feat: evaluate deterministic cross-branch corroboration

2026-08-03 07:19:20 +01:00
parent 1d64144e01
commit b2ffc54964
6 changed files with 547 additions and 7 deletions
@@ -463,6 +463,129 @@ function confidenceFromAssessment(assessment) {
  return assessment?.conclusionConfidence ?? "medium";
 }

+function unique(values = []) {
+  return [...new Set(values.filter(Boolean))];
+}
+
+function branchEvidenceKeys(node) {
+  return unique([...(node?.evidenceIds || []), node?.value]);
+}
+
+function sharedMeaningfulTokens(aText, bText) {
+  const stop = new Set([
+    "the",
+    "and",
+    "for",
+    "that",
+    "this",
+    "with",
+    "from",
+    "because",
+    "need",
+    "unknown",
+    "possible",
+  ]);
+  const a = splitSemanticTokens(aText).filter((token) => !stop.has(token));
+  const b = splitSemanticTokens(bText).filter((token) => !stop.has(token));
+  return [...new Set(a.filter((token) => b.includes(token)))];
+}
+
+function branchConflictSignature(node) {
+  return normaliseText(
+    `${node?.label || ""} ${node?.description || ""} ${node?.value || ""}`,
+  );
+}
+
+function branchesConflict(aNode, bNode) {
+  const aText = branchConflictSignature(aNode);
+  const bText = branchConflictSignature(bNode);
+  const oppositePolarity =
+    (aText.includes("correctly") && bText.includes("incorrectly")) ||
+    (aText.includes("incorrectly") && bText.includes("correctly")) ||
+    aNode?.status === "contradicted" ||
+    bNode?.status === "contradicted";
+
+  if (!oppositePolarity) return false;
+
+  return sharedMeaningfulTokens(aText, bText).length >= 2;
+}
+
+export function evaluateBranchInteractions({ parentNode, graph }) {
+  const directBranches = findDirectChildUnknowns(graph, parentNode.id).filter(
+    (node) => ["resolved", "provisional", "contradicted"].includes(node.status),
+  );
+  const duplicateEvidenceGroups = [];
+  const conflictingBranches = [];
+  const corroboratingBranches = [];
+  const duplicateBranchIds = new Set();
+  const conflictingBranchIds = new Set();
+
+  const evidenceGroups = new Map();
+  for (const branch of directBranches) {
+    for (const evidenceKey of branchEvidenceKeys(branch)) {
+      const ids = evidenceGroups.get(evidenceKey) || [];
+      ids.push(branch.id);
+      evidenceGroups.set(evidenceKey, ids);
+    }
+  }
+
+  for (const [evidenceKey, branchIds] of evidenceGroups.entries()) {
+    if (branchIds.length > 1) {
+      duplicateEvidenceGroups.push({
+        evidenceKey,
+        branchIds: unique(branchIds),
+      });
+      for (const id of branchIds) duplicateBranchIds.add(id);
+    }
+  }
+
+  for (let index = 0; index < directBranches.length; index += 1) {
+    for (let inner = index + 1; inner < directBranches.length; inner += 1) {
+      const aNode = directBranches[index];
+      const bNode = directBranches[inner];
+      if (branchesConflict(aNode, bNode)) {
+        conflictingBranches.push([aNode.id, bNode.id]);
+        conflictingBranchIds.add(aNode.id);
+        conflictingBranchIds.add(bNode.id);
+        continue;
+      }
+
+      const aEvidence = branchEvidenceKeys(aNode);
+      const bEvidence = branchEvidenceKeys(bNode);
+      const sharesEvidence = aEvidence.some((key) => bEvidence.includes(key));
+      if (
+        !sharesEvidence &&
+        aNode.status === "resolved" &&
+        bNode.status === "resolved"
+      ) {
+        corroboratingBranches.push([aNode.id, bNode.id]);
+      }
+    }
+  }
+
+  const interactionBranchIds = new Set([
+    ...duplicateBranchIds,
+    ...conflictingBranchIds,
+    ...corroboratingBranches.flat(),
+  ]);
+  const independentBranches = directBranches
+    .map((branch) => branch.id)
+    .filter((id) => !interactionBranchIds.has(id));
+
+  return {
+    corroboratingBranches,
+    conflictingBranches,
+    duplicateEvidenceGroups,
+    independentBranches,
+    interactionSummary: {
+      corroboratingBranchCount: corroboratingBranches.length,
+      conflictingBranchCount: conflictingBranches.length,
+      duplicateEvidenceCount: duplicateEvidenceGroups.length,
+      independentBranchCount: independentBranches.length,
+    },
+  };
+}
+
 function upsertProposalNodeUpdate(proposalSnapshot, update) {
  const existing = proposalSnapshot.updatedNodes.find(
    (candidate) => candidate.nodeId === update.nodeId,
@@ -585,6 +708,13 @@ function computeParentProgressState(graph, parentNode) {
    (child) => child.status !== "resolved",
  ).length;
  const beforeAssessment = getNodeConfidenceAssessment(parentNode);
+  const interactions = evaluateBranchInteractions({ parentNode, graph });
+  const corroborationCount =
+    interactions.interactionSummary.corroboratingBranchCount;
+  const duplicateEvidenceCount =
+    interactions.interactionSummary.duplicateEvidenceCount;
+  const conflictingBranchCount =
+    interactions.interactionSummary.conflictingBranchCount;

