@equationalapplications/core-llm-wiki 7.4.0 → 7.6.0

package/dist/testing.js

@@ -593,6 +593,19 @@ function validateTags(tags) {
   if (!Array.isArray(tags)) return [];
   return tags.filter((t) => typeof t === "string").map((t) => t.trim().toLowerCase()).filter((t) => t.length > 0 && t.length <= 40).slice(0, 6);
 }
+var MAX_EVIDENCE_QUOTES = 10;
+function normalizeEvidence(raw) {
+  if (!Array.isArray(raw)) return void 0;
+  const out = [];
+  for (const entry of raw) {
+    if (typeof entry !== "string") continue;
+    const trimmed = entry.trim();
+    if (!trimmed) continue;
+    out.push(trimmed);
+    if (out.length > MAX_EVIDENCE_QUOTES) break;
+  }
+  return out;
+}
 function validateFact(fact) {
   if (typeof fact?.title !== "string" || typeof fact?.body !== "string") return null;
   const title = clip(fact.title, 80);
@@ -600,13 +613,17 @@ function validateFact(fact) {
   if (!title || !body) return null;
   let confidence = fact.confidence;
   if (confidence !== "certain" && confidence !== "tentative") confidence = "inferred";
-  return {
+  const valid = {
     ...fact,
     title,
     body,
     confidence,
     tags: validateTags(fact.tags)
   };
+  const evidence = normalizeEvidence(fact.evidence);
+  if (evidence) valid.evidence = evidence;
+  else delete valid.evidence;
+  return valid;
 }
 function validateTask(task) {
   if (typeof task?.description !== "string") return null;
@@ -817,8 +834,7 @@ var EmbeddingService = class {
     }
   }
   async tryEmbedFact(fact, ctx) {
-    const embedFn = this.options.llmProvider.embed;
-    if (typeof embedFn !== "function") return { ok: false, kind: "no_provider" };
+    if (typeof this.options.llmProvider.embed !== "function") return { ok: false, kind: "no_provider" };
     let tagsStr;
     if (Array.isArray(fact.tags)) {
       tagsStr = fact.tags.join(" ");
@@ -835,7 +851,7 @@ var EmbeddingService = class {
     const text = clip(`${fact.title} ${fact.body} ${tagsStr}`.trim(), maxEmbedChars);
     let float32Vector;
     try {
-      const vector = await embedFn(text);
+      const vector = await this.options.llmProvider.embed(text);
       if (vector.length === 0 || !vector.every((v) => typeof v === "number" && isFinite(v))) {
         console.warn(`[WikiMemory] embedFact: embed() returned an invalid vector for ${fact.id}; skipping.`);
         this.reportEmbed(ctx, fact, "embedding_failed", "invalid_vector");
@@ -1510,17 +1526,94 @@ Return ONLY a valid JSON object matching this schema:
 }
 If no manifest type fits a fact, omit that fact from "classifications" entirely \u2014 do not guess.
 When echoing an existing fact's title verbatim into "target_title", preserve every JSON escape sequence (\\", \\n, \\\\, \\/) exactly as it appeared in the input body \u2014 do not strip backslashes, do not add unescaped quotes. Do not return markdown, just raw JSON.`;
+var GROUNDING_SOURCE = {
+  ingest: { key: "facts", source: "the document chunk" },
+  librarian: { key: "facts", source: 'the "summary" text of the events' },
+  heal: { key: "newFacts", source: 'the "summary" text of the recent events or the "body" text of the document anchors' }
+};
+function groundingEvidenceBlock(writer, cfg) {
+  const { key, source } = GROUNDING_SOURCE[writer];
+  return `EVIDENCE REQUIREMENT: every object in "${key}" must also carry an "evidence" array of 1 to ${cfg.maxEvidence} quotes. Each quote must be an exact substring copied character-for-character from ${source} (the SOURCE section), at least ${cfg.minEvidenceChars} characters long. Do not paraphrase. Do not quote these instructions, the ontology manifest, or any existing fact. A fact whose quotes cannot be found in the SOURCE section is stored as an unreviewed draft.
+"evidence": ["exact substring copied from the SOURCE section"]`;
+}
 
 // src/utils/healConstants.ts
 var HEAL_MAX_ANCHORS = 50;
 var HEAL_ANCHORS_PER_CANDIDATE = 4;
 var HEAL_MAX_FACT_BODY_CHARS_L3 = 4e3;
 var HEAL_MAX_TASKS = 50;
+var HEAL_ANCHOR_BODY_CHARS = 800;
+
+// src/utils/grounding.ts
+var GROUNDING_VERIFIER = "process:grounding-check";
+var WRITERS = ["ingest", "librarian", "heal"];
+function positiveInt(value, fallback) {
+  return typeof value === "number" && Number.isFinite(value) && value >= 1 ? Math.floor(value) : fallback;
+}
+function resolveGrounding(config) {
+  if (!config || config.mode !== "draft") return null;
+  const writers = Array.isArray(config.writers) ? config.writers.filter((w) => WRITERS.includes(w)) : ["ingest"];
+  return {
+    writers: new Set(writers),
+    minEvidenceChars: positiveInt(config.minEvidenceChars, 20),
+    // Never ask for more quotes than checkGrounding accepts (MAX_EVIDENCE_QUOTES).
+    maxEvidence: Math.min(MAX_EVIDENCE_QUOTES, positiveInt(config.maxEvidence, 3)),
+    maxEvidenceChars: positiveInt(config.maxEvidenceChars, 300)
+  };
+}
+function normalizeForGrounding(text) {
+  return text.normalize("NFKC").replace(/\s+/g, " ").trim();
+}
+function buildGroundingCorpus(parts) {
+  return parts.filter((p) => typeof p === "string").map(normalizeForGrounding);
+}
+function checkGrounding(evidence, normalizedCorpus, cfg) {
+  const quotes = evidence ?? [];
+  if (quotes.length > MAX_EVIDENCE_QUOTES) return { status: "failed", reason: "too_many_quotes" };
+  if (quotes.length === 0) return { status: "missing", reason: "no_evidence" };
+  const qualifying = quotes.map(normalizeForGrounding).filter((q) => q.length >= cfg.minEvidenceChars);
+  if (qualifying.length === 0) return { status: "missing", reason: "evidence_too_short" };
+  if (qualifying.some((q) => !normalizedCorpus.some((part) => part.includes(q)))) {
+    return { status: "failed", reason: "quote_not_found" };
+  }
+  return {
+    status: "grounded",
+    retained: qualifying.slice(0, cfg.maxEvidence).map((q) => safeSlice(q, 0, cfg.maxEvidenceChars))
+  };
+}
+function groundingOutcome(verdict, now) {
+  if (verdict.status === "grounded") {
+    return {
+      trust: {
+        lifecycle_status: "stable",
+        okf_verified: [{ by: GROUNDING_VERIFIER, at: new Date(now).toISOString() }],
+        last_verified_at: now,
+        last_verified_by: GROUNDING_VERIFIER
+      }
+    };
+  }
+  return {
+    trust: { lifecycle_status: "draft" },
+    diagnostic: { code: verdict.status === "missing" ? "grounding_missing" : "grounding_failed", reason: verdict.reason }
+  };
+}
 
 // src/services/PromptService.ts
 var PromptService = class {
-  constructor(globalOverrides) {
+  constructor(globalOverrides, grounding = null) {
     this.globalOverrides = globalOverrides;
+    this.grounding = grounding;
+  }
+  /** The resolved grounding config when `writer` is in `grounding.writers`; otherwise null (spec §6.2). */
+  groundingFor(writer) {
+    return this.grounding?.writers.has(writer) ? this.grounding : null;
+  }
+  /** Appended after any override and after ontology context, so it is always last. */
+  appendGrounding(systemPrompt, writer) {
+    const cfg = this.groundingFor(writer);
+    return cfg ? `${systemPrompt}
+
+${groundingEvidenceBlock(writer, cfg)}` : systemPrompt;
   }
   hydrate(template, variables) {
     return template.replace(/\{\{\s*(\w+)\s*\}\}/g, (_match, key) => {
@@ -1550,13 +1643,13 @@ ${ctx.ontologyModeInstructions}`;
     const hasDocumentChunk = /\{\{\s*documentChunk\s*\}\}/.test(template);
     if (hasDocumentChunk || this.hasOntologyPlaceholders(template)) {
       return {
-        systemPrompt: this.buildSystemPrompt(template, { documentChunk }, ontologyContext),
+        systemPrompt: this.appendGrounding(this.buildSystemPrompt(template, { documentChunk }, ontologyContext), "ingest"),
         userPrompt: hasDocumentChunk ? "Please extract the facts." : `Document Chunk:
 ${documentChunk}`
       };
     }
     return {
-      systemPrompt: this.appendOntology(template, ontologyContext),
+      systemPrompt: this.appendGrounding(this.appendOntology(template, ontologyContext), "ingest"),
       userPrompt: `Document Chunk:
 ${documentChunk}`
     };
@@ -1565,23 +1658,27 @@ ${documentChunk}`
     const template = runtimeOverride ?? this.globalOverrides?.librarianSystemPrompt ?? LIBRARIAN_SYSTEM_PROMPT;
     const hasEvents = /\{\{\s*events\s*\}\}/.test(template);
     const hasCurrentFacts = /\{\{\s*currentFacts\s*\}\}/.test(template);
+    const eventsShown = hasEvents || !hasCurrentFacts;
+    const corpusField = this.groundingFor("librarian") ? { groundingCorpus: buildGroundingCorpus(eventsShown ? events.map(summaryOf) : []) } : {};
     if (hasEvents || hasCurrentFacts || this.hasOntologyPlaceholders(template)) {
       return {
-        systemPrompt: this.buildSystemPrompt(template, { events, currentFacts }, ontologyContext),
+        systemPrompt: this.appendGrounding(this.buildSystemPrompt(template, { events, currentFacts }, ontologyContext), "librarian"),
         userPrompt: hasEvents || hasCurrentFacts ? "Please synthesize the context." : `Events:
 ${JSON.stringify(events, null, 2)}
 
 Current Facts:
-${JSON.stringify(currentFacts, null, 2)}`
+${JSON.stringify(currentFacts, null, 2)}`,
+        ...corpusField
       };
     }
     return {
-      systemPrompt: this.appendOntology(template, ontologyContext),
+      systemPrompt: this.appendGrounding(this.appendOntology(template, ontologyContext), "librarian"),
       userPrompt: `Events:
 ${JSON.stringify(events, null, 2)}
 
 Current Facts:
-${JSON.stringify(currentFacts, null, 2)}`
+${JSON.stringify(currentFacts, null, 2)}`,
+      ...corpusField
     };
   }
   /**
@@ -1611,40 +1708,54 @@ ${JSON.stringify(currentFacts, null, 2)}`
     const effectiveEvents = attemptLevel >= 2 ? [] : recentEvents;
     const maxAnchors = Math.max(1, Math.min(HEAL_MAX_ANCHORS, healCandidates.length * HEAL_ANCHORS_PER_CANDIDATE));
     const effectiveAnchors = documentAnchors.slice(0, maxAnchors);
+    const template = runtimeOverride ?? this.globalOverrides?.healSystemPrompt ?? HEAL_SYSTEM_PROMPT;
+    const hasRecentEvents = /\{\{\s*recentEvents\s*\}\}/.test(template);
+    const hasDocumentAnchors = /\{\{\s*documentAnchors\s*\}\}/.test(template);
+    const usesPlaceholders = /\{\{\s*healCandidates\s*\}\}/.test(template) || hasDocumentAnchors || /\{\{\s*allTasks\s*\}\}/.test(template) || hasRecentEvents;
+    const healGrounding = this.groundingFor("heal");
+    const promptAnchors = healGrounding ? effectiveAnchors.map(toGroundingAnchor) : effectiveAnchors;
+    const eventsShown = !usesPlaceholders || hasRecentEvents;
+    const anchorsShown = !usesPlaceholders || hasDocumentAnchors;
+    const groundingCorpus = healGrounding ? buildGroundingCorpus([
+      ...eventsShown ? effectiveEvents.map(summaryOf) : [],
+      ...anchorsShown ? effectiveAnchors.map((a, i) => ({ status: a?.lifecycle_status, shown: promptAnchors[i] })).filter((x) => x.status !== "draft").map((x) => x.shown?.body) : []
+    ]) : null;
     const { shapedCandidates, degraded } = applyBodyTruncation(
       healCandidates,
       attemptLevel,
       bodyTruncationChars
     );
-    const template = runtimeOverride ?? this.globalOverrides?.healSystemPrompt ?? HEAL_SYSTEM_PROMPT;
-    if (/\{\{\s*healCandidates\s*\}\}/.test(template) || /\{\{\s*documentAnchors\s*\}\}/.test(template) || /\{\{\s*allTasks\s*\}\}/.test(template) || /\{\{\s*recentEvents\s*\}\}/.test(template)) {
+    const corpusField = groundingCorpus !== null ? { groundingCorpus } : {};
+    if (usesPlaceholders) {
       return {
         prompts: {
-          systemPrompt: this.hydrate(template, {
+          systemPrompt: this.appendGrounding(this.hydrate(template, {
             healCandidates: shapedCandidates,
-            documentAnchors: effectiveAnchors,
+            documentAnchors: promptAnchors,
             allTasks: effectiveTasks,
             recentEvents: effectiveEvents
-          }),
+          }), "heal"),
           userPrompt: "Please heal the memory graph."
         },
-        degraded
+        degraded,
+        ...corpusField
       };
     }
     return {
       prompts: {
-        systemPrompt: template,
+        systemPrompt: this.appendGrounding(template, "heal"),
         userPrompt: `Heal Candidates:
 ${JSON.stringify(shapedCandidates, null, 2)}
 Document Anchors (DO NOT MODIFY OR DELETE):
-${JSON.stringify(effectiveAnchors, null, 2)}
+${JSON.stringify(promptAnchors, null, 2)}
 All Tasks:
 ${JSON.stringify(effectiveTasks, null, 2)}
 Recent Events:
 ${JSON.stringify(effectiveEvents, null, 2)}
 The following document anchors are provided for contradiction detection only. Do not include them in \`downgraded\`, \`deleted\`, or \`newFacts\`.`
       },
-      degraded
+      degraded,
+      ...corpusField
     };
   }
   buildOntologyBackfillPrompt(facts, runtimeOverride, ontologyContext) {
@@ -1694,6 +1805,19 @@ function applyBodyTruncation(candidates, attemptLevel, bodyTruncationChars) {
   }
   return { shapedCandidates, degraded };
 }
+function summaryOf(event) {
+  return event?.summary;
+}
+function toGroundingAnchor(anchor) {
+  if (typeof anchor !== "object" || anchor === null) return anchor;
+  const a = anchor;
+  return {
+    id: a.id,
+    title: a.title,
+    source_ref: a.source_ref,
+    body: typeof a.body === "string" ? safeSlice(a.body, 0, HEAL_ANCHOR_BODY_CHARS) : ""
+  };
+}
 
 // src/utils/chunkingDefaults.ts
 var DEFAULT_MAX_CHUNK_LENGTH = 12e3;
@@ -1723,7 +1847,7 @@ var IngestionService = class {
     this.jobManager = jobManager;
     this.embeddingService = embeddingService;
     this.ontologyService = ontologyService;
-    this.promptService = promptService ?? new PromptService(this.options.config?.prompts);
+    this.promptService = promptService ?? new PromptService(this.options.config?.prompts, resolveGrounding(this.options.config?.grounding));
   }
   async ingestDocument(entityId, params, opts) {
     const sourceRef = normalizeSourceRef(params.sourceRef);
@@ -1756,6 +1880,7 @@ var IngestionService = class {
       const { chunks, truncated } = chunkText(params.documentChunk, maxChunkLength, chunkOverlap);
       if (chunks.length === 0) return zeroChunkResult();
       const ontologyContext = await this.ontologyService?.buildPromptContext(entityId) ?? null;
+      const ingestGrounding = this.promptService.groundingFor("ingest");
       const chunkResults = await withConcurrency(
         chunks.map((chunk, chunkIndex) => async () => {
           const { systemPrompt, userPrompt } = this.promptService.buildIngestPrompt(
@@ -1779,11 +1904,16 @@ var IngestionService = class {
                 rejected.push({ itemIndex, reason: factRejectionReason(raw) });
               }
             });
+            const verdicts = ingestGrounding ? (() => {
