@agentplat/mesh-sim 0.3.0-beta.2 → 0.3.0-beta.4

package/dist/verified-benchmark-runtime.js

@@ -0,0 +1,199 @@
+import { evaluatorMetricProjectionV2, validateBenchmarkArtifactSourceLockV1, validateBenchmarkTraceEvidenceV1, validateEvaluationParticipantDescriptorV1, