@claritylabs/cl-sdk 3.2.9 → 3.2.10

package/README.md

@@ -37,7 +37,6 @@ const extractor = createExtractor({
     return { object: result.object, usage: result.usage };
   },
   concurrency: 3,
-  pageMapConcurrency: 3,
   extractorConcurrency: 4,
   formatConcurrency: 2,
   reviewMode: "auto",
@@ -81,7 +80,7 @@ const extractor = createExtractor({ generateText, generateObject, sourceStore })
 const result = await extractor.extract(pdfBase64, "policy-123", { sourceSpans });
 ```
 
-When source spans are available, extraction returns `sourceTree`, `sourceSpans`, `operationalProfile`, `warnings`, `tokenUsage`, and `performanceReport`. The source tree is canonical for policy wording and hierarchy. The extractor uses form/page-range hints to group declarations, policy forms, and endorsements, promotes title elements into section/schedule nodes that can span page breaks, and runs a bounded organizer over small batches of existing node IDs so itemized declarations and embedded later section headings can be lifted out of over-eager table or previous-section nesting. The organizer keeps generic table rows out of model context unless the row carries structural markers such as declaration item labels, section headings, premium rows, aggregate rows, or endorsement references. The operational profile contains model-extracted policy metadata, parties, coverage units, nested coverage limit terms, deductibles/retentions, premiums, key dates, and endorsement-support facts, each with `sourceNodeIds` and `sourceSpanIds`. Source-backed identity records such as insurer, producer, contacts, named insured rows, endorsement parties, and standalone insured addresses must include non-empty `sourceSpanIds`. After extraction, a bounded model cleanup pass may keep, drop, or update existing coverage rows and terms by index; the SDK enforces the schema and rejects source IDs that are not present in the current source tree/spans. Endorsement schedules are modeled as whole endorsement coverage units with their own `limits[]` terms instead of unrelated flat rows like `Aggregate Limit`.
+When source spans are available, extraction returns `sourceTree`, `sourceSpans`, `operationalProfile`, `warnings`, `tokenUsage`, and `performanceReport`. The source tree is canonical for policy wording and hierarchy. The extractor groups parser-provided page, title, table, row, and text spans into declarations, policy forms, endorsements, sections, and schedules without running a separate form-inventory, page-map, or source-tree-organizer model call. The operational profile model receives a bounded packet of high-value declaration, schedule, premium, producer, endorsement, and coverage evidence rather than the full document wording. The operational profile contains model-extracted policy metadata, parties, coverage units, nested coverage limit terms, deductibles/retentions, premiums, key dates, and endorsement-support facts, each with `sourceNodeIds` and `sourceSpanIds`. Source-backed identity records such as insurer, producer, contacts, named insured rows, endorsement parties, and standalone insured addresses must include non-empty `sourceSpanIds`. After extraction, a bounded model cleanup pass may keep, drop, or update existing coverage rows and terms by index; the SDK enforces the schema and rejects source IDs that are not present in the current source tree/spans. Endorsement schedules are modeled as whole endorsement coverage units with their own `limits[]` terms instead of unrelated flat rows like `Aggregate Limit`.
 
 Store `result.sourceTree` in a retrievable node index and embed node `description` values for search. Keep `result.sourceSpans` as the exact PDF highlighting layer. `result.document` and its `documentOutline` remain compatibility projections for existing host screens; do not treat broad structured policy JSON as canonical extraction truth.
 
@@ -140,17 +139,16 @@ If your callback ignores those fields, the model will only see the text prompt.
 
 ## Model routing metadata
 
-Every SDK model callback may receive `taskKind`, `budgetDiagnostics`, and `trace`. Hosts can use these provider-agnostic fields for cheap-first routing, fallback, and telemetry without the SDK hardcoding model names. Example task kinds include `extraction_classify`, `extraction_focused`, `extraction_review`, `query_reason`, `application_extract_fields`, and `pce_impact_analysis`. `budgetDiagnostics` includes the resolved output-token budget, preferred task budget, truncation-risk warnings, and any model/provider caps that constrained the current subtask. `maxOutputTokens` is treated as a ceiling, not a default request size, so small organizer and classifier calls stay small even on models with large output windows. `trace` identifies the current extractor, page range, format batch, or source-backed call so host logs can show what was being generated instead of a generic model-call label.
