProbability?.toFixed(2) ?? 'network'} margin=${winnerMargin.toFixed(2)}${clearsWinnerMargin ? '✅' : '❌'}` : '▸ (none)'} · prefix ${currentWord.length}/${DISPLAY_MIN_PREFIX_LENGTH}${displayPrefixReady ? '✅' : '⏳'} · grouped pending ${hasUnresolvedPotential ? 'yes' : 'no'}  |  ${fmtType('word')}  ${fmtType('bigram')}  ${fmtType('phrase')}`,
+		);
+
+		// STABILIZE — post-accept cooldown + whole-surface repetition (QI-2)
+		const repetitionBlocked = ranked
+			.filter((c) => hasRequiredEvidence(c) && duplicatesPreceding(c))
+			.map((c) => c.word);
+		const cooldownLabel = cooledSurface
+			? `cooldown "${cooledSurface}" (${acceptCooldown?.predictionsSince ?? 0}/${COOLDOWN_KEYSTROKES} keys · ${COOLDOWN_MS}ms)`
+			: 'cooldown —';
+		ctcSection(
+			'STABILIZE',
+			`${cooldownLabel} · repetition-blocked: ${repetitionBlocked.length > 0 ? repetitionBlocked.join(', ') : '—'}`,
+		);
 
+		// CANDIDATES — full scored table (verbose only)
+		if (verbose && ranked.length > 0) {
 			const tableData = ranked.slice(0, 10).map((r) => {
 				let rawLogit: string | number = 'Not in Payload';
-				if (prefixLmLogits) {
+				if (r.lmEvidence?.rawLogit !== undefined) {
+					rawLogit = Number(r.lmEvidence.rawLogit.toFixed(5));
+				} else if (prefixLmLogits) {
 					const val = prefixLmLogits[r.word.toLowerCase()];
 					if (val !== undefined) {
 						rawLogit = Number(val.toFixed(5));
@@ -660,9 +2143,19 @@ export const predict = (textBefore: string): string | null => {
 				}
 
 				const original = scoringCandidates.find((sc) => sc.word === r.word);
+				const runtime = runtimeBySurface.get(r.word);
+				const progress = runtime ? getProgressiveSurfaceEvidence(runtime.contextKey, r.word) : null;
+				const optimisticTotal = optimisticTotalByWord.get(r.word);
+				const normalizer = runtime ? logSumExpByContext.get(runtime.contextKey) : undefined;
+				const shortlistSize = runtime ? (contextTotals.get(runtime.contextKey)?.length ?? 0) : 0;
+				const contextLeader = runtime ? bestCandidateByContext.get(runtime.contextKey) : undefined;
 				let source = 'Unknown';
 				if (original) {
-					if (original.docFreq === 0 && original.tenantFreq === L3_BASELINE_FREQ) {
+					if (original.termType === 'bigram') {
+						source = '🔗 L2 (Bigram)';
+					} else if (original.termType === 'phrase') {
+						source = '🧩 L2 (Phrase)';
+					} else if (original.docFreq === 0 && original.tenantFreq === L3_BASELINE_FREQ) {
 						source = '🌍 L3 (Generic)';
 					} else if (original.sessionFreq > 0 && original.tenantFreq === 0) {
 						source = '👤 L1 (Session Only)';
@@ -671,13 +2164,70 @@ export const predict = (textBefore: string): string | null => {
 					}
 				}
 