+              const corpus = buildGroundingCorpus([chunk]);
+              return facts.map((f) => checkGrounding(f.evidence, corpus, ingestGrounding));
+            })() : [];
             return {
               status: "ok",
               facts,
               itemIndexes,
               rejected,
+              verdicts,
               ontology_updates: result2.ontology_updates
             };
           } catch (e) {
@@ -1828,6 +1958,7 @@ var IngestionService = class {
       const failures = [];
       const seen = /* @__PURE__ */ new Set();
       const orderedChunkFacts = [];
+      const groundingLedger = /* @__PURE__ */ new Map();
       const diagBuffer = new DiagnosticBuffer();
       const diagBase = { entityId, operation: "ingest", trigger: "call" };
       for (const [chunkIndex, slot] of chunkResults.entries()) {
@@ -1846,6 +1977,7 @@ var IngestionService = class {
           if (!seen.has(normalizedTitle)) {
             seen.add(normalizedTitle);
             dedupedFacts.push(fact);
+            if (ingestGrounding) groundingLedger.set(fact, { verdict: slot.verdicts[k], chunkIndex, itemIndex: slot.itemIndexes[k] });
           } else {
             diagBuffer.push({
               ...diagBase,
@@ -1882,7 +2014,7 @@ var IngestionService = class {
       try {
         if (failedChunks === 0) {
           const fullResult = await this.db.withTransactionAsync(async (tx) => {
-            return await this.runFullUpsertGraph(entityId, sourceRef, sourceHash, orderedChunkFacts, tx, diagBuffer);
+            return await this.runFullUpsertGraph(entityId, sourceRef, sourceHash, orderedChunkFacts, tx, diagBuffer, groundingLedger);
           });
           deletedSourceFactIds.push(...fullResult.deletedSourceFactIds);
           insertedFacts.push(...fullResult.insertedFacts);
@@ -1890,7 +2022,7 @@ var IngestionService = class {
           const partialResult = await this.db.withTransactionAsync(async (tx) => {
             const flat = [];
             for (const slot of orderedChunkFacts) flat.push(...slot.facts);
-            return await this.appendPartialFacts(entityId, sourceRef, flat, tx, diagBuffer);
+            return await this.appendPartialFacts(entityId, sourceRef, flat, tx, diagBuffer, groundingLedger);
           });
           insertedFacts.push(...partialResult.insertedDescriptors);
         }
@@ -2032,7 +2164,8 @@ var IngestionService = class {
         last_accessed_at: null,
         access_count: 0,
         deleted_at: null,
-        okf_type: normalized.okf_type
+        okf_type: normalized.okf_type,
+        ...opts?.nodeTrust?.get(node.id)
       };
       wikiFacts.push(wikiFact);
     }
@@ -2123,9 +2256,10 @@ var IngestionService = class {
    * On the partial path the caller never invokes this method; the empty
    * arrays stay empty.
    */
-  async runFullUpsertGraph(entityId, sourceRef, sourceHash, orderedChunkFacts, tx, diagBuffer) {
+  async runFullUpsertGraph(entityId, sourceRef, sourceHash, orderedChunkFacts, tx, diagBuffer, groundingLedger = /* @__PURE__ */ new Map()) {
     const deletedSourceFactIds = [];
     const insertedFacts = [];
+    const nodeTrust = /* @__PURE__ */ new Map();
     deletedSourceFactIds.push(...await this.entryRepo.findIdsBySource(entityId, sourceRef, null, tx, false));
     const titleIndex = /* @__PURE__ */ new Map();
     const existingFacts = await this.entryRepo.findRecentByEntityId(entityId, 500, tx, sourceRef);
@@ -2169,6 +2303,20 @@ var IngestionService = class {
           tags: fact.tags,
           confidence: fact.confidence
         });
+        const grounding = groundingLedger.get(fact);
+        if (grounding) {
+          const outcome = groundingOutcome(grounding.verdict, now);
+          nodeTrust.set(id, outcome.trust);
+          if (outcome.diagnostic) {
+            diagBuffer.push({
+              entityId,
+              operation: "ingest",
+              trigger: "call",
+              code: outcome.diagnostic.code,
+              detail: { factId: id, sourceRef, chunkIndex: grounding.chunkIndex, itemIndex: grounding.itemIndex, reason: outcome.diagnostic.reason }
+            });
+          }
+        }
         insertedFacts.push({ id, entity_id: entityId, title: fact.title, body: fact.body, tags: JSON.stringify(fact.tags) });
         titleIndex.set(normalizeTitleKey(fact.title), { id, okf_type: normalized.okf_type });
         if (normalized.edges.length > 0) {
@@ -2200,7 +2348,7 @@ var IngestionService = class {
       entityId,
       { sourceRef, sourceHash, nodes: hostNodes, edges: hostEdges },
       tx,
-      { strict: false, diag: { buffer: diagBuffer, operation: "ingest" } }
+      { strict: false, diag: { buffer: diagBuffer, operation: "ingest" }, ...nodeTrust.size > 0 ? { nodeTrust } : {} }
     );
     return { deletedSourceFactIds, insertedFacts };
   }
@@ -2224,7 +2372,7 @@ var IngestionService = class {
    * Runs INSIDE the caller's `tx`. Does not open a nested transaction.