+Every SDK model callback may receive `taskKind`, `budgetDiagnostics`, and `trace`. Hosts can use these provider-agnostic fields for cheap-first routing, fallback, and telemetry without the SDK hardcoding model names. Example task kinds include `extraction_operational_profile`, `extraction_coverage_cleanup`, `query_reason`, `application_extract_fields`, and `pce_impact_analysis`. `budgetDiagnostics` includes the resolved output-token budget, preferred task budget, truncation-risk warnings, and any model/provider caps that constrained the current subtask. `maxOutputTokens` is treated as a ceiling, not a default request size, so small classifier and lookup calls stay small even on models with large output windows. `trace` identifies the current source-backed or workflow call so host logs can show what was being generated instead of a generic model-call label.
 
 ## Bounded Agentic Workflows
 
 CL-SDK uses deterministic scaffolding with agentic decision points rather than fixed all-tools-all-the-time chains:
 
-- Extraction page mapping and review choose focused follow-up extractors from the live extractor catalog. Definitions and covered reasons can fall back through section extraction when a focused run returns no usable records.
-- Supplementary extraction runs only when page assignments, form inventory, existing extracted text, or review follow-up tasks indicate regulatory, claims, notice, cancellation, or contact facts are likely present.
-- Referential coverage resolution tries cheap local section/form matches first, then uses bounded target-specific actions for declarations, schedules, sections, page-location lookup, or skip.
-- Page mapping, focused extractors, referential lookup, and formatting use separate concurrency controls. Page-scoped PDF and image ranges are cached so overlapping extractor tasks do not repeatedly slice or render the same pages.
-- Formatting skips the LLM cleanup pass for plain prose and formats long or noisy markdown/table/list content in parallel batches.
+- Policy extraction requires caller-provided source spans and does not run the older form-inventory, page-map, focused-extractor, or source-tree-organizer model calls.
+- Source-tree construction stays deterministic from parser spans; the model is used for the operational profile and optional coverage cleanup because those are end-user facts.
+- Referential coverage resolution and application/query workflows still use bounded target-specific actions when those workflows need lookup or explanation steps.
+- Formatting skips the LLM cleanup pass for plain prose and formats long or noisy markdown/table/list content in parallel batches outside the v3 source-span extraction path.
 - `reviewMode: "auto"` skips the expensive LLM review pass when deterministic checks are clean and source spans are available. Use `"always"` for maximum review coverage or `"skip"` when the host owns quality review separately.
 - Application processing plans optional backfill, context auto-fill, document search, batching, reply parsing, lookup, explanations, and next-batch email generation based on current active question state. Conditional fields that are not active are skipped until their parent answers trigger them.
 

package/dist/index.js

@@ -4583,36 +4583,6 @@ function spanPageEnd(span) {
 function spanSourceUnit(span) {
   return span.sourceUnit ?? span.metadata?.sourceUnit ?? span.metadata?.elementType;
 }
-function formNumberFromText(value) {
-  return cleanText(value, "").match(/\b[A-Z]{2,}(?:-[A-Z0-9]+)+\s+\d{2}\s+\d{2}\b/)?.[0]?.replace(/\s+/g, " ");
-}
-function editionDateFromFormNumber(formNumber) {
-  const match = formNumber?.match(/\b(\d{2})\s+(\d{2})$/);
-  return match ? `${match[1]}/${match[2]}` : void 0;
-}
-function pageTitleFromText(text, fallback) {
-  const normalized = cleanText(text, fallback);
-  const patterns = [
-    /\bIMPORTANT NOTICE\s+[—-]\s+HOW TO REPORT A CLAIM\b/i,
-    /\bPRIVACY NOTICE TO POLICYHOLDERS\b/i,