+				const floor = MIN_LM_POSTERIOR[r.termType];
 				return {
 					Candidate: r.word,
 					Source: source,
+					Type: r.termType,
 					'Final Score': Number(r.finalScore.toFixed(4)),
+					'Bounded Score': Number(confidenceScore(r).toFixed(4)),
+					'Exact Gate Score': hasExactEvidence(r)
+						? Number(exactConfidenceScore(r).toFixed(4))
+						: '—',
+					Floor: floor,
+					'Posterior Floor': optimisticPosterior(r) >= floor ? '✅' : '❌',
+					'LM Floor': hasExactEvidence(r) ? (passesLmFloor(r) ? '✅' : '❌') : '⏳',
+					'Gate Reason': gateReasonFor(r),
+					'Display Eligible':
+						hasRequiredEvidence(r) && isStabilized(r) && judgedPosterior(r) >= floor ? '✅' : '❌',
 					Semantics: Number(r.semanticScore.toFixed(4)),
 					Freq: Number(r.freqScore.toFixed(4)),
 					'LM Score': Number(r.lmScore.toFixed(4)),
+					'LM Evidence': r.lmEvidence?.source ?? 'absent',
+					'Verified Tokens': r.lmEvidence
+						? `${r.lmEvidence.verifiedTokenCount}/${r.lmEvidence.totalTokenCount}`
+						: '0/0',
+					'Progressive Tokens': progress
+						? `${progress.scoredTokenCount}/${progress.totalTokenCount}`
+						: '0/0',
+					'LM Mean Upper':
+						progress !== null ? Number(progress.meanTokenLogProbabilityUpperBound.toFixed(5)) : '—',
+					'Optimistic Total':
+						optimisticTotal !== undefined ? Number(optimisticTotal.toFixed(5)) : '—',
+					'Context logSumExp':
+						normalizer !== undefined && Number.isFinite(normalizer)
+							? Number(normalizer.toFixed(5))
+							: '—',
+					'Shortlist Size': shortlistSize,
+					'Posterior (bound)': Number(optimisticPosterior(r).toFixed(5)),
+					'Posterior (judged)': hasJudgeableEvidence(r)
+						? Number(judgedPosterior(r).toFixed(5))
+						: '—',
+					'Context Leader': contextLeader ? `${contextLeader.termType}:${contextLeader.word}` : '—',
+					'Verified Chars': r.lmEvidence
+						? `${r.lmEvidence.verifiedCharCount}/${r.lmEvidence.totalSurfaceCharCount}`
+						: `0/${r.word.length}`,
+					Separator: runtime?.separatorKind ?? 'unknown',
+					'Canonical Eligibility':
+						runtime?.separatorKind !== 'whitespace'
+							? 'incompatible-separator'
+							: runtime.canonicalTokenIds === null
+								? 'missing-artifact'
+								: 'eligible',
+					'Artifact Tokens': runtime?.canonicalTokenIds?.join(',') ?? 'absent',
+					'Context Key': runtime?.contextKey ?? '—',
+					'Exact Total':
+						r.lmEvidence?.totalLogProbability !== undefined
+							? Number(r.lmEvidence.totalLogProbability.toFixed(5))
+							: '—',
+					'Exact Mean':
+						r.lmEvidence?.meanTokenLogProbability !== undefined
+							? Number(r.lmEvidence.meanTokenLogProbability.toFixed(5))
+							: '—',
+					Assoc:
+						typeof original?.association === 'number'
+							? Number(original.association.toFixed(4))
+							: '—',
 					'Raw Logit': rawLogit,
 					'Session Freq': original?.sessionFreq || 0,
 				};
@@ -685,16 +2235,38 @@ export const predict = (textBefore: string): string | null => {
 
 			// eslint-disable-next-line no-console
 			console.table(tableData);
-		} else {
-			// eslint-disable-next-line no-console
-			console.log('No candidates found.');
 		}
 
 		// eslint-disable-next-line no-console
 		console.groupEnd();
 	}
 
-	return suggestion && suggestion.length > 0 ? suggestion : null;
+	const selectedContextKey = selected ? runtimeBySurface.get(selected.word)?.contextKey : undefined;
+
+	return suggestion && suggestion.length > 0 && selected && selectedEvidence
+		? {
+				text: suggestion,
+				surface: selected.word,
+				termType: selected.termType,
+				posterior: judgedPosterior(selected),
+				meanTokenLogProbability: judgedEvidenceFor(selected)?.mean ?? NaN,
+				poolHeldExtension: poolExtensionOf(selected) !== undefined,
+				shortlistSize:
+					selectedContextKey === undefined
+						? 1
+						: (contextTotals.get(selectedContextKey)?.length ?? 1),
+				rankScore: selectedDisplayScore,
+				evidenceTier:
+					selectedEvidence?.source === 'network-logit' ? 'network-logit' : 'canonical-full-surface',
+				evidenceDepth: {
+					verifiedTokens: selectedEvidence?.verifiedTokenCount ?? 0,
+					totalTokens: selectedEvidence?.totalTokenCount ?? 0,
+					verifiedChars: selectedEvidence?.verifiedCharCount ?? 0,
+					totalChars: selected.word.length,
+				},
+				winnerMargin,
+			}
+		: null;
 };
 
 // ─── Data Loading ────────────────────────────────────────────────────────────
@@ -710,6 +2282,100 @@ interface VocabularyJson {
 	>;
 }
 
+const isVocabularyJson = (payload: unknown): payload is VocabularyJson => {
+	if (payload == null || typeof payload !== 'object') {
+		return false;
+	}
+	const words = (payload as { words?: unknown }).words;
+	return words != null && typeof words === 'object';
+};
+
+const isStringArray = (payload: unknown): payload is string[] =>
+	Array.isArray(payload) && payload.every((entry) => typeof entry === 'string');
+
+type WordIndexPayload = Record<string, number> | { index: Record<string, number> };
+
+const isWordToOffsetMap = (value: unknown): value is Record<string, number> =>
+	value != null &&
+	typeof value === 'object' &&
+	Object.values(value as Record<string, unknown>).every((entry) => typeof entry === 'number');
+
+const isWordIndexPayload = (payload: unknown): payload is WordIndexPayload => {
+	if (payload == null || typeof payload !== 'object') {
+		return false;
+	}
+	const index = (payload as { index?: unknown }).index;
+	return index === undefined ? isWordToOffsetMap(payload) : isWordToOffsetMap(index);
+};
+
+/**
+ * The word index has shipped both bare and wrapped in `{ index: ... }`, so
+ * accept either rather than silently building an empty vector store.
+ */
+const unwrapWordIndex = (payload: WordIndexPayload): Record<string, number> => {
+	const wrapped = (payload as { index?: Record<string, number> }).index;
+	return wrapped ?? (payload as Record<string, number>);
+};
+
+/**
+ * Per-term stats shape for `bigrams.json` / `phrases.json`. Field names/types
+ * confirmed against the producer (`l2_vocabulary_creation._build_ngram_stat_payload`
+ * + `artifact_release._validate_ngram_payload`): `freq`/`doc_freq`/`author_freq`
+ * are ints, `pos_seq` a string array, `head_pos` a string or null, `association`
+ * a number.
+ */
+interface PhraseStatsJson {
+	association?: number;
+	author_freq: number;
+	doc_freq: number;
+	freq: number;
+	head_pos?: string | null;
+	pos_seq?: string[];
+}
+
+/** Normalized `{term: stats}` map after unwrapping any producer wrapper key. */
+type PhraseArtifactJson = Record<string, PhraseStatsJson>;
+
+const isPhraseStats = (value: unknown): value is PhraseStatsJson =>
+	value != null &&
+	typeof value === 'object' &&
+	typeof (value as { freq?: unknown }).freq === 'number';
+
+const isTermStatsMap = (value: unknown): value is PhraseArtifactJson => {
+	if (value == null || typeof value !== 'object') {
+		return false;
+	}
+	const values = Object.values(value as Record<string, unknown>);
+	return values.length > 0 && values.every(isPhraseStats);
+};
+
+/**
+ * Normalize a `bigrams.json` / `phrases.json` payload into a flat
+ * `{term: stats}` map. Tolerant of the exact wire shape (not yet finalized by
+ * the producer): accepts either the flat map from spec §9 or a single wrapper
+ * key (`terms`/`bigrams`/`phrases`/`words`). Returns `null` if unrecognized.
+ *
+ * :params:
+ *   payload: Raw JSON parsed from the artifact endpoint
+ * :returns:
+ *   A `{term: stats}` map, or `null` when the shape is not a term-stats map
+ */
+const normalizePhraseArtifact = (payload: unknown): PhraseArtifactJson | null => {
+	if (payload == null || typeof payload !== 'object') {
+		return null;
+	}
+	if (isTermStatsMap(payload)) {
+		return payload;
+	}
+	const obj = payload as Record<string, unknown>;
+	for (const key of ['terms', 'bigrams', 'phrases', 'words']) {
+		if (isTermStatsMap(obj[key])) {
+			return obj[key] as PhraseArtifactJson;
+		}
+	}
+	return null;
+};
+
 export const loadVectorsAsync = async (options?: {
 	isLocalLLM?: boolean;
 	surface?: string;
@@ -725,57 +2391,26 @@ export const loadVectorsAsync = async (options?: {
 		...(surface ? { surface } : {}),
 	});
 