    * Returns `{ inserted, skippedDuplicate }` for observability.
    */
-  async appendPartialFacts(entityId, sourceRef, dedupedFacts, tx, diagBuffer) {
+  async appendPartialFacts(entityId, sourceRef, dedupedFacts, tx, diagBuffer, groundingLedger = /* @__PURE__ */ new Map()) {
     const liveIds = await this.entryRepo.findIdsBySource(entityId, sourceRef, null, tx, false);
     const liveFacts = liveIds.length === 0 ? [] : await this.entryRepo.findByIds(liveIds, void 0, tx);
     const liveTitles = new Set(liveFacts.map((f) => normalizeTitleKey(f.title)));
@@ -2247,6 +2395,17 @@ var IngestionService = class {
       }
       liveTitles.add(normalizedTitle);
       const id = generateId("fact_");
+      const grounding = groundingLedger.get(fact);
+      const outcome = grounding ? groundingOutcome(grounding.verdict, now) : null;
+      if (grounding && outcome?.diagnostic) {
+        diagBuffer.push({
+          entityId,
+          operation: "ingest",
+          trigger: "call",
+          code: outcome.diagnostic.code,
+          detail: { factId: id, sourceRef, chunkIndex: grounding.chunkIndex, itemIndex: grounding.itemIndex, reason: outcome.diagnostic.reason }
+        });
+      }
       const wikiFact = {
         id,
         entity_id: entityId,
@@ -2265,7 +2424,8 @@ var IngestionService = class {
         last_accessed_at: null,
         access_count: 0,
         deleted_at: null,
-        okf_type: null
+        okf_type: null,
+        ...outcome?.trust
       };
       await this.entryRepo.upsert(wikiFact, tx);
       insertedDescriptors.push({
@@ -2281,6 +2441,51 @@ var IngestionService = class {
   }
 };
 
+// src/utils/classifier.ts
+var isUnit = (v) => typeof v === "number" && Number.isFinite(v) && v >= 0 && v <= 1;
+function validateClassifierAnswer(response, key, question) {
+  try {
+    if (response === null || typeof response !== "object") return { ok: false, reason: "malformed" };
+    const answers = response.answers;
+    if (answers === null || typeof answers !== "object") return { ok: false, reason: "missing_answer" };
+    if (!Object.prototype.hasOwnProperty.call(answers, key)) return { ok: false, reason: "missing_answer" };
+    const raw = answers[key];
+    if (raw === null || typeof raw !== "object") return { ok: false, reason: "missing_answer" };
+    const a = raw;
+    if (a.kind !== question.kind) return { ok: false, reason: "kind_mismatch" };
+    if (question.kind === "choice") {
+      if (typeof a.choice !== "string" || !question.options.includes(a.choice)) return { ok: false, reason: "choice_not_offered" };
+      if (!isUnit(a.confidence)) return { ok: false, reason: "invalid_probability" };
+      const probs = a.probabilities;
+      if (probs === null || typeof probs !== "object" || Array.isArray(probs)) return { ok: false, reason: "invalid_probability" };
+      const probabilities = {};
+      for (const [k, v] of Object.entries(probs)) {
+        if (!isUnit(v)) return { ok: false, reason: "invalid_probability" };
+        probabilities[k] = v;
+      }
+      return { ok: true, answer: { kind: "choice", choice: a.choice, confidence: a.confidence, probabilities } };
+    }
+    if (question.kind === "binary") {
+      if (!isUnit(a.probability)) return { ok: false, reason: "invalid_probability" };
+      return { ok: true, answer: { kind: "binary", probability: a.probability } };
+    }
+    const maxScore = question.levels.length - 1;
+    if (typeof a.score !== "number" || !Number.isFinite(a.score) || a.score < 0 || a.score > maxScore) {
+      return { ok: false, reason: "score_out_of_range" };
+    }
+    if (!isUnit(a.confidence)) return { ok: false, reason: "invalid_probability" };
+    if (!Array.isArray(a.probabilities) || !a.probabilities.every(isUnit)) return { ok: false, reason: "invalid_probability" };
+    return { ok: true, answer: { kind: "score", score: a.score, confidence: a.confidence, probabilities: a.probabilities.slice() } };
+  } catch {
+    return { ok: false, reason: "malformed" };
+  }
+}
+function classifierStateForFact(fact) {
+  const parts = [fact.title, fact.body];
+  if (Array.isArray(fact.tags) && fact.tags.length > 0) parts.push(`Tags: ${fact.tags.join(", ")}`);
+  return parts.join("\n\n");
+}
+
 // src/utils/embedding.ts
 function parseEmbedding(blob, text) {
   if (blob && blob.byteLength > 0) {
@@ -2416,7 +2621,7 @@ async function runBatched(args) {
     }
     let result;
     try {
-      result = parse(responseText, batch);
+      result = parse(responseText, batch, prompts);
     } catch (err) {
       await onFailure(batch, err, attemptLevel, false);
       return;