-    /\bOFAC ADVISORY NOTICE\b/i,
-    /\bTERRORISM RISK INSURANCE ACT\s*\(TRIA\)\s*DISCLOSURE AND REJECTION\b/i,
-    /\bDECLARATIONS PAGE\b/i,
-    /\bTECHNOLOGY ERRORS?\s*&\s*OMISSIONS AND CYBER LIABILITY INSURANCE POLICY\b/i,
-    /\bTRADE OR ECONOMIC SANCTIONS LIMITATION\b/i,
-    /\bFORMS? AND ENDORSEMENTS\b/i
-  ];
-  for (const pattern of patterns) {
-    const match = normalized.match(pattern)?.[0];
-    if (match) return cleanText(match, fallback);
-  }
-  const endorsement = normalized.match(/\bENDORSEMENT\s+(?:NO\.?|NUMBER|#)\s*[A-Z0-9][A-Z0-9.-]*\b/i)?.[0];
-  if (endorsement) return cleanText(endorsement, fallback);
-  const firstSentence = normalized.split(/(?<=\.)\s+/)[0];
-  if (/^page\s+\d+\b/i.test(firstSentence)) return fallback;
-  if (firstSentence && firstSentence.length <= 120) return firstSentence.replace(/[.]$/, "");
-  return fallback;
-}
 function pageHeadingTitleFromText(text, fallback) {
   const normalized = cleanText(text, "");
   const headingText = normalized.replace(/^page\s+\d+\s*(?:\|\s*page\s*\|\s*page\s+\d+\s*\|?)?/i, "").slice(0, 700);
@@ -4632,194 +4602,9 @@ function pageHeadingTitleFromText(text, fallback) {
   }
   return fallback;
 }
-function pageFormTypeFromText(text) {
-  if (hasSubstantiveDeclarationsScheduleText(text)) return "declarations";
-  if (/\b(declarations?\s+page|declarations?\s+schedule)\b/i.test(text)) return "declarations";
-  if (/\b(endorsement\s+(?:no\.?|number|#)|this endorsement changes the policy|[A-Z]{2,}-END\s+\d{2,})\b/i.test(text)) return "endorsement";
-  if (/\b(technology errors?\s*&?\s*omissions.*liability insurance policy|policy form|coverage form|insuring agreement|definitions?|exclusions?|conditions?)\b/i.test(text)) return "coverage";
-  if (/\b(important notice|privacy notice|ofac advisory|terrorism risk insurance act|tria|trade or economic sanctions)\b/i.test(text)) return "notice";
-  return "other";
-}
 function hasSubstantiveDeclarationsScheduleText(text) {
   return /\bitem\s+\d+\.?\s*(?:named insured|policy number|policy period|renewal|form of business|coverage parts?|limits?|premium|extended reporting|producer|forms? and endorsements?)\b/i.test(text) || /\bforms? and endorsements attached at inception\b/i.test(text) || /\bcoverage parts?,?\s+limits? of liability,?\s+deductibles?,?\s+and retroactive dates\b/i.test(text) || /\bannual premium\s*\(all coverage parts?\)\b/i.test(text) || /\berp option\b/i.test(text) || /\bproducer\b[\s\S]{0,240}\blicense\b/i.test(text);
 }
-function administrativeFormTypeFromText(text) {
-  if (hasSubstantiveDeclarationsScheduleText(text)) return void 0;
-  if (/\b(important notice|privacy notice|ofac advisory|terrorism risk insurance act|tria|trade or economic sanctions|economic sanctions limitation|how to report a claim)\b/i.test(text)) {
-    return "notice";
-  }
-  if (/\b(specimen policy|policy jacket|countersigned|countersignature|licensed resident agent|corporate secretary|president and ceo|application of insurance executed)\b/i.test(text)) {
-    return "other";
-  }
-  return void 0;
-}
-function pageTextByNumber(sourceSpans) {
-  const pageTexts = /* @__PURE__ */ new Map();
-  const pageSpanTexts = /* @__PURE__ */ new Map();
-  for (const span of sourceSpans) {
-    const start = spanPageStart(span);
-    if (typeof start !== "number") continue;
-    const end = spanPageEnd(span) ?? start;
-    for (let page = start; page <= end; page += 1) {
-      if (spanSourceUnit(span) === "page") {
-        pageSpanTexts.set(page, cleanText(span.text, ""));
-        continue;
-      }
-      const existing = pageTexts.get(page) ?? "";
-      if (existing.length < 4e3) pageTexts.set(page, cleanText([existing, span.text].filter(Boolean).join(" "), ""));
-    }
-  }
-  return new Map([.../* @__PURE__ */ new Set([...pageTexts.keys(), ...pageSpanTexts.keys()])].map((page) => [
-    page,
-    pageSpanTexts.get(page) ?? pageTexts.get(page) ?? ""
-  ]));
-}
-function reconcileFormTypeWithSourceText(form, pageTexts) {