-	let url: string;
-	try {
-		url = await getArtifactUrl(ARTIFACT_NAME.WORD_VECTORS);
-	} catch (e) {
-		vectorsLoadStarted = false;
-		failExp(EXPERIENCE_NAME.LOAD_VECTORS, 'singleton', {
-			isLocalLLM,
-			errorType: 'resolve_url',
-			...(surface ? { surface } : {}),
-		});
-		// eslint-disable-next-line no-console
-		console.warn('[text-predictor] Failed to resolve vectors URL:', e);
-		return;
-	}
-
 	try {
-		const res = await fetch(url);
-		if (!res.ok) {
-			vectorsLoadStarted = false;
-			failExp(EXPERIENCE_NAME.LOAD_VECTORS, 'singleton', {
-				isLocalLLM,
-				status: res.status,
-				errorType: 'http_error',
-				...(surface ? { surface } : {}),
-			});
-			// eslint-disable-next-line no-console
-			console.warn(`[text-predictor] Failed to load vectors: ${res.status}`);
-			return;
-		}
-		const buffer = await res.arrayBuffer();
+		const buffer = await fetchAutocompleteArtifactBinary(
+			ARTIFACT_NAME.WORD_VECTORS,
+			'word_vectors_10k',
+		);
 		const float32 = new Float32Array(buffer);
-
-		// The word index is published either bare or wrapped as `{ "index": {…} }`
-		// (the wrapper exists so no real entry, e.g. the word "default", can shadow
-		// a bundler-synthesised ESM export). Detect the wrapper by value type, not
-		// by key presence — "index" is itself a plausible vocabulary word, and in a
-		// bare index every value is a number.
-		const wordIndexPayload = await fetchArtifactJson<Record<string, unknown>>(
+		const wordIndexPayload = await fetchAutocompleteArtifactJson<WordIndexPayload>(
 			ARTIFACT_NAME.WORD_INDEX,
+			{
+				summarize: (payload) => `${Object.keys(unwrapWordIndex(payload)).length} entries`,
+				validate: isWordIndexPayload,
+			},
 		);
-		const wrappedIndex = wordIndexPayload?.index;
-		const wordIndex = (
-			typeof wrappedIndex === 'object' && wrappedIndex !== null
-				? wrappedIndex
-				: (wordIndexPayload ?? {})
-		) as Record<string, number>;
+		const wordIndex = unwrapWordIndex(wordIndexPayload);
 		const nWords = Object.keys(wordIndex).length;
 		if (nWords === 0) {
-			// eslint-disable-next-line no-console
-			console.warn(
-				`[text-predictor] ${ARTIFACT_NAME.WORD_INDEX} was empty — semantic scoring will be a no-op.`,
+			ctcTag(
+				'init',
+				'⚠️ word_index_10k.json was empty — semantic scoring will be a no-op',
+				CTC_STYLES.warn,
 			);
 		}
 		const dim = float32.length / nWords;
@@ -788,14 +2423,7 @@ export const loadVectorsAsync = async (options?: {
 			sizeBytes: float32.byteLength,
 			...(surface ? { surface } : {}),
 		});
-		if (isAutocompleteDebugEnabled()) {
-			// eslint-disable-next-line no-console
-			console.log('[text-predictor] Vectors loaded:', {
-				wordCount: nWords,
-				dim,
-				sizeBytes: float32.byteLength,
-			});
-		}
+		ctcTag('init', `vectors loaded: ${nWords} words · dim ${dim} · ${float32.byteLength}B`);
 	} catch (e) {
 		vectorsLoadStarted = false;
 		failExp(EXPERIENCE_NAME.LOAD_VECTORS, 'singleton', {
@@ -803,8 +2431,7 @@ export const loadVectorsAsync = async (options?: {
 			errorType: 'network',
 			...(surface ? { surface } : {}),
 		});
-		// eslint-disable-next-line no-console
-		console.warn('[text-predictor] Failed to load vectors:', e);
+		ctcTag('init', `⚠️ failed to load vectors: ${String(e)}`, CTC_STYLES.bad);
 	}
 };
 
@@ -812,10 +2439,92 @@ export const initVectors = (store: VectorStore): void => {
 	vectorStore = store;
 };
 