@@ -2497,7 +2702,7 @@ var MaintenanceService = class {
     this.jobManager = jobManager;
     this.embeddingService = embeddingService;
     this.ontologyService = ontologyService;
-    this.promptService = promptService ?? new PromptService(this.options.config?.prompts);
+    this.promptService = promptService ?? new PromptService(this.options.config?.prompts, resolveGrounding(this.options.config?.grounding));
   }
   async runPrune(entityId, options) {
     this.jobManager.acquireLock("prune", entityId);
@@ -2802,8 +3007,10 @@ var MaintenanceService = class {
       };
     });
     const ontologyContext = await this.ontologyService?.buildPromptContext(entityId) ?? null;
-    const { systemPrompt, userPrompt } = this.promptService.buildLibrarianPrompt(
-      events.reverse(),
+    const promptEvents = events.reverse();
+    const librarianGrounding = this.promptService.groundingFor("librarian");
+    const { systemPrompt, userPrompt, groundingCorpus: librarianCorpus = [] } = this.promptService.buildLibrarianPrompt(
+      promptEvents,
       currentFacts,
       promptOverride,
       ontologyContext
@@ -2872,6 +3079,10 @@ var MaintenanceService = class {
         const validationDrops = [];
         const normalized = this.ontologyService?.validateAndNormalizeFact(ontologyFact, manifest, { strict: false, drops: validationDrops }) ?? { okf_type: null, edges: [] };
         for (const drop of validationDrops) diagBuffer.push(edgeDropDiagnostic(drop, { ...diagBase, factId: id }));
+        const grounding = librarianGrounding ? groundingOutcome(checkGrounding(fact.evidence, librarianCorpus, librarianGrounding), now) : null;
+        if (grounding?.diagnostic) {
+          diagBuffer.push({ ...diagBase, code: grounding.diagnostic.code, detail: { factId: id, itemIndex: validFactItemIndexes[k], reason: grounding.diagnostic.reason } });
+        }
         const factObj = {
           id,
           entity_id: entityId,
@@ -2887,7 +3098,8 @@ var MaintenanceService = class {
           last_accessed_at: null,
           access_count: 0,
           deleted_at: null,
-          okf_type: normalized.okf_type
+          okf_type: normalized.okf_type,
+          ...grounding?.trust
         };
         await this.entryRepo.upsert(factObj, tx);
         insertedFacts.push({ id, entity_id: entityId, title: fact.title, body: fact.body, tags: JSON.stringify(fact.tags) });
@@ -3003,11 +3215,14 @@ var MaintenanceService = class {
     }
     const allTasks = await this.taskRepo.findAllPending([entityId], HEAL_MAX_TASKS);
     const recentEvents = await this.eventRepo.getRecent(entityId, 20);
+    const healGrounding = this.promptService.groundingFor("heal");
     const toPromptShape = (f) => {
       const { embedding: _embedding, embedding_blob: _blob, ...rest } = f;
-      return { ...rest, tags: typeof rest.tags === "string" ? JSON.parse(rest.tags) : rest.tags };
+      const shown = healGrounding ? (({ lifecycle_status: _l, okf_verified: _v, last_verified_at: _a, last_verified_by: _b, ...r }) => r)(rest) : rest;
+      return { ...shown, tags: typeof shown.tags === "string" ? JSON.parse(shown.tags) : shown.tags };
     };
     const anchorCache = /* @__PURE__ */ new Map();
+    const corpusByPrompt = /* @__PURE__ */ new WeakMap();
     const degraded = [];
     const outcome = await runBatched({
       items: healCandidates,
@@ -3020,9 +3235,10 @@ var MaintenanceService = class {
           // not overfetch the same 200 keyword hits a 25-fact batch once did.
           // buildHealPrompt applies the matching cap on its side — values must match.
           Math.min(HEAL_MAX_ANCHORS, batch.length * HEAL_ANCHORS_PER_CANDIDATE),
-          anchorCache
+          anchorCache,
+          healGrounding !== null
         );
-        const { prompts, degraded: batchDegraded } = await this.promptService.buildHealPrompt(
+        const { prompts, degraded: batchDegraded, groundingCorpus } = await this.promptService.buildHealPrompt(
           batch.map(toPromptShape),
           documentAnchors,
           allTasks,
@@ -3032,16 +3248,18 @@ var MaintenanceService = class {
           bodyTruncationChars
         );
         degraded.push(...batchDegraded);
+        if (groundingCorpus !== void 0) corpusByPrompt.set(prompts, groundingCorpus);
         return prompts;
       },
       call: (prompts) => this.options.llmProvider.generateText(prompts),
-      parse: (responseText, batch) => {
+      parse: (responseText, batch, prompts) => {
         const result = parseJsonResponse(responseText);
         return {
           batch,
           downgraded: Array.isArray(result.downgraded) ? result.downgraded : [],
