chainlesschain 0.166.91 → 0.166.93

package/src/lib/evolution/rrsi-runtime-composition.js

@@ -0,0 +1,79 @@
+/** Initial governed PM preparation composition; no signer/provider/promotion defaults. */
+import { isProxy } from "node:util/types";
+import { verifyRrsiCampaign } from "./rrsi-contracts.js";
+import { recheckRrsiEffectiveParent } from "./rrsi-parent-binding.js";
+import { bindRrsiPmRuntime } from "./rrsi-pm-execution-bridge.js";
+import { snapshotRrsiData, rrsiEnvelope, rrsiFail } from "./rrsi-data.js";
+
+export const RRSI_RUNTIME_COMPOSITION_SCHEMA =
+  "chainlesschain.rrsi-runtime-composition/v1";
+const COMPOSITIONS = new WeakMap();
+
+export function createRrsiRuntimeComposition(input) {
+  const keys = ["campaign", "pmBridge", "parentBinding", "mode"];
+  if (
+    !input ||
+    typeof input !== "object" ||
+    isProxy(input) ||
+    Object.getPrototypeOf(input) !== Object.prototype ||
+    Reflect.ownKeys(input).length !== keys.length
+  )
+    rrsiFail("RRSI runtime composition requires plain own fields");
+  const ports = Object.fromEntries(
+    keys.map((key) => {
+      const field = Object.getOwnPropertyDescriptor(input, key);
+      if (!field?.enumerable || !("value" in field))
+        rrsiFail("RRSI runtime composition cannot use accessors");
+      return [key, field.value];
+    }),
+  );
+  const campaign = verifyRrsiCampaign(snapshotRrsiData(ports.campaign));
+  recheckRrsiEffectiveParent(ports.parentBinding);
+  if (ports.parentBinding.descriptor.campaignDigest !== campaign.campaignDigest)
+    rrsiFail("RRSI runtime parent belongs to another campaign");
+  // Mutating binding happens last, after all independent validation.
+  const bridge = bindRrsiPmRuntime(
+    ports.pmBridge,
+    ports.parentBinding,
+    ports.mode,
+  );
+  const composition = Object.freeze({
+    descriptor: rrsiEnvelope(
+      RRSI_RUNTIME_COMPOSITION_SCHEMA,
+      "compositionDigest",
+      {
+        campaignDigest: campaign.campaignDigest,
+        parentBindingDigest: bridge.parentBindingDigest,
+        bridgeDigest: bridge.descriptor.bridgeDigest,
+        mode: bridge.mode,
+        supportedOperation: bridge.descriptor.supportedOperation,
+        effectiveParentRecheckedBeforeReservationAndDispatch: true,
+        fullProductionAdmissionVerified: false,
+        monetaryBudgetAuthorityVerified: false,
+        supportsEnforcedMode: false,
+        supportsPromotion: false,
+      },
+    ),
+  });
+  COMPOSITIONS.set(composition, bridge);
+  return composition;
+}
+
+function portsFor(composition) {
+  const ports = COMPOSITIONS.get(composition);
+  if (!ports) rrsiFail("a branded live RRSI runtime composition is required");
+  return ports;
+}
+
+export function reserveRrsiRuntimePmBroadRound(composition, journal, input) {
+  return portsFor(composition).reserveBroadRound(journal, input);
+}
+
+export async function executeRrsiRuntimePmBroadRound(composition, response) {
+  return portsFor(composition).executeBroadRound(response);
+}
+
+/** Stale parents and off mode must not prevent existing execution reconciliation. */
+export function inspectRrsiRuntimeHistory(composition) {
+  return portsFor(composition).inspectHistory();
+}

package/src/lib/evolution/rrsi-selector.js

@@ -0,0 +1,434 @@
+/** Synthetic replay only. No aggregate supplied here is authenticated evidence. */
+import { verifyRrsiCampaign, verifyRrsiCandidate } from "./rrsi-contracts.js";
+import {
+  snapshotRrsiData,
+  rrsiExact,
+  rrsiDigest,
+  rrsiInteger,
+  rrsiFinite,
+  rrsiBoolean,
+  rrsiId,