  if (totalChildren === 0) {
    const nextAssessment = {
@@ -625,18 +755,32 @@ function computeParentProgressState(graph, parentNode) {
    confidenceCapReason = "no_resolved_direct_children";
  } else if (resolvedCount < totalChildren) {
    nextAssessment = {
-      evidenceConfidence: "high",
+      evidenceConfidence: corroborationCount > 0 ? "high" : "medium",
      completenessStatus: "partial",
      conclusionConfidence: "medium",
    };
-    confidenceCapReason = "unresolved_direct_children_cap_conclusion";
+    confidenceCapReason =
+      conflictingBranchCount > 0
+        ? "conflicting_branches_cap_conclusion"
+        : duplicateEvidenceCount > 0
+          ? "duplicate_evidence_no_extra_confidence"
+          : corroborationCount > 0
+            ? "independent_corroboration_with_incomplete_parent"
+            : "unresolved_direct_children_cap_conclusion";
  } else {
    nextAssessment = {
      evidenceConfidence: "high",
      completenessStatus: "complete",
-      conclusionConfidence: "high",
+      conclusionConfidence: conflictingBranchCount > 0 ? "low" : "high",
    };
-    confidenceCapReason = null;
+    confidenceCapReason =
+      conflictingBranchCount > 0
+        ? "conflicting_branches_cap_conclusion"
+        : duplicateEvidenceCount > 0
+          ? "duplicate_evidence_no_extra_confidence"
+          : corroborationCount > 0
+            ? "independent_corroboration_supported_conclusion"
+            : null;
  }

  if (resolvedChildren.length === totalChildren) {
@@ -652,6 +796,7 @@ function computeParentProgressState(graph, parentNode) {
      resolvedDirectChildren: resolvedCount,
      unresolvedDirectChildren: unresolvedCount,
      contradictoryDirectChildren: contradictoryChildren.length,
+      branchInteractions: interactions,
      confidenceCapReason,
      reason:
        "All direct child unknowns are resolved, so the parent can now resolve deterministically.",
@@ -671,6 +816,7 @@ function computeParentProgressState(graph, parentNode) {
      resolvedDirectChildren: resolvedCount,
      unresolvedDirectChildren: unresolvedCount,
      contradictoryDirectChildren: contradictoryChildren.length,
+      branchInteractions: interactions,
      confidenceCapReason,
      reason:
        "At least one direct child has been progressed, so the parent becomes provisional but remains unresolved until all direct children are resolved.",
@@ -689,6 +835,7 @@ function computeParentProgressState(graph, parentNode) {
    resolvedDirectChildren: resolvedCount,
    unresolvedDirectChildren: unresolvedCount,
    contradictoryDirectChildren: contradictoryChildren.length,
+    branchInteractions: interactions,
    confidenceCapReason,
    reason: "No direct child progress exists yet for the parent.",
  };
@@ -807,6 +954,19 @@ export function propagateResolvedChildEvidence({
        resolvedDirectChildren: progressState.resolvedDirectChildren,
        unresolvedDirectChildren: progressState.unresolvedDirectChildren,
        contradictoryDirectChildren: progressState.contradictoryDirectChildren,
+        corroboratingBranchCount:
+          progressState.branchInteractions.interactionSummary
+            .corroboratingBranchCount,
+        conflictingBranchCount:
+          progressState.branchInteractions.interactionSummary
+            .conflictingBranchCount,
+        duplicateEvidenceCount:
+          progressState.branchInteractions.interactionSummary
+            .duplicateEvidenceCount,
+        independentBranchCount:
+          progressState.branchInteractions.interactionSummary
+            .independentBranchCount,
+        interactionSummary: progressState.branchInteractions.interactionSummary,
        confidenceCapReason: progressState.confidenceCapReason,
        parentResolved: progressState.parentResolved,
        reason: progressState.reason,
@@ -846,6 +1006,11 @@ export function propagateResolvedChildEvidence({
    resolvedDirectChildren: firstEvent?.resolvedDirectChildren ?? 0,
    unresolvedDirectChildren: firstEvent?.unresolvedDirectChildren ?? 0,
    contradictoryDirectChildren: firstEvent?.contradictoryDirectChildren ?? 0,
+    corroboratingBranchCount: firstEvent?.corroboratingBranchCount ?? 0,
+    conflictingBranchCount: firstEvent?.conflictingBranchCount ?? 0,
+    duplicateEvidenceCount: firstEvent?.duplicateEvidenceCount ?? 0,
+    independentBranchCount: firstEvent?.independentBranchCount ?? 0,
+    interactionSummary: firstEvent?.interactionSummary ?? null,
    confidenceCapReason: firstEvent?.confidenceCapReason ?? null,
    ancestorPropagationStoppedReason,
    affectedAncestorIds: [...affectedAncestorIds],
@@ -1915,6 +2080,11 @@ export function applyValidatedProposal({
  const affectedAncestorIds = propagationResult.affectedAncestorIds;
  const nextSelectedSibling = propagationResult.nextSelectedSibling;
  const parentResolved = propagationResult.parentResolved;
+  const corroboratingBranchCount = propagationResult.corroboratingBranchCount;
+  const conflictingBranchCount = propagationResult.conflictingBranchCount;
+  const duplicateEvidenceCount = propagationResult.duplicateEvidenceCount;
+  const independentBranchCount = propagationResult.independentBranchCount;
+  const interactionSummary = propagationResult.interactionSummary;
  const propagationReason = propagationResult.reason;

  if (
@@ -2048,6 +2218,11 @@ export function applyValidatedProposal({
    resolvedDirectChildren,
    unresolvedDirectChildren,
    contradictoryDirectChildren,
+    corroboratingBranchCount,
+    conflictingBranchCount,
+    duplicateEvidenceCount,
+    independentBranchCount,
+    interactionSummary,
    confidenceCapReason,
    ancestorPropagationStoppedReason,
    affectedAncestorIds,