           deleted: Array.isArray(result.deleted) ? result.deleted : [],
-          newFacts: Array.isArray(result.newFacts) ? result.newFacts : []
+          newFacts: Array.isArray(result.newFacts) ? result.newFacts : [],
+          corpus: corpusByPrompt.get(prompts) ?? null
         };
       },
       maxOutputTokens: this.options.llmProvider.maxOutputTokens,
@@ -3056,6 +3274,7 @@ var MaintenanceService = class {
     const safeDeletedSet = /* @__PURE__ */ new Set();
     const newFacts = [];
     const newFactItemIndexes = [];
+    const newFactCorpora = [];
     for (const batchResult of outcome.results) {
       const mutableIds = new Set(batchResult.batch.map((f) => f.id));
       for (const id of batchResult.downgraded) if (mutableIds.has(id)) safeDowngradedSet.add(id);
@@ -3063,6 +3282,7 @@ var MaintenanceService = class {
       batchResult.newFacts.forEach((raw, itemIndex) => {
         newFacts.push(raw);
         newFactItemIndexes.push(itemIndex);
+        newFactCorpora.push(batchResult.corpus);
       });
     }
     const safeDowngraded = Array.from(safeDowngradedSet);
@@ -3070,12 +3290,14 @@ var MaintenanceService = class {
     const diagBuffer = new DiagnosticBuffer();
     const validNewFacts = [];
     const validNewFactItemIndexes = [];
+    const validNewFactCorpora = [];
     newFacts.forEach((raw, k) => {
       const itemIndex = newFactItemIndexes[k];
       const valid = validateFact(raw);
       if (valid) {
         validNewFacts.push(valid);
         validNewFactItemIndexes.push(itemIndex);
+        validNewFactCorpora.push(newFactCorpora[k]);
       } else {
         diagBuffer.push({ ...diagBase, code: "fact_rejected", detail: { itemIndex, reason: factRejectionReason(raw) } });
       }
@@ -3105,6 +3327,10 @@ var MaintenanceService = class {
           continue;
         }
         const id = generateId("fact_");
+        const grounding = healGrounding ? groundingOutcome(checkGrounding(fact.evidence, validNewFactCorpora[k] ?? [], healGrounding), now) : null;
+        if (grounding?.diagnostic) {
+          diagBuffer.push({ ...diagBase, code: grounding.diagnostic.code, detail: { factId: id, itemIndex: validNewFactItemIndexes[k], reason: grounding.diagnostic.reason } });
+        }
         const factObj = {
           id,
           entity_id: entityId,
@@ -3119,7 +3345,8 @@ var MaintenanceService = class {
           updated_at: now,
           last_accessed_at: null,
           access_count: 0,
-          deleted_at: null
+          deleted_at: null,
+          ...grounding?.trust
         };
         await this.entryRepo.upsert(factObj, tx);
         insertedFacts.push({ id, entity_id: entityId, title: fact.title, body: fact.body, tags: JSON.stringify(fact.tags) });
@@ -3191,7 +3418,7 @@ var MaintenanceService = class {
     if (!ontologyService) {
       return { ...zeroed, remaining: 0, deferred: 0 };
     }
-    const { mode } = await ontologyService.getEffectiveState(entityId);
+    const { mode, manifest: effectiveManifest } = await ontologyService.getEffectiveState(entityId);
     if (mode === "off") {
       const counts2 = await this.entryRepo.countUntypedByEntityId(entityId, recheckCutoff);
       return { ...zeroed, remaining: 0, deferred: counts2.deferred };
@@ -3201,6 +3428,14 @@ var MaintenanceService = class {
       const counts2 = await this.entryRepo.countUntypedByEntityId(entityId, recheckCutoff);
       return { ...zeroed, remaining: counts2.eligible, deferred: counts2.deferred };
     }
+    const classifierMode = options?.classifier ?? this.options.config?.ontology?.backfillClassifier ?? "llm";
+    const classify = this.options.llmProvider.classify;
+    if (classifierMode === "auto" && typeof classify === "function") {
+      const slugs = effectiveManifest.node_types.map((n) => n.type);
+      if (slugs.length > 0 && slugs.length <= 255) {
+        return this._runClassifierBackfill(entityId, candidates, effectiveManifest, now, recheckCutoff);
+      }
+    }
     const ontologyContext = await ontologyService.buildPromptContext(entityId);
     const toPromptShape = (f) => ({ id: f.id, title: f.title, body: f.body, tags: f.tags });
     const buildPrompt = (facts) => this.promptService.buildOntologyBackfillPrompt(
@@ -3279,6 +3514,87 @@ var MaintenanceService = class {
       deferred: counts.deferred
     };
   }
+  /**
+   * Classifier-mode backfill (spec §7.3): one `choice` question per untyped
+   * fact over the effective manifest's node-type slugs. Accepted answers go
+   * through `_applyOntologyBackfillBatch` exactly like LLM classifications.
+   * No edges are proposed.