+  rrsiFail,
+  rrsiEnvelope,
+} from "./rrsi-data.js";
+
+export const RRSI_REPLAY_SCHEMA = "chainlesschain.rrsi-selection-replay/v1";
+const OBSERVATION_KEYS = [
+  "candidateDigest",
+  "trainDelta",
+  "selectionDelta",
+  "baselineCostMicrounits",
+  "candidateCostMicrounits",
+  "baselineP95LatencyMs",
+  "candidateP95LatencyMs",
+  "baselineTokens",
+  "candidateTokens",
+  "baselineToolCalls",
+  "candidateToolCalls",
+  "perturbationDeltas",
+  "wallClockMs",
+  "executions",
+  "history",
+  "plannedSlots",
+  "completedSlots",
+  "independentGroups",
+  "safetyViolations",
+  "permissionViolations",
+  "dataLeakage",
+  "environmentMatches",
+  "cleanupConfirmed",
+  "costComplete",
+  "ancestorPassed",
+];
+const COST_METRICS = [
+  "baselineCostMicrounits",
+  "candidateCostMicrounits",
+  "baselineP95LatencyMs",
+  "candidateP95LatencyMs",
+  "baselineTokens",
+  "candidateTokens",
+  "baselineToolCalls",
+  "candidateToolCalls",
+];
+
+function observation(input, campaign) {
+  rrsiExact(input, OBSERVATION_KEYS, "synthetic observation");
+  const result = {
+    candidateDigest: rrsiDigest(input.candidateDigest, "candidate digest"),
+  };
+  result.trainDelta =
+    input.trainDelta === null
+      ? null
+      : rrsiFinite(input.trainDelta, "training delta", -1, 1);
+  if (input.selectionDelta === null) result.selectionDelta = null;
+  else {
+    result.selectionDelta = Object.fromEntries(
+      ["mean", "lower", "upper"].map((key) => {
+        rrsiExact(
+          input.selectionDelta,
+          ["mean", "lower", "upper"],
+          "synthetic delta interval",
+        );
+        return [
+          key,
+          rrsiFinite(input.selectionDelta[key], `delta ${key}`, -1, 1),
+        ];
+      }),
+    );
+    if (
+      result.selectionDelta.lower > result.selectionDelta.mean ||
+      result.selectionDelta.mean > result.selectionDelta.upper
+    )
+      rrsiFail("invalid synthetic delta interval");
+  }
+  for (const key of COST_METRICS)
+    result[key] = input[key] === null ? null : rrsiInteger(input[key], key);
+  for (const key of ["wallClockMs", "executions"])
+    result[key] = input[key] === null ? null : rrsiInteger(input[key], key);
+  for (const key of [
+    "plannedSlots",
+    "completedSlots",
+    "independentGroups",
+    "safetyViolations",
+    "permissionViolations",
+  ])
+    result[key] = rrsiInteger(input[key], key);
+  if (
+    result.completedSlots > result.plannedSlots ||
+    result.independentGroups > campaign.dataset.pools.select.componentCount ||
+    result.independentGroups > campaign.dataset.pools.select.taskCount
+  )
+    rrsiFail("synthetic observation invents slots or independent groups");
+  for (const key of [
+    "dataLeakage",
+    "environmentMatches",
+    "cleanupConfirmed",
+    "costComplete",
+    "ancestorPassed",
+  ])
+    result[key] = rrsiBoolean(input[key], key);
+  rrsiExact(
+    input.history,
+    ["oscillations", "repeatedFailures"],
+    "synthetic history",
+  );
+  result.history = {
+    oscillations: rrsiInteger(
+      input.history.oscillations,
+      "oscillations",
+      0,
+      1000,
+    ),
+    repeatedFailures: rrsiInteger(
+      input.history.repeatedFailures,
+      "repeated failures",
+      0,
+      1000,
+    ),
+  };
+  if (
+    !Array.isArray(input.perturbationDeltas) ||
+    input.perturbationDeltas.length > 8
+  )
+    rrsiFail("invalid synthetic perturbations");
+  const seen = new Set();
+  result.perturbationDeltas = input.perturbationDeltas
+    .map((entry) => {