+/**
+ * Load the producer's `bigrams.json` + `phrases.json` completion-term artifacts
+ * over the TDP-OS gateway and insert them into the phrase trie.
+ *
+ * Each file is fetched independently (`Promise.allSettled`) so a missing or
+ * late-published artifact only skips that term type — word completion and the
+ * other term type are unaffected. Phrase VECTORS are not fetched here: per the
+ * producer's `extend` decision they are appended into the existing
+ * `word-index-10k.json` / `word-vectors-10k.bin`, so `loadVectorsAsync` already
+ * covers them.
+ *
+ * :params:
+ *   options.isLocalLLM: Tags the UFO experience so latency/success feeds the same SLO
+ * :returns:
+ *   A promise that resolves once both fetches have settled
+ */
+export const loadPhraseArtifacts = async (options?: { isLocalLLM?: boolean }): Promise<void> => {
+	if (phrasesLoadStarted) {
+		return;
+	}
+	phrasesLoadStarted = true;
+	const isLocalLLM = options?.isLocalLLM ?? false;
+	startExp(EXPERIENCE_NAME.LOAD_PHRASES, 'singleton', { isLocalLLM });
+
+	const loadOne = async (
+		artifactName: ArtifactName,
+		termType: TermType,
+		label: string,
+	): Promise<number> => {
+		const payload = await fetchAutocompleteArtifactJson<unknown>(artifactName, {
+			label,
+			validate: (p): p is unknown => p != null && typeof p === 'object',
+		});
+		const normalized = normalizePhraseArtifact(payload);
+		if (!normalized) {
+			throw new Error(`[text-predictor] ${label} payload was not a recognised term-stats map`);
+		}
+		return initPhrases(normalized, termType);
+	};
+
+	const [bigramsResult, phrasesResult] = await Promise.allSettled([
+		loadOne(ARTIFACT_NAME.BIGRAMS, 'bigram', 'bigrams'),
+		loadOne(ARTIFACT_NAME.PHRASES, 'phrase', 'phrases'),
+	]);
+
+	const bigramCount = bigramsResult.status === 'fulfilled' ? bigramsResult.value : 0;
+	const phraseCount = phrasesResult.status === 'fulfilled' ? phrasesResult.value : 0;
+
+	if (bigramsResult.status === 'rejected') {
+		ctcTag(
+			'init',
+			`⚠️ failed to load bigrams.json: ${String(bigramsResult.reason)}`,
+			CTC_STYLES.warn,
+		);
+	}
+	if (phrasesResult.status === 'rejected') {
+		ctcTag(
+			'init',
+			`⚠️ failed to load phrases.json: ${String(phrasesResult.reason)}`,
+			CTC_STYLES.warn,
+		);
+	}
+
+	if (bigramsResult.status === 'fulfilled' || phrasesResult.status === 'fulfilled') {
+		phrasesLoaded = true;
+		succeedExp(EXPERIENCE_NAME.LOAD_PHRASES, 'singleton', {
+			isLocalLLM,
+			bigramCount,
+			phraseCount,
+		});
+		ctcTag('init', `phrase artifacts loaded: ${bigramCount} bigrams · ${phraseCount} phrases`);
+	} else {
+		// Both failed — e.g. artifacts not published to the tenant manifest yet.
+		// Reset the guard so a later focus/predict can retry once they land.
+		phrasesLoadStarted = false;
+		failExp(EXPERIENCE_NAME.LOAD_PHRASES, 'singleton', { isLocalLLM, errorType: 'network' });
+	}
+};
+
 let vocabularyLoadPromise: Promise<void> | undefined;
 
+type VocabularyLoadSource = 'focus' | 'predict';
+
 export const loadDefaultVocabulary = (options?: {
 	isLocalLLM?: boolean;
+	source?: VocabularyLoadSource;
 	surface?: string;
 }): Promise<void> => {
 	if (isInitialized) {
@@ -826,6 +2535,7 @@ export const loadDefaultVocabulary = (options?: {
 	}
 
 	const isLocalLLM = options?.isLocalLLM ?? false;
+	const source = options?.source ?? 'predict';
 	const surface = options?.surface;
 	vocabularyLoadPromise = (async () => {
 		startExp(EXPERIENCE_NAME.LOAD_VOCABULARY, 'singleton', {
@@ -834,19 +2544,56 @@ export const loadDefaultVocabulary = (options?: {
 		});
 
 		try {
-			const [vocabularyData, l3VocabularyData] = await Promise.all([
-				fetchArtifactJson<VocabularyJson>(ARTIFACT_NAME.VOCABULARY),
-				fetchArtifactJson<string[]>(ARTIFACT_NAME.L3_VOCABULARY),
-			]);
-
-			if (vocabularyData?.words == null || !Array.isArray(l3VocabularyData)) {
-				throw new Error(
-					`[text-predictor] unexpected vocabulary payload — ` +
-						`${ARTIFACT_NAME.VOCABULARY} ${vocabularyData?.words == null ? 'is missing its `words` wrapper' : 'is valid'}, ` +
-						`${ARTIFACT_NAME.L3_VOCABULARY} ${Array.isArray(l3VocabularyData) ? 'is valid' : 'is not an array'}`,
+			ctcTag('init', `loading artifacts (trigger: ${source} · localLLM: ${isLocalLLM})`);
+			if (isAutocompleteDebugVerbose()) {
+				// eslint-disable-next-line no-console
+				console.groupCollapsed(
+					'%c[CTC:init]%c artifact manifest',
+					CTC_STYLES.brand,
+					CTC_STYLES.body,
 				);
+				// eslint-disable-next-line no-console
+				console.log('Eager:', [
+					ARTIFACT_NAME.VOCABULARY,
+					ARTIFACT_NAME.L3_VOCABULARY,
+					ARTIFACT_NAME.POS_TAGS,
+					ARTIFACT_NAME.GHOST_POS_TAGS,
+					ARTIFACT_NAME.GRAMMAR_TRANSITIONS,
+				]);
+				// eslint-disable-next-line no-console
+				console.log('Conditional:', [
+					`${ARTIFACT_NAME.WORD_INDEX} (only when vectors load; phrase/bigram surfaces share this index)`,
+					`${ARTIFACT_NAME.BIGRAMS} + ${ARTIFACT_NAME.PHRASES} (completion-term stats; fire-and-forget)`,
+					`${ARTIFACT_NAME.PHRASE_CONTINUATION_TOKENS} (local model only; canonical token IDs for the full served union)`,
+				]);
+				// eslint-disable-next-line no-console
+				console.groupEnd();
 			}
 
+			const [vocabularyData, l3VocabularyData] = await Promise.all([
+				fetchAutocompleteArtifactJson<VocabularyJson>(ARTIFACT_NAME.VOCABULARY, {
+					summarize: (payload) => `${Object.keys(payload.words).length} words`,
+					validate: isVocabularyJson,
+				}),
+				fetchAutocompleteArtifactJson<string[]>(ARTIFACT_NAME.L3_VOCABULARY, {
+					summarize: (payload) => `${payload.length} words`,
+					validate: isStringArray,
+				}),
+				// Awaited alongside the vocabulary so the grammar filter is settled
+				// before the first suggestion can be produced. Without this the
+				// predictor initialises on the smaller vocabulary payload and serves
+				// candidates the filter would have dropped until the grammar lands.
+				// Only settlement is required: a failed load leaves the filter as a
+				// pass-through, which is the same degradation as an absent artifact.
+				loadGrammarDataAsync({ isLocalLLM, surface }).catch((error) => {
+					ctcTag(
+						'init',
+						`⚠️ grammar data unavailable; filtering skipped: ${String(error)}`,
+						CTC_STYLES.warn,
+					);
+				}),
+			]);
+
 			const terms = Object.entries(vocabularyData.words).map(([word, stats]) => ({
 				word,
 				freq: stats.freq,
@@ -860,6 +2607,10 @@ export const loadDefaultVocabulary = (options?: {
 			initL3Vocabulary(l3VocabularyData);
 			initVocabulary({ terms });
 
+			// Phrase/bigram artifacts load independently and must never affect the
+			// word path — fire-and-forget with its own error handling inside.
+			void loadPhraseArtifacts({ isLocalLLM }).catch(() => {});
+
 			succeedExp(EXPERIENCE_NAME.LOAD_VOCABULARY, 'singleton', {
 				isLocalLLM,
 				l2WordCount: terms.length,
@@ -869,7 +2620,7 @@ export const loadDefaultVocabulary = (options?: {
 		} catch (e) {
 			failExp(EXPERIENCE_NAME.LOAD_VOCABULARY, 'singleton', {
 				isLocalLLM,
-				errorType: 'load_error',
+				errorType: 'parse_error',
 				...(surface ? { surface } : {}),
 			});
 			// Allow a later call to retry the load rather than caching the failure.