@remnic/core 9.65.6 → 9.65.7

package/src/eval-trajectory.ts

@@ -0,0 +1,773 @@
+/**
+ * Issue #2345 — privacy-safe action-to-outcome trajectory evaluator (offline slice).
+ *
+ * Pure, local, deterministic evaluator that joins sanitized action rows to
+ * delayed outcome rows inside ONE namespace and ONE bounded time window, then
+ * emits a redacted shadow artifact (digests, enums, numbers, counts only).
+ *
+ * This module never:
+ *   - reads live stores (inputs are synthetic or caller-sanitized rows),
+ *   - calls a model or network,
+ *   - mutates memories, counters, utility runtime, or policy decisions.
+ *
+ * Delivery note (issue "Delivery order"): schema + offline path first. The
+ * live `EvalTrajectoryReader` projections and eval-queue wiring arrive with
+ * #2348; this module ships the contract they will feed.
+ *
+ * Contracts and shared helpers live in `eval-trajectory-contracts.ts`;
+ * validation, persistence, and the promotion gate live in
+ * `eval-trajectory-store.ts`. Both are re-exported here so consumers import
+ * one public surface.
+ */
+
+import assert from "node:assert/strict";
+import {
+  DEFAULT_EVAL_TRAJECTORY_OUTCOME_WINDOW_MS,
+  EVAL_TRAJECTORY_REPORT_SCHEMA_VERSION,
+  type EvalActionInput,
+  type EvalActionOutcomeRow,
+  type EvalActionProjection,
+  type EvalOutcomeInput,
+  type EvalOutcomeProjection,
+  type EvalStaleUseEvidence,
+  type EvalStaleUseEvidenceInput,
+  type EvalTrajectoryAggregates,
+  type EvalTrajectoryJoinKind,
+  type EvalTrajectoryMetricName,
+  type EvalTrajectoryRejectionReason,
+  type EvalTrajectoryReport,
+  type EvalTrajectoryRowStatus,
+  type EvalTrajectoryRunConfig,
+  STALE_LIFECYCLE_STATES,
+  STALE_OUTCOMES,
+  canonicalJson,
+  compareRows,
+  compareStrings,
+  digestOf,
+  isFiniteNumber,
+  isIsoTimestamp,
+  isMemoryActionType,
+  isNonEmptyString,
+  resolveEvalTrajectoryRunConfig,
+  serializeEvalTrajectoryReport,
+  sha256Hex,
+} from "./eval-trajectory-contracts.js";
+import type { MemoryActionType } from "./types.js";
+
+export {
+  DEFAULT_EVAL_TRAJECTORY_MAX_FUTURE_SKEW_MS,
+  DEFAULT_EVAL_TRAJECTORY_MAX_ROWS,
+  DEFAULT_EVAL_TRAJECTORY_OUTCOME_WINDOW_MS,
+  DEFAULT_EVAL_TRAJECTORY_REPORT_MAX_AGE_MS,
+  DEFAULT_EVAL_TRAJECTORY_WEIGHTS,
+  EVAL_TRAJECTORY_PROMOTION_THRESHOLDS,
+  EVAL_TRAJECTORY_REPORT_SCHEMA_VERSION,
+  MAX_EVAL_TRAJECTORY_ROWS_LIMIT,
+  resolveEvalTrajectoryRunConfig,
+  serializeEvalTrajectoryReport,
+  type EvalActionInput,
+  type EvalActionProjection,
+  type EvalOutcomeInput,
+  type EvalOutcomeProjection,
+  type EvalStaleUseEvidence,
+  type EvalStaleUseEvidenceInput,
+  type EvalStaleUseLifecycleState,
+  type EvalStaleUseOutcome,
+  type EvalTrajectoryAggregates,
+  type EvalTrajectoryJoinKind,
+  type EvalTrajectoryMetricName,
+  type EvalTrajectoryOutcomeSource,
+  type EvalTrajectoryRejectionReason,
+  type EvalTrajectoryReport,
+  type EvalTrajectoryRowStatus,
+  type EvalTrajectoryRunConfig,
+  type EvalTrajectoryTaskResult,
+  type EvalTrajectoryWeights,
+  type UnifiedMemoryPromotionReport,
+  type UnifiedMemoryPromotionState,
+} from "./eval-trajectory-contracts.js";
+export * from "./eval-trajectory-store.js";
+
+/**
+ * Future live-reader seam (#2348). Must receive a trusted principal and a
+ * resolved namespace, verify access, and return sanitized rows only.