+      rrsiExact(entry, ["id", "mean"], "synthetic perturbation");
+      const id = rrsiId(entry.id, "perturbation ID");
+      if (seen.has(id) || !campaign.experiment.perturbations.includes(id))
+        rrsiFail("duplicate or unregistered perturbation");
+      seen.add(id);
+      return {
+        id,
+        mean:
+          entry.mean === null
+            ? null
+            : rrsiFinite(entry.mean, "perturbation delta", -1, 1),
+      };
+    })
+    .sort((a, b) => (a.id < b.id ? -1 : a.id > b.id ? 1 : 0));
+  return result;
+}
+
+function decision(candidate, raw, policy, campaign) {
+  const result = {
+    candidateDigest: candidate.candidateDigest,
+    contentDigest: candidate.contentDigest,
+    status: "hold",
+    reasons: [],
+    rawMetrics: raw,
+    penalties: null,
+    contributions: null,
+    score: null,
+  };
+  if (!raw) {
+    result.reasons.push("INSUFFICIENT_EVIDENCE");
+    return result;
+  }
+  const rejects = [];
+  if (raw.safetyViolations || raw.permissionViolations)
+    rejects.push("SAFETY_HARD_GATE");
+  if (raw.dataLeakage) rejects.push("DATA_LEAKAGE");
+  if (!raw.ancestorPassed) rejects.push("ANCESTOR_REGRESSION");
+  const expectedSlots =
+    campaign.dataset.pools.select.taskCount *
+    campaign.experiment.seeds.length *
+    (1 + campaign.experiment.perturbations.length);
+  if (
+    raw.plannedSlots !== expectedSlots ||
+    raw.completedSlots !== expectedSlots ||
+    raw.independentGroups < policy.thresholds.minIndependentGroups ||
+    raw.trainDelta === null ||
+    raw.selectionDelta === null ||
+    raw.perturbationDeltas.length !==
+      campaign.experiment.perturbations.length ||
+    raw.perturbationDeltas.some((entry) => entry.mean === null)
+  )
+    result.reasons.push("INSUFFICIENT_EVIDENCE");
+  if (!raw.costComplete || COST_METRICS.some((key) => raw[key] === null))
+    result.reasons.push("COST_UNKNOWN");
+  if (raw.executions === null || raw.wallClockMs === null)
+    result.reasons.push("BUDGET_UNKNOWN");
+  else if (raw.executions < 2 * raw.completedSlots)
+    result.reasons.push("INSUFFICIENT_EVIDENCE");
+  if (!raw.environmentMatches) result.reasons.push("ENVIRONMENT_DRIFT");
+  if (!raw.cleanupConfirmed) result.reasons.push("CLEANUP_UNCONFIRMED");
+  if (!policy.calibrated) result.reasons.push("POLICY_NOT_CALIBRATED");
+  if (rejects.length)
+    return {
+      ...result,
+      status: "rejected",
+      reasons: [...rejects, ...result.reasons],
+    };
+  if (result.reasons.length) return result;
+  const noise = Math.max(
+    ...raw.perturbationDeltas.map((entry) =>
+      Math.max(0, raw.selectionDelta.mean - entry.mean),
+    ),
+  );
+  const size = candidate.artifacts.reduce(
+    (sum, artifact) => sum + artifact.bytes,
+    0,
+  );
+  const nodes = candidate.artifacts.reduce(
+    (sum, artifact) => sum + artifact.workflowNodes,
+    0,
+  );
+  const excess = (actual, target) => Math.max(0, (actual - target) / target);
+  const penalties = {
+    generalization: Math.max(0, raw.trainDelta - raw.selectionDelta.mean),
+    cost:
+      policy.cost.costWeight *
+        Math.max(
+          0,
+          (raw.candidateCostMicrounits - raw.baselineCostMicrounits) /
+            policy.cost.costScaleMicrounits,
+        ) +
+      policy.cost.latencyWeight *
+        Math.max(
+          0,
+          (raw.candidateP95LatencyMs - raw.baselineP95LatencyMs) /
+            policy.cost.latencyScaleMs,
+        ),
+    noise,
+    history:
+      (policy.history.oscillationWeight * raw.history.oscillations) /