-  const pages = [];
-  const start = form.pageStart;
-  const end = form.pageEnd ?? start;
-  if (typeof start === "number" && typeof end === "number") {
-    for (let page = start; page <= end; page += 1) pages.push(pageTexts.get(page) ?? "");
-  }
-  const text = cleanText([form.title, form.formNumber, ...pages].filter(Boolean).join(" "), "");
-  const administrativeType = administrativeFormTypeFromText(text);
-  if (administrativeType && form.formType !== administrativeType) return administrativeType;
-  return form.formType;
-}
-function inferFormHintsFromSourceSpans(sourceSpans) {
-  const pageTexts = pageTextByNumber(sourceSpans);
-  if (pageTexts.size === 0) return [];
-  const pageHints = [...pageTexts.keys()].sort((left, right) => left - right).map((page) => {
-    const text = pageTexts.get(page) ?? "";
-    const formNumber = formNumberFromText(text);
-    return {
-      formNumber,
-      editionDate: editionDateFromFormNumber(formNumber),
-      title: pageTitleFromText(text, `Page ${page}`),
-      formType: pageFormTypeFromText(text),
-      pageStart: page,
-      pageEnd: page
-    };
-  });
-  const merged = [];
-  for (const hint of pageHints) {
-    const previous = merged[merged.length - 1];
-    const startsNewEndorsement = hint.formType === "endorsement" && /\bendorsement\s+(?:no\.?|number|#)\s*[A-Z0-9]|this endorsement changes the policy/i.test(hint.title ?? "");
-    const canMerge = previous && previous.formType === hint.formType && previous.pageEnd !== void 0 && hint.pageStart === previous.pageEnd + 1 && (hint.formType === "declarations" || hint.formType === "coverage" || hint.formType === "endorsement" && !startsNewEndorsement);
-    if (!canMerge) {
-      merged.push(hint);
-      continue;
-    }
-    previous.pageEnd = hint.pageEnd;
-    previous.title = previous.title ?? hint.title;
-    previous.formNumber = previous.formNumber ?? hint.formNumber;
-    previous.editionDate = previous.editionDate ?? hint.editionDate;
-  }
-  return merged;
-}
-function normalizeFormHints(forms, sourceSpans) {
-  const pageTexts = pageTextByNumber(sourceSpans);
-  const provided = (forms ?? []).filter(
-    (form) => typeof form.pageStart === "number" && typeof form.pageEnd === "number" && form.pageStart > 0 && form.pageEnd >= form.pageStart
-  ).map((form) => ({
-    ...form,
-    formType: reconcileFormTypeWithSourceText(form, pageTexts),
-    title: form.title ? cleanText(form.title, "") : void 0
-  })).sort(
-    (left, right) => (left.pageStart ?? Number.MAX_SAFE_INTEGER) - (right.pageStart ?? Number.MAX_SAFE_INTEGER) || (left.pageEnd ?? Number.MAX_SAFE_INTEGER) - (right.pageEnd ?? Number.MAX_SAFE_INTEGER)
-  );
-  return provided.length ? provided : inferFormHintsFromSourceSpans(sourceSpans);
-}
-function formHintForPage(forms, page) {
-  return forms.find(
-    (form) => typeof form.pageStart === "number" && typeof form.pageEnd === "number" && page >= form.pageStart && page <= form.pageEnd
-  );
-}
-function titleFromFormHint(form, fallback) {
-  if (form.formType === "declarations") return "Declarations";
-  if (form.formType === "coverage") return "Policy Form";
-  if (form.formType === "endorsement") return endorsementTitle([form.title, form.formNumber].filter(Boolean).join(" ")) ?? cleanText(form.title, fallback);
-  return cleanText(form.title, fallback);
-}
-function formGroupConfig(form) {
-  if (form.formType === "declarations") {
-    return {
-      kind: "page_group",
-      title: "Declarations",
-      description: "Declarations pages and schedules grouped from form inventory",
-      organizer: "form_inventory_declarations_grouping"
-    };
-  }
-  if (form.formType === "coverage") {
-    return {
-      kind: "form",
-      title: "Policy Form",
-      description: "Policy form pages grouped from form inventory",
-      organizer: "form_inventory_policy_form_grouping"
-    };
-  }
-  if (form.formType === "endorsement") {
-    const title = titleFromFormHint(form, "Endorsement");