+ */
+export interface EvalTrajectoryReaderResult {
+  actions: EvalActionInput[];
+  outcomes: EvalOutcomeInput[];
+  staleUseEvidence: EvalStaleUseEvidenceInput[];
+}
+
+export type EvalTrajectoryReader = (input: {
+  trustedPrincipal: string;
+  resolvedNamespace: string;
+  asOf: string;
+}) => Promise<EvalTrajectoryReaderResult> | EvalTrajectoryReaderResult;
+
+/** Synthetic reader over caller-supplied clean rows (tests, offline runs). */
+export function createSyntheticEvalTrajectoryReader(rows: EvalTrajectoryReaderResult): EvalTrajectoryReader {
+  return () => rows;
+}
+
+// ---------------------------------------------------------------------------
+// Input normalization with per-reason rejection counting
+// ---------------------------------------------------------------------------
+
+function emptyRejectionCounts(): Record<EvalTrajectoryRejectionReason, number> {
+  return {
+    missing_action_id: 0,
+    namespace_mismatch: 0,
+    invalid_timestamp: 0,
+    future_beyond_skew: 0,
+    malformed_row: 0,
+    invalid_context_budget: 0,
+    duplicate_row: 0,
+    max_rows_exceeded: 0,
+  };
+}
+
+interface NormalizeContext {
+  namespace: string;
+  namespaceDigest: string;
+  salt: string;
+  asOfMs: number;
+  config: EvalTrajectoryRunConfig;
+  rejections: Record<EvalTrajectoryRejectionReason, number>;
+}
+
+function normalizeAction(raw: unknown, context: NormalizeContext): EvalActionProjection | null {
+  if (typeof raw !== "object" || raw === null) {
+    context.rejections.malformed_row += 1;
+    return null;
+  }
+  const row = raw as Partial<EvalActionInput>;
+  if (!isNonEmptyString(row.actionId)) {
+    context.rejections.missing_action_id += 1;
+    return null;
+  }
+  if (row.namespace !== context.namespace) {
+    context.rejections.namespace_mismatch += 1;
+    return null;
+  }
+  if (!isIsoTimestamp(row.actionAt)) {
+    context.rejections.invalid_timestamp += 1;
+    return null;
+  }
+  if (Date.parse(row.actionAt) > context.asOfMs + context.config.maxFutureSkewMs) {
+    context.rejections.future_beyond_skew += 1;
+    return null;
+  }
+  if (!isMemoryActionType(row.action)) {
+    context.rejections.malformed_row += 1;
+    return null;
+  }
+  const immediateStatus = row.immediateStatus;
+  if (immediateStatus !== "applied" && immediateStatus !== "skipped" && immediateStatus !== "failed") {
+    context.rejections.malformed_row += 1;
+    return null;
+  }
+  if (
+    row.policyDecision !== undefined &&
+    row.policyDecision !== "allow" &&
+    row.policyDecision !== "defer" &&
+    row.policyDecision !== "deny"
+  ) {
+    context.rejections.malformed_row += 1;
+    return null;
+  }
+
+  const tokensPresent =
+    row.contextTokensBefore !== undefined ||
+    row.contextTokensAfter !== undefined ||
+    row.contextTokenBudget !== undefined;
+  let contextTokenCost: number | null = null;
+  if (tokensPresent) {
+    const before = row.contextTokensBefore;
+    const after = row.contextTokensAfter;
+    const budget = row.contextTokenBudget;
+    if (
+      (before !== undefined && (!isFiniteNumber(before) || before < 0)) ||
+      (after !== undefined && (!isFiniteNumber(after) || after < 0)) ||
+      budget === undefined ||
+      !isFiniteNumber(budget)
+    ) {
+      context.rejections.malformed_row += 1;
+      return null;
+    }
+    if (budget === 0) {
+      context.rejections.invalid_context_budget += 1;
+      return null;
+    }
+    contextTokenCost = after === undefined ? null : Math.min(1, Math.max(0, after / Math.max(1, budget)));
+  }
+
+  const appliedDeny = immediateStatus === "applied" && row.policyDecision === "deny";
+  const trustViolation: 0 | 1 =
+    appliedDeny || row.replayMismatch === true || row.missingReversibleGuard === true || row.outsideEligibility === true
+      ? 1
+      : 0;
+
+  return {
+    schemaVersion: 1,
+    actionDigest: digestOf(context.salt, {
+      kind: "action",
+      namespace: context.namespace,
+      actionId: row.actionId,
+    }),
+    namespaceDigest: context.namespaceDigest,
+    actionAt: row.actionAt,
+    action: row.action,
+    immediateStatus,
+    policyDecision: row.policyDecision ?? null,
+    policyVersion: isNonEmptyString(row.policyVersion) ? row.policyVersion : null,
+    source: isNonEmptyString(row.subsystem) ? row.subsystem : null,
+    actionIdDigest: digestOf(context.salt, {
+      kind: "actionId",
+      namespace: context.namespace,
+      actionId: row.actionId,
+    }),
+    sessionKeyDigest: isNonEmptyString(row.sessionKey)
+      ? digestOf(context.salt, { kind: "sessionKey", namespace: context.namespace, sessionKey: row.sessionKey })
+      : null,
+    memoryIdDigest: isNonEmptyString(row.memoryId)
+      ? digestOf(context.salt, { kind: "memoryId", namespace: context.namespace, memoryId: row.memoryId })
+      : null,
+    trajectoryIdDigest: isNonEmptyString(row.trajectoryId)
+      ? digestOf(context.salt, {
+          kind: "trajectoryId",
+          namespace: context.namespace,
+          trajectoryId: row.trajectoryId,
+        })
+      : null,
+    contextTokenCost,
+    trustViolation,
+  };
+}
+
+function normalizeOutcome(raw: unknown, context: NormalizeContext): EvalOutcomeProjection | null {
+  if (typeof raw !== "object" || raw === null) {
+    context.rejections.malformed_row += 1;
+    return null;
+  }
+  const row = raw as Partial<EvalOutcomeInput>;
+  if (!isNonEmptyString(row.outcomeId)) {
+    context.rejections.malformed_row += 1;
+    return null;
+  }
+  if (row.namespace !== context.namespace) {
+    context.rejections.namespace_mismatch += 1;
+    return null;
+  }
+  if (row.source !== "eval_artifact" && row.source !== "causal_trajectory" && row.source !== "memory_outcome") {
+    context.rejections.malformed_row += 1;
+    return null;
+  }
+  if (!isIsoTimestamp(row.outcomeAt)) {
+    context.rejections.invalid_timestamp += 1;
+    return null;
+  }
+  if (Date.parse(row.outcomeAt) > context.asOfMs + context.config.maxFutureSkewMs) {
+    context.rejections.future_beyond_skew += 1;
+    return null;
+  }
+  if (row.result !== undefined && row.result !== "success" && row.result !== "partial" && row.result !== "failure") {
+    context.rejections.malformed_row += 1;
+    return null;
+  }
+  const hasJoinKey =
+    isNonEmptyString(row.actionId) ||
+    isNonEmptyString(row.trajectoryId) ||
+    isNonEmptyString(row.sessionKey) ||
+    (isNonEmptyString(row.memoryId) && isNonEmptyString(row.sessionKey));
+  if (!hasJoinKey) {
+    context.rejections.malformed_row += 1;
+    return null;
+  }
+
+  const value =
+    row.result === undefined || row.source === "memory_outcome"
+      ? null
+      : row.result === "success"
+        ? 1
+        : row.result === "partial"
+          ? 0.5
+          : 0;
+  const ref = row.actionId ?? row.trajectoryId ?? row.memoryId ?? row.sessionKey ?? row.outcomeId;
+
+  return {
+    schemaVersion: 1,
+    outcomeDigest: digestOf(context.salt, {
+      kind: "outcome",
+      namespace: context.namespace,
+      outcomeId: row.outcomeId,
+    }),
+    namespaceDigest: context.namespaceDigest,
+    recordedAt: row.outcomeAt,