+        policy.history.maxOscillations +
+      (policy.history.repeatedFailureWeight * raw.history.repeatedFailures) /
+        policy.history.maxRepeatedFailures,
+    complexity:
+      policy.complexity.bytesWeight *
+        excess(size, policy.complexity.targetArtifactBytes) +
+      policy.complexity.nodesWeight *
+        excess(nodes, policy.complexity.targetWorkflowNodes) +
+      policy.complexity.filesWeight *
+        excess(
+          candidate.artifacts.length,
+          policy.complexity.targetChangedFiles,
+        ),
+  };
+  const contributions = Object.fromEntries(
+    Object.keys(penalties).map((key) => [
+      key,
+      penalties[key] * policy.weights[key],
+    ]),
+  );
+  const score =
+    raw.selectionDelta.lower -
+    Object.values(contributions).reduce((sum, term) => sum + term, 0);
+  const reduction = (base, next) => (base > 0 ? (base - next) / base : 0);
+  const efficient =
+    raw.selectionDelta.lower >= 0 &&
+    raw.candidateCostMicrounits <= raw.baselineCostMicrounits &&
+    Math.max(
+      reduction(raw.baselineTokens, raw.candidateTokens),
+      reduction(raw.baselineP95LatencyMs, raw.candidateP95LatencyMs),
+      reduction(raw.baselineToolCalls, raw.candidateToolCalls),
+    ) >= policy.thresholds.minimumEfficiencyReduction;
+  const reasons = [];
+  if (noise > policy.thresholds.maxNoiseRegression)
+    reasons.push("NO_ROBUST_GAIN");
+  if (
+    raw.selectionDelta.lower < policy.thresholds.minimumQualityDelta &&
+    !efficient
+  )
+    reasons.push("NO_ROBUST_GAIN");
+  if (score < 0) reasons.push("REGULARIZED_GAIN_NOT_MET");
+  if (
+    raw.history.oscillations >= policy.history.maxOscillations ||
+    raw.history.repeatedFailures >= policy.history.maxRepeatedFailures
+  )
+    reasons.push("HISTORY_LIMIT");
+  if (
+    raw.candidateCostMicrounits + raw.baselineCostMicrounits >
+      campaign.budget.selectionPerExploringArm.maxCostMicrounits ||
+    raw.candidateTokens + raw.baselineTokens >
+      campaign.budget.selectionPerExploringArm.maxTokens ||
+    raw.candidateToolCalls + raw.baselineToolCalls >
+      campaign.budget.selectionPerExploringArm.maxToolCalls
+  )
+    reasons.push("BUDGET_EXCEEDED");
+  return {
+    ...result,
+    penalties,
+    contributions,
+    score,
+    reasons: [...new Set(reasons)],
+    status: reasons.length ? "rejected" : "shadow-eligible",
+  };
+}
+
+/** Caller aggregates are synthetic examples, never a production Gate receipt. */
+export function replayRrsiSelection(input) {
+  const value = snapshotRrsiData(input);
+  rrsiExact(
+    value,
+    ["campaign", "candidates", "observations"],
+    "RRSI replay input",
+  );
+  const campaign = verifyRrsiCampaign(value.campaign);
+  if (
+    !Array.isArray(value.candidates) ||
+    value.candidates.length < 1 ||
+    value.candidates.length > campaign.policy.limits.maxCandidates ||
+    value.candidates.length > campaign.policy.limits.maxSelectionQueries
+  )
+    rrsiFail("replay exceeds the frozen candidate or query limit");
+  const candidates = value.candidates.map((candidate) =>
+    verifyRrsiCandidate(campaign, candidate),
+  );
+  const digests = new Set();
+  const ids = new Set();
+  for (const candidate of candidates) {
+    if (
+      digests.has(candidate.candidateDigest) ||
+      ids.has(candidate.candidateId)
+    )
+      rrsiFail("duplicate candidate identity");
+    digests.add(candidate.candidateDigest);
+    ids.add(candidate.candidateId);
+  }
+  if (
+    !Array.isArray(value.observations) ||
+    value.observations.length > candidates.length