+   */
+  async _runClassifierBackfill(entityId, candidates, manifest, now, recheckCutoff) {
+    const provider = this.options.llmProvider;
+    const rawMin = this.options.config?.ontology?.classifyMinConfidence;
+    const minConfidence = typeof rawMin === "number" && Number.isFinite(rawMin) && rawMin >= 0 && rawMin <= 1 ? rawMin : 0.5;
+    const rawConcurrency = this.options.config?.chunkConcurrency ?? 1;
+    const concurrency = Number.isFinite(rawConcurrency) && rawConcurrency >= 1 ? Math.floor(rawConcurrency) : 1;
+    const question = {
+      kind: "choice",
+      options: manifest.node_types.map((n) => n.type),
+      instructions: "Choose the ontology node type that best describes this fact.\n" + manifest.node_types.map((n) => `- ${n.type}: ${n.description}`).join("\n")
+    };
+    const outcomes = await withConcurrency(
+      candidates.map((fact) => async () => {
+        let response;
+        try {
+          response = await provider.classify({ state: classifierStateForFact(fact), questions: { okf_type: question } });
+        } catch {
+          return { fact, kind: "threw" };
+        }
+        const checked = validateClassifierAnswer(response, "okf_type", question);
+        if (!checked.ok) return { fact, kind: "invalid", reason: checked.reason };
+        if (checked.answer.kind !== "choice") return { fact, kind: "invalid", reason: "kind_mismatch" };
+        if (checked.answer.confidence < minConfidence) return { fact, kind: "low_confidence" };
+        return { fact, kind: "accepted", okfType: checked.answer.choice };
+      }),
+      concurrency
+    );
+    const attempted = outcomes.filter((o) => o.kind !== "threw").map((o) => o.fact);
+    const skipped = outcomes.length - attempted.length;
+    const classifications = outcomes.filter((o) => o.kind === "accepted").map((o) => ({ id: o.fact.id, okf_type: o.okfType }));
+    const invalidCount = outcomes.filter((o) => o.kind === "invalid").length;
+    let typed = 0;
+    let failedValidation = 0;
+    let edgesAdded = 0;
+    let aborted = false;
+    if (attempted.length > 0) {
+      const applied = await this._applyOntologyBackfillBatch(
+        entityId,
+        { batch: attempted, classifications, ontologyUpdates: void 0 },
+        now
+      );
+      aborted = applied.abortedOntologyOff;
+      if (!aborted) {
+        typed = applied.typed;
+        failedValidation = invalidCount + applied.failedValidation;
+        edgesAdded = applied.edgesAdded;
+      }
+    }
+    const diagBuffer = new DiagnosticBuffer();
+    const diagBase = { entityId, operation: "ontologyBackfill", trigger: "call" };
+    for (const o of outcomes) {
+      if (o.kind === "low_confidence") {
+        diagBuffer.push({ ...diagBase, code: "classification_low_confidence", detail: { factId: o.fact.id, reason: "below_threshold" } });
+      } else if (o.kind === "invalid") {
+        diagBuffer.push({ ...diagBase, code: "classification_invalid", detail: { factId: o.fact.id, reason: o.reason } });
+      } else if (o.kind === "threw") {
+        diagBuffer.push({ ...diagBase, code: "classification_invalid", detail: { factId: o.fact.id, reason: "classify_threw" } });
+      }
+    }
+    diagBuffer.flush(this.options);
+    const counts = await this.entryRepo.countUntypedByEntityId(entityId, recheckCutoff);
+    if (aborted) {
+      return { scanned: 0, typed: 0, failedValidation: 0, edgesAdded: 0, skipped, remaining: 0, deferred: counts.deferred };
+    }
+    this.searchService.evictCache(entityId);
+    return {
+      scanned: candidates.length,
+      typed,
+      failedValidation,
+      edgesAdded,
+      skipped,
+      remaining: counts.eligible,
+      deferred: counts.deferred
+    };
+  }
   /**
    * Applies one parsed backfill batch in its own transaction. Per-batch rather
    * than one transaction for the pass, so mergeEmergentUpdates semantics and
@@ -3394,7 +3710,7 @@ var MaintenanceService = class {
    * batches that reduce to the same query share one lookup. Caller-owned and
    * per-pass — see the call site in doRunHeal.
    */
-  async _selectHealAnchors(entityId, batch, cap = HEAL_MAX_ANCHORS, cache) {
+  async _selectHealAnchors(entityId, batch, cap = HEAL_MAX_ANCHORS, cache, withBodies = false) {
     const query = batch.map((f) => f.title).join(" ").trim();
     if (!query) return [];
     const cached = cache?.get(query);
@@ -3407,7 +3723,7 @@ var MaintenanceService = class {
     const hitIds = hits.map((h) => h.id);
     const anchors = [];
     if (hitIds.length > 0) {
-      const rows = await this.entryRepo.findAnchorRowsByIds(entityId, hitIds);
+      const rows = withBodies ? await this.entryRepo.findAnchorRowsByIds(entityId, hitIds, void 0, { withBody: true }) : await this.entryRepo.findAnchorRowsByIds(entityId, hitIds);
       const byId = new Map(rows.map((r) => [r.id, r]));
       for (const id of hitIds) {
         const row = byId.get(id);
@@ -3563,7 +3879,6 @@ var RetrievalService = class {
     const hybridWeight = options?.hybridWeight ?? config?.hybridWeight;
     const weight = hybridWeight !== void 0 && !Number.isNaN(hybridWeight) ? Math.max(0, Math.min(1, hybridWeight)) : void 0;
     const skipEmbed = weight === 0;
-    const embedFn = this.options.llmProvider.embed;
     let facts = [];
     let scoreByFactId;
     if (maxResults === 0) ; else if (trimmedQuery) {
@@ -3574,11 +3889,11 @@ var RetrievalService = class {
       const padLimit = (n) => n >= Number.MAX_SAFE_INTEGER ? n : n + draftPad;
       if (scoredEntityIds.length === 0) {
         usedEmbed = true;
-      } else if (!skipEmbed && embedFn) {
+      } else if (!skipEmbed && typeof this.options.llmProvider.embed === "function") {
         let rankerShouldRethrow = false;
         let pendingRankerFallbackError;
         try {
-          const queryVec = await embedFn(trimmedQuery);
+          const queryVec = await this.options.llmProvider.embed(trimmedQuery);
           if (queryVec.length === 0 || !queryVec.every((v) => typeof v === "number" && isFinite(v))) {
             throw new Error(
               "embed() returned an empty or non-finite vector. Falling back to keyword search."