+    kind: row.source,
+    value,
+    sourceRefDigest: digestOf(context.salt, { kind: "sourceRef", namespace: context.namespace, ref }),
+    actionIdDigest: isNonEmptyString(row.actionId)
+      ? digestOf(context.salt, { kind: "actionId", namespace: context.namespace, actionId: row.actionId })
+      : null,
+    trajectoryIdDigest: isNonEmptyString(row.trajectoryId)
+      ? digestOf(context.salt, {
+          kind: "trajectoryId",
+          namespace: context.namespace,
+          trajectoryId: row.trajectoryId,
+        })
+      : null,
+    memoryIdDigest: isNonEmptyString(row.memoryId)
+      ? digestOf(context.salt, { kind: "memoryId", namespace: context.namespace, memoryId: row.memoryId })
+      : null,
+    sessionKeyDigest: isNonEmptyString(row.sessionKey)
+      ? digestOf(context.salt, { kind: "sessionKey", namespace: context.namespace, sessionKey: row.sessionKey })
+      : null,
+  };
+}
+
+function normalizeStaleEvidence(raw: unknown, context: NormalizeContext): EvalStaleUseEvidence | null {
+  if (typeof raw !== "object" || raw === null) {
+    context.rejections.malformed_row += 1;
+    return null;
+  }
+  const row = raw as Partial<EvalStaleUseEvidenceInput>;
+  if (!isNonEmptyString(row.actionId) || !isNonEmptyString(row.memoryId)) {
+    context.rejections.malformed_row += 1;
+    return null;
+  }
+  if (row.namespace !== context.namespace) {
+    context.rejections.namespace_mismatch += 1;
+    return null;
+  }
+  if (typeof row.lifecycleState !== "string" || STALE_LIFECYCLE_STATES[row.lifecycleState] !== true) {
+    context.rejections.malformed_row += 1;
+    return null;
+  }
+  if (!isIsoTimestamp(row.usedAt) || typeof row.outcome !== "string" || STALE_OUTCOMES[row.outcome] !== true) {
+    context.rejections.malformed_row += 1;
+    return null;
+  }
+  return {
+    schemaVersion: 1,
+    actionDigest: digestOf(context.salt, {
+      kind: "action",
+      namespace: context.namespace,
+      actionId: row.actionId,
+    }),
+    memoryDigest: digestOf(context.salt, {
+      kind: "memoryId",
+      namespace: context.namespace,
+      memoryId: row.memoryId,
+    }),
+    lifecycleState: row.lifecycleState,
+    usedAt: row.usedAt,
+    outcome: row.outcome,
+  };
+}
+
+// ---------------------------------------------------------------------------
+// Join + metric core
+// ---------------------------------------------------------------------------
+
+interface JoinCandidateLevel {
+  kind: EvalTrajectoryJoinKind;
+  matches: (action: EvalActionProjection, outcome: EvalOutcomeProjection) => boolean;
+}
+
+const JOIN_LEVELS: JoinCandidateLevel[] = [
+  {
+    kind: "action_id",
+    matches: (action, outcome) => outcome.actionIdDigest !== null && outcome.actionIdDigest === action.actionIdDigest,
+  },
+  {
+    kind: "trajectory_id",
+    matches: (action, outcome) =>
+      action.trajectoryIdDigest !== null &&
+      outcome.trajectoryIdDigest !== null &&
+      outcome.trajectoryIdDigest === action.trajectoryIdDigest,
+  },
+  {
+    kind: "memory_scope",
+    matches: (action, outcome) =>
+      outcome.kind === "memory_outcome" &&
+      action.memoryIdDigest !== null &&
+      outcome.memoryIdDigest !== null &&
+      outcome.memoryIdDigest === action.memoryIdDigest &&
+      action.sessionKeyDigest !== null &&
+      outcome.sessionKeyDigest !== null &&
+      outcome.sessionKeyDigest === action.sessionKeyDigest,
+  },
+  {
+    kind: "session_scope",
+    matches: (action, outcome) =>
+      outcome.kind !== "memory_outcome" &&
+      action.sessionKeyDigest !== null &&