+  )
+    rrsiFail("invalid observation count");
+  const observed = new Map();
+  for (const input of value.observations) {
+    const raw = observation(input, campaign);
+    if (!digests.has(raw.candidateDigest) || observed.has(raw.candidateDigest))
+      rrsiFail("unknown or duplicate observed candidate");
+    observed.set(raw.candidateDigest, raw);
+  }
+  const contentCounts = new Map();
+  for (const candidate of candidates)
+    contentCounts.set(
+      candidate.contentDigest,
+      (contentCounts.get(candidate.contentDigest) ?? 0) + 1,
+    );
+  const decisions = candidates.map((candidate) => {
+    const result = decision(
+      candidate,
+      observed.get(candidate.candidateDigest) ?? null,
+      campaign.policy,
+      campaign,
+    );
+    if (contentCounts.get(candidate.contentDigest) > 1)
+      return {
+        ...result,
+        status: "rejected",
+        reasons: [...result.reasons, "DUPLICATE_CONTENT"],
+      };
+    return result;
+  });
+  const incompleteCodes = new Set([
+    "INSUFFICIENT_EVIDENCE",
+    "COST_UNKNOWN",
+    "BUDGET_UNKNOWN",
+    "ENVIRONMENT_DRIFT",
+    "CLEANUP_UNCONFIRMED",
+    "POLICY_NOT_CALIBRATED",
+  ]);
+  const blockingReasons = new Set(
+    decisions.flatMap((result) =>
+      result.reasons.filter((reason) => incompleteCodes.has(reason)),
+    ),
+  );
+  for (const [resource, fields] of [
+    [
+      "maxCostMicrounits",
+      ["baselineCostMicrounits", "candidateCostMicrounits"],
+    ],
+    ["maxTokens", ["baselineTokens", "candidateTokens"]],
+    ["maxToolCalls", ["baselineToolCalls", "candidateToolCalls"]],
+    ["maxWallClockMs", ["wallClockMs"]],
+    ["maxExecutions", ["executions"]],
+  ]) {
+    let total = 0n;
+    for (const raw of observed.values())
+      for (const field of fields)
+        if (raw[field] !== null) total += BigInt(raw[field]);
+    if (total > BigInt(campaign.budget.selectionPerExploringArm[resource]))
+      blockingReasons.add("BUDGET_EXCEEDED");
+  }
+  // A partial campaign cannot quietly drop an expensive or uncertain competitor.
+  const held = blockingReasons.size > 0;
+  const ranked = held
+    ? []
+    : decisions
+        .filter((result) => result.status === "shadow-eligible")
+        .sort(
+          (a, b) =>
+            b.score - a.score ||
+            a.rawMetrics.candidateCostMicrounits -
+              b.rawMetrics.candidateCostMicrounits ||
+            candidates
+              .find((c) => c.candidateDigest === a.candidateDigest)
+              .artifacts.reduce((n, x) => n + x.bytes, 0) -
+              candidates
+                .find((c) => c.candidateDigest === b.candidateDigest)
+                .artifacts.reduce((n, x) => n + x.bytes, 0) ||
+            (a.contentDigest < b.contentDigest
+              ? -1
+              : a.contentDigest > b.contentDigest
+                ? 1
+                : 0),
+        );
+  return rrsiEnvelope(RRSI_REPLAY_SCHEMA, "replayDigest", {
+    campaignDigest: campaign.campaignDigest,
+    evidenceKind: "synthetic-offline",
+    statisticalProtocolValidated: false,
+    budgetCoverage: "synthetic-all-five-resource-aggregates",
+    blockingReasons: [...blockingReasons].sort(),
+    status: held || !ranked.length ? "hold" : "shadow-selected",
+    selectedCandidateDigest: ranked[0]?.candidateDigest ?? null,
+    decisions: decisions.sort((a, b) =>
+      a.candidateDigest < b.candidateDigest
+        ? -1
+        : a.candidateDigest > b.candidateDigest
+          ? 1
+          : 0,
+    ),
+  });
+}