-    return {
-      kind: "endorsement",
-      title,
-      description: `${title} grouped from form inventory`,
-      organizer: "form_inventory_endorsement_grouping"
-    };
-  }
-  return void 0;
-}
-function applyFormInventoryHints(sourceTree, forms) {
-  if (forms.length === 0) return sourceTree;
-  const rootId = sourceTreeRootId(sourceTree);
-  if (!rootId) return sourceTree;
-  const byParent = nodesByParent(sourceTree);
-  const children = (byParent.get(rootId) ?? []).filter((node) => node.kind !== "document").sort((left, right) => left.order - right.order);
-  const rootPages = children.filter((node) => node.kind === "page" && typeof node.pageStart === "number");
-  let nextTree = sourceTree.map((node) => {
-    if (node.kind !== "page" || typeof node.pageStart !== "number") return node;
-    const form = formHintForPage(forms, node.pageStart);
-    if (!form) return node;
-    const isStartPage = form.pageStart === node.pageStart;
-    const shouldRetitle = isStartPage && form.formType !== "other";
-    return {
-      ...node,
-      title: shouldRetitle ? titleFromFormHint(form, node.title) : node.title,
-      metadata: {
-        ...node.metadata,
-        formInventoryHint: {
-          formType: form.formType,
-          formNumber: form.formNumber,
-          title: form.title,
-          pageStart: form.pageStart,
-          pageEnd: form.pageEnd
-        }
-      }
-    };
-  });
-  const claimed = /* @__PURE__ */ new Set();
-  for (const form of forms) {
-    const config = formGroupConfig(form);
-    const pageStart2 = form.pageStart;
-    const pageEnd2 = form.pageEnd;
-    if (!config || typeof pageStart2 !== "number" || typeof pageEnd2 !== "number") continue;
-    const childIds = rootPages.filter(
-      (page) => !claimed.has(page.id) && typeof page.pageStart === "number" && page.pageStart >= pageStart2 && page.pageStart <= pageEnd2
-    ).map((page) => page.id);
-    if (childIds.length === 0) continue;
-    nextTree = groupAdjacentChildren({
-      sourceTree: nextTree,
-      children,
-      childIds,
-      kind: config.kind,
-      title: config.title,
-      description: config.description,
-      organizer: config.organizer
-    });
-    childIds.forEach((id) => claimed.add(id));
-  }
-  return normalizeDocumentSourceTreePaths(nextTree);
-}
 function looksLikeDeclarationsStart(node) {
   const title = cleanText(node.title, "");
   const text = sourceNodeText(node);
@@ -4830,13 +4615,13 @@ function looksLikeDeclarationsStart(node) {
 }
 function looksLikeDeclarationsContinuation(node) {
   const text = sourceNodeText(node);
-  return looksLikeDeclarationsStart(node) || /\b(item\s+\d+\.|coverage part|each claim limit|aggregate limit|retroactive date|self-insured retention|premium|payment plan|producer|broker|forms? and endorsements?|extended reporting period|discovery period)\b/i.test(text);
+  return looksLikeDeclarationsStart(node) || /\b(item\s+\d+\.|coverage part|limits?,?\s+sub-limits?|each claim limit|aggregate limit|retroactive date|self-insured retention|premium|payment plan|producer|broker|forms? and endorsements?|attached at inception|extended reporting period|discovery period)\b/i.test(text) || /\b(these declarations|policy form|[A-Z]{2,}-END\s+\d{3}|endorsement\s+(?:no\.?|number|#)?\s*\d+)\b/i.test(text);
 }
 function looksLikePolicyFormStart(node) {
   const text = sourceNodeText(node);
   const excerpt = cleanText(node.textExcerpt, "");
   if (isAdministrativeNoticeNode(node) || looksLikeDeclarationsStart(node)) return false;
-  return /\bpolicy form\b/i.test(node.title) || /^policy\s+form\b/i.test(excerpt) || /\btechnology errors?\s*&?\s*omissions\b/i.test(text) && /\bplease read this entire policy carefully\b/i.test(text) || /\bform\s+[A-Z]{2,}-[A-Z0-9-]+\s+\d{2}\s+\d{2}\b/i.test(text);