+      outcome.sessionKeyDigest !== null &&
+      outcome.sessionKeyDigest === action.sessionKeyDigest,
+  },
+];
+
+/** Half-open delayed window: `actionAt < outcomeAt < actionAt + outcomeWindowMs`. */
+function outcomeInWindow(
+  action: EvalActionProjection,
+  outcome: EvalOutcomeProjection,
+  config: EvalTrajectoryRunConfig
+): boolean {
+  const outcomeMs = Date.parse(outcome.recordedAt);
+  return Date.parse(action.actionAt) < outcomeMs && outcomeMs < Date.parse(action.actionAt) + config.outcomeWindowMs;
+}
+
+function anyKeyMatch(action: EvalActionProjection, outcome: EvalOutcomeProjection): boolean {
+  return JOIN_LEVELS.some((level) => level.matches(action, outcome));
+}
+
+function staleHarmFor(action: EvalActionProjection, evidence: EvalStaleUseEvidence[]): 0 | 0.5 | 1 {
+  let harm: 0 | 0.5 | 1 = 0;
+  for (const item of evidence) {
+    if (item.actionDigest !== action.actionDigest) continue;
+    if (item.lifecycleState !== "stale" && item.lifecycleState !== "archived") continue;
+    if (Date.parse(item.usedAt) <= Date.parse(action.actionAt)) continue;
+    if (item.outcome === "failure") return 1;
+    if (item.outcome === "partial" || item.outcome === "unknown") harm = 0.5;
+  }
+  return harm;
+}
+
+function buildRow(
+  action: EvalActionProjection,
+  outcomes: EvalOutcomeProjection[],
+  evidence: EvalStaleUseEvidence[],
+  config: EvalTrajectoryRunConfig,
+  asOfMs: number
+): EvalActionOutcomeRow {
+  let selected: EvalOutcomeProjection | null = null;
+  let joinKind: EvalTrajectoryJoinKind | null = null;
+  let ambiguous = false;
+
+  for (const level of JOIN_LEVELS) {
+    const inWindow = outcomes.filter(
+      (outcome) =>
+        level.matches(action, outcome) &&
+        outcomeInWindow(action, outcome, config) &&
+        Date.parse(outcome.recordedAt) <= asOfMs
+    );
+    if (inWindow.length === 1) {
+      selected = inWindow[0] ?? null;
+      joinKind = level.kind;
+      break;
+    }
+    if (inWindow.length > 1) {
+      ambiguous = true;
+      joinKind = level.kind;
+      break;
+    }
+  }
+
+  let status: EvalTrajectoryRowStatus;
+  let outcomeDigest: string | null = null;
+  let outcomeRecordedAt: string | null = null;
+  let lateProjection: EvalOutcomeProjection | null = null;
+  if (ambiguous) {
+    status = "ambiguous";
+  } else if (selected !== null) {
+    status = "joined";
+    outcomeDigest = selected.outcomeDigest;
+    outcomeRecordedAt = selected.recordedAt;
+  } else {
+    const keyMatches = outcomes.filter((outcome) => anyKeyMatch(action, outcome));
+    // Prefer the in-data off-window result when one exists; a future result
+    // (recorded after asOf) is reported but never matched.
+    const lateCandidates = keyMatches
+      .filter(
+        (outcome) =>
+          Date.parse(outcome.recordedAt) >= Date.parse(action.actionAt) + config.outcomeWindowMs &&
+          Date.parse(outcome.recordedAt) <= asOfMs
+      )
+      .sort((a, b) => compareStrings(a.recordedAt, b.recordedAt));
+    lateProjection = lateCandidates.length > 0 ? (lateCandidates[lateCandidates.length - 1] ?? null) : null;
+    if (lateProjection !== null) {
+      status = "late_outcome";
+      outcomeDigest = lateProjection.outcomeDigest;
+      outcomeRecordedAt = lateProjection.recordedAt;
+    } else if (keyMatches.some((outcome) => Date.parse(outcome.recordedAt) > asOfMs)) {
+      status = "future_outcome";
+    } else {
+      status = "missing_within_window";
+    }
+  }
+
+  // Only task-source outcomes carry a task value; memory outcomes never do.