+  return /\bpolicy form\b/i.test(node.title) || /^policy\s+form\b/i.test(excerpt) || /\btechnology errors?\s*&?\s*omissions\b/i.test(text) && /\bplease read this entire policy carefully\b/i.test(text) || /\bsection\s+[IVX0-9]+\s*[—-]\s*(insuring agreements?|definitions?|exclusions?|conditions?)\b/i.test(text) || /\bform\s+[A-Z]{2,}-[A-Z0-9-]+\s+\d{2}\s+\d{2}\b/i.test(text);
 }
 function looksLikePolicyFormContinuation(node) {
   const text = sourceNodeText(node);
@@ -5598,25 +5383,51 @@ function isOperationalEvidenceAnchor(span) {
   if (/\$[\d,.]+|[0-9]{1,2}\/[0-9]{1,2}\/[0-9]{2,4}|[0-9]{1,2}\s+[A-Za-z]{3,9}\s+[0-9]{4}/.test(text)) return true;
   return false;
 }
+function operationalEvidencePages(sourceSpans) {
+  const scores = /* @__PURE__ */ new Map();
+  for (const span of sourceSpans) {
+    const page = spanPageStart(span);
+    if (typeof page !== "number") continue;
+    const text = cleanText([span.text, span.formNumber, spanSourceUnit(span)].filter(Boolean).join(" "), "");
+    if (!text) continue;
+    let score = Math.max(0, operationalEvidenceScore(span));
+    if (/\bdeclarations?\s+(page|schedule)?\b/i.test(text)) score += 24;
+    if (/\bitem\s+\d+\.?\s*(?:named insured|policy number|policy period|coverage parts?|limits?|premium|producer|forms? and endorsements?)\b/i.test(text)) score += 20;
+    if (/\bcoverage parts?,?\s+limits? of liability,?\s+deductibles?,?\s+and retroactive dates\b/i.test(text)) score += 24;
+    if (/\b(policy\s*(number|period|term)|effective date|expiration date|expiry date|named insured|insurer|carrier|security)\b/i.test(text)) score += 12;
+    if (/\b(premium|total due|tax|fee|producer|broker)\b/i.test(text)) score += 10;
+    if (/\b(endorsement\s+(?:no\.?|number|#)?\s*[A-Z0-9]|attached at inception|forms? and endorsements?)\b/i.test(text)) score += 8;
+    if (/\b(definitions?|exclusions?|conditions?|duties in the event|action against|cancellation by)\b/i.test(text)) score -= 12;
+    if (score > 0) scores.set(page, (scores.get(page) ?? 0) + score);
+  }
+  return new Set(
+    [...scores.entries()].filter(([, score]) => score >= 18).sort((left, right) => right[1] - left[1] || left[0] - right[0]).slice(0, 12).map(([page]) => page)
+  );
+}
 function operationalProfileEvidence(sourceTree, sourceSpans) {
   const sorted = [...sourceSpans].sort(
     (left, right) => (spanPageStart(left) ?? Number.MAX_SAFE_INTEGER) - (spanPageStart(right) ?? Number.MAX_SAFE_INTEGER) || (left.location?.charStart ?? Number.MAX_SAFE_INTEGER) - (right.location?.charStart ?? Number.MAX_SAFE_INTEGER) || left.id.localeCompare(right.id)
   );
+  const selectedPages = operationalEvidencePages(sorted);
   const selected = /* @__PURE__ */ new Set();
   const selectedTableIds = /* @__PURE__ */ new Set();
   for (let index = 0; index < sorted.length; index += 1) {
-    const score = operationalEvidenceScore(sorted[index]);
+    const span = sorted[index];
+    const score = operationalEvidenceScore(span);
     if (score < 8) continue;
-    if (!isOperationalEvidenceAnchor(sorted[index])) continue;
-    const tableId2 = spanTableId(sorted[index]);
-    if (tableId2 && !isTableCellSpan(sorted[index])) selectedTableIds.add(tableId2);
-    const page = spanPageStart(sorted[index]);
-    for (let offset = -2; offset <= 2; offset += 1) {
+    if (!isOperationalEvidenceAnchor(span)) continue;
+    const page = spanPageStart(span);
+    if (selectedPages.size > 0 && typeof page === "number" && !selectedPages.has(page) && score < 30) continue;
+    const tableId2 = spanTableId(span);
+    if (tableId2 && !isTableCellSpan(span)) selectedTableIds.add(tableId2);
+    const neighborWindow = score >= 24 ? 2 : 1;
+    for (let offset = -neighborWindow; offset <= neighborWindow; offset += 1) {
       const neighborIndex = index + offset;
       const neighbor = sorted[neighborIndex];
       if (!neighbor || spanPageStart(neighbor) !== page) continue;
+      if (selectedPages.size > 0 && typeof page === "number" && !selectedPages.has(page) && operationalEvidenceScore(neighbor) < 30) continue;
       const neighborText = cleanText(neighbor.text, "");
-      if (!neighborText || neighborText.length > 5e3) continue;
+      if (!neighborText || neighborText.length > 3e3) continue;
       selected.add(neighborIndex);
     }
   }
@@ -5625,7 +5436,7 @@ function operationalProfileEvidence(sourceTree, sourceSpans) {
       const tableId2 = spanTableId(span);
       if (!tableId2 || !selectedTableIds.has(tableId2) || isTableCellSpan(span)) return;
       const text = cleanText(span.text, "");
-      if (text && text.length <= 5e3) selected.add(index);
+      if (text && text.length <= 3e3) selected.add(index);
     });
   }
   const nodeIdsBySpanId = sourceNodeIdsBySpanId(sourceTree);
@@ -5643,12 +5454,12 @@ function operationalProfileEvidence(sourceTree, sourceSpans) {
       pageEnd: spanPageEnd(span),
       sourceUnit: spanSourceUnit(span),
       formNumber: span.formNumber,
-      text: text.slice(0, span.sourceUnit === "page" ? 1200 : 900)
+      text: text.slice(0, span.sourceUnit === "page" ? 900 : 700)
     }];
   });
   const detailEntries = entries.filter((entry) => entry.sourceUnit !== "page");
   const pageEntries = entries.filter((entry) => entry.sourceUnit === "page");
-  return [...detailEntries, ...pageEntries].slice(0, 180);
+  return [...detailEntries.slice(0, 80), ...pageEntries.slice(0, 8)];
 }
 function sourceTreeRootId(sourceTree) {
   return sourceTree.find((node) => node.kind === "document")?.id;
@@ -5697,7 +5508,7 @@ Rules:
 - If a coverage schedule continues onto the next page before the next item marker, include the continuation rows in the same coverage or declaration item.
 - If one schedule row or continuation row states the same amount with multiple bases, such as "$1,000,000 Each Claim / Aggregate", return separate limit terms for each basis using the same value instead of one combined "Each Claim / Aggregate" term.
 - LiteParse text can fragment visual table cells into adjacent lines. Before extracting coverage terms, mentally join adjacent lines in the same declaration item or schedule row. For example, "$2,000,000 Policy Each Claim" followed immediately by "Aggregate" means "$2,000,000 Policy Aggregate"; a line ending with "/" followed by "Aggregate ..." means the limit cell continues, not a new coverage.
-- Forms-and-endorsements schedules are operational form inventory evidence, not coverage limits. Do not turn form schedule rows into coverage units unless the row also states a coverage-specific limit or deductible.
+- Forms-and-endorsements schedules are form schedule evidence, not coverage limits. Do not turn form schedule rows into coverage units unless the row also states a coverage-specific limit or deductible.
 - Keep each coverage unit tied to one evidence scope: a declaration/core schedule row, a core policy form section, or one specific endorsement schedule. Do not merge declaration facts and endorsement schedule facts into the same coverage unit, even when they use the same coverage name.
 - If the declarations schedule and an endorsement schedule both list Network Security, Social Engineering Fraud, Regulatory Proceedings, or another same-named coverage, return separate coverage units for each supported source scope.
 - Use the declaration coverage name for declaration/core schedule rows. Use the endorsement title or endorsement schedule coverage name for endorsement rows, and include formNumber and endorsementNumber when source-backed.
@@ -5936,8 +5747,8 @@ async function cleanupOperationalCoverageSchedules(params) {
 }
 async function runSourceTreeExtraction(params) {
   const sourceSpans = normalizeSourceSpans(params.sourceSpans);
-  const formHints = normalizeFormHints(params.formInventory?.forms, sourceSpans);
-  let sourceTree = applySemanticPageGrouping(applyFormInventoryHints(buildDocumentSourceTree(sourceSpans, params.id), formHints));
+  const formHints = [];
+  let sourceTree = applySemanticPageGrouping(buildDocumentSourceTree(sourceSpans, params.id));
   const warnings = [];
   let modelCalls = 0;
   let callsWithUsage = 0;