+  const matched = selected ?? (status === "late_outcome" ? lateProjection : null);
+  const taskOutcome: number | null =
+    status === "joined" || status === "late_outcome"
+      ? matched !== null && matched.kind !== "memory_outcome"
+        ? matched.value
+        : null
+      : null;
+
+  const staleMemoryHarm = staleHarmFor(action, evidence);
+  const weights = config.weights;
+  const missingDimensions: EvalTrajectoryMetricName[] = [];
+  if (taskOutcome === null && weights.task > 0) missingDimensions.push("taskOutcome");
+  if (action.contextTokenCost === null && weights.context > 0) missingDimensions.push("contextTokenCost");
+
+  const actionConditionedUtility =
+    status !== "joined" || missingDimensions.length > 0
+      ? null
+      : Math.min(
+          1,
+          Math.max(
+            -1,
+            weights.task * (taskOutcome ?? 0) -
+              weights.context * (action.contextTokenCost ?? 0) -
+              weights.trust * action.trustViolation -
+              weights.stale * staleMemoryHarm
+          )
+        );
+
+  return {
+    schemaVersion: 1,
+    actionDigest: action.actionDigest,
+    outcomeDigest: status === "future_outcome" ? null : outcomeDigest,
+    join: status === "joined" || status === "ambiguous" ? joinKind : null,
+    status,
+    outcomeRecordedAt: status === "future_outcome" ? null : outcomeRecordedAt,
+    taskOutcome,
+    contextTokenCost: action.contextTokenCost,
+    trustViolation: action.trustViolation,
+    staleMemoryHarm,
+    actionConditionedUtility,
+    missingDimensions,
+  };
+}
+
+// ---------------------------------------------------------------------------
+// Report builder (pure — no I/O, no model, no memory writes)
+// ---------------------------------------------------------------------------
+
+export interface BuildEvalTrajectoryReportOptions {
+  /** Single resolved namespace for the whole run. Never "all namespaces". */
+  namespace: string;
+  asOf: string;
+  /** Per-run salt. Required, never exported. */
+  salt: string;
+  actions: unknown[];
+  outcomes?: unknown[];
+  staleUseEvidence?: unknown[];
+  config?: Partial<EvalTrajectoryRunConfig>;
+}
+
+export function buildEvalTrajectoryReport(options: BuildEvalTrajectoryReportOptions): EvalTrajectoryReport {
+  if (!isNonEmptyString(options.namespace)) throw new Error("eval trajectory run requires a namespace");
+  if (!isNonEmptyString(options.salt)) throw new Error("eval trajectory run requires a salt");
+  if (!isIsoTimestamp(options.asOf)) throw new Error("asOf must be a normalized ISO-8601 UTC timestamp");
+  const config = resolveEvalTrajectoryRunConfig(options.config);
+  const asOfMs = Date.parse(options.asOf);
+  const context: NormalizeContext = {
+    namespace: options.namespace,
+    namespaceDigest: digestOf(options.salt, { kind: "namespace", namespace: options.namespace }),
+    salt: options.salt,
+    asOfMs,
+    config,
+    rejections: emptyRejectionCounts(),
+  };
+
+  const actions: EvalActionProjection[] = [];
+  const seenActionDigests = new Set<string>();
+  for (const raw of options.actions ?? []) {
+    const projection = normalizeAction(raw, context);
+    if (projection === null) continue;
+    if (seenActionDigests.has(projection.actionDigest)) {
+      context.rejections.duplicate_row += 1;
+      continue;
+    }
+    seenActionDigests.add(projection.actionDigest);
+    actions.push(projection);
+  }
+  actions.sort((a, b) => compareStrings(a.actionDigest, b.actionDigest));
+  if (actions.length > config.maxRows) {
+    context.rejections.max_rows_exceeded += actions.length - config.maxRows;
+    actions.length = config.maxRows;
+  }
+
+  const outcomes: EvalOutcomeProjection[] = [];
+  const seenOutcomeDigests = new Set<string>();
+  for (const raw of options.outcomes ?? []) {
+    const projection = normalizeOutcome(raw, context);
+    if (projection === null) continue;
+    if (seenOutcomeDigests.has(projection.outcomeDigest)) {
+      context.rejections.duplicate_row += 1;
+      continue;
+    }
+    seenOutcomeDigests.add(projection.outcomeDigest);
+    outcomes.push(projection);
+  }
+  outcomes.sort((a, b) => compareStrings(a.outcomeDigest, b.outcomeDigest));
+
+  const evidence: EvalStaleUseEvidence[] = [];
+  for (const raw of options.staleUseEvidence ?? []) {
+    const projection = normalizeStaleEvidence(raw, context);
+    if (projection !== null) evidence.push(projection);
+  }
+  evidence.sort((a, b) => {
+    const byUsedAt = compareStrings(a.usedAt, b.usedAt);
+    if (byUsedAt !== 0) return byUsedAt;
+    return compareStrings(a.memoryDigest, b.memoryDigest);
+  });
+
+  const rows = actions.map((action) => buildRow(action, outcomes, evidence, config, asOfMs));
+  rows.sort(compareRows);
+
+  const inputFingerprint = sha256Hex(
+    canonicalJson({
+      schemaVersion: EVAL_TRAJECTORY_REPORT_SCHEMA_VERSION,
+      namespace: options.namespace,
+      asOf: options.asOf,
+      salt: options.salt,
+      config,
+      actions,
+      outcomes,
+      evidence,
+    })
+  );
+  const reportId = sha256Hex(
+    canonicalJson({
+      kind: "eval-trajectory-report",
+      inputFingerprint,
+      namespaceDigest: context.namespaceDigest,
+      asOf: options.asOf,
+    })
+  );
+
+  const report: EvalTrajectoryReport = {
+    schemaVersion: EVAL_TRAJECTORY_REPORT_SCHEMA_VERSION,
+    reportId,
+    namespaceDigest: context.namespaceDigest,
+    asOf: options.asOf,
+    config,
+    sourceCounts: { actions: actions.length, outcomes: outcomes.length, staleUseEvidence: evidence.length },
+    rejectionCounts: context.rejections,
+    rows,
+    aggregates: computeAggregates(rows, actions),
+    inputFingerprint,
+    contentHash: "",
+  };
+  report.contentHash = sha256Hex(canonicalJson({ ...report, contentHash: undefined }));
+  return report;
+}
+
+function computeAggregates(rows: EvalActionOutcomeRow[], actions: EvalActionProjection[]): EvalTrajectoryAggregates {
+  const statusCounts: Record<EvalTrajectoryRowStatus, number> = {
+    joined: 0,
+    ambiguous: 0,
+    late_outcome: 0,
+    future_outcome: 0,
+    missing_within_window: 0,
+  };
+  for (const row of rows) statusCounts[row.status] += 1;
+
+  const digestToAction = new Map(actions.map((action) => [action.actionDigest, action]));
+  const joinedRows = rows.filter((row) => row.status === "joined");
+  const mean = (values: number[]): number | null =>
+    values.length === 0 ? null : values.reduce((sum, value) => sum + value, 0) / values.length;
+
+  interface GroupTally {
+    rowCount: number;
+    utilitySum: number;
+    utilityCount: number;
+  }
+  const bump = (map: Map<string, GroupTally>, key: string, utility: number | null): void => {
+    const tally = map.get(key) ?? { rowCount: 0, utilitySum: 0, utilityCount: 0 };
+    tally.rowCount += 1;
+    if (utility !== null) {
+      tally.utilitySum += utility;
+      tally.utilityCount += 1;
+    }
+    map.set(key, tally);
+  };
+  const byActionMap = new Map<string, GroupTally>();
+  const byPolicyMap = new Map<string, GroupTally>();
+  for (const row of joinedRows) {
+    const action = digestToAction.get(row.actionDigest);
+    if (!action) continue;
+    bump(byActionMap, action.action, row.actionConditionedUtility);
+    bump(byPolicyMap, action.policyDecision ?? "unrecorded", row.actionConditionedUtility);
+  }
+
+  const reachDenominator = rows.length - statusCounts.future_outcome;
+  return {
+    rowCount: rows.length,
+    statusCounts,
+    metricMeans: {
+      taskOutcome: mean(joinedRows.map((row) => row.taskOutcome).filter(isFiniteNumber)),
+      contextTokenCost: mean(joinedRows.map((row) => row.contextTokenCost).filter(isFiniteNumber)),
+      trustViolation: mean(joinedRows.map((row) => row.trustViolation).filter(isFiniteNumber)),
+      staleMemoryHarm: mean(joinedRows.map((row) => row.staleMemoryHarm).filter(isFiniteNumber)),
+      actionConditionedUtility: mean(joinedRows.map((row) => row.actionConditionedUtility).filter(isFiniteNumber)),
+    },
+    coverage: {
+      reach: reachDenominator > 0 ? statusCounts.joined / reachDenominator : null,
+      joined: statusCounts.joined,
+      missingWithinWindow: statusCounts.missing_within_window,
+      lateOutcome: statusCounts.late_outcome,
+      futureOutcome: statusCounts.future_outcome,
+      ambiguous: statusCounts.ambiguous,
+      excludedFromMeans: rows.length - statusCounts.joined,
+    },
+    byAction: [...byActionMap.entries()]
+      .sort(([a], [b]) => compareStrings(a, b))
+      .map(([action, tally]) => ({
+        action: action as MemoryActionType,
+        rowCount: tally.rowCount,
+        meanUtility: tally.utilityCount > 0 ? tally.utilitySum / tally.utilityCount : null,
+      })),
+    byPolicyDecision: [...byPolicyMap.entries()]
+      .sort(([a], [b]) => compareStrings(a, b))
+      .map(([policyDecision, tally]) => ({
+        policyDecision,
+        rowCount: tally.rowCount,
+        meanUtility: tally.utilityCount > 0 ? tally.utilitySum / tally.utilityCount : null,
+      })),
+  };
+}
+
+// ---------------------------------------------------------------------------
+// Self-check (minimal made-up fixture): npx tsx packages/remnic-core/src/eval-trajectory.ts
+// ---------------------------------------------------------------------------
+
+if (process.argv[1]?.endsWith("eval-trajectory.ts") === true) {
+  const report = buildEvalTrajectoryReport({
+    namespace: "self-check",
+    asOf: "2026-08-15T00:00:00.000Z",
+    salt: "self-check-salt",
+    actions: [
+      {
+        schemaVersion: 1,
+        actionId: "act-1",
+        namespace: "self-check",
+        actionAt: "2026-08-10T00:00:00.000Z",
+        action: "store_note",
+        immediateStatus: "applied",
+        sessionKey: "agent:main",
+      },
+    ],
+    outcomes: [
+      {
+        schemaVersion: 1,
+        namespace: "self-check",
+        outcomeId: "out-1",
+        outcomeAt: "2026-08-12T00:00:00.000Z",
+        source: "causal_trajectory",
+        result: "partial",
+        sessionKey: "agent:main",
+      },
+    ],
+  });
+  const row = report.rows[0];
+  assert.equal(row?.status, "joined", "self-check row must join");
+  assert.equal(row?.taskOutcome, 0.5, "partial must score 0.5");
+  assert.equal(report.aggregates.coverage.reach, 1, "reach must be 1");
+  assert.ok(serializeEvalTrajectoryReport(report).includes('"contentHash"'), "report must serialize");
+  assert.ok(DEFAULT_EVAL_TRAJECTORY_OUTCOME_WINDOW_MS === 604_800_000, "default window must stay 7 days");
+  console.log("eval-trajectory self-check OK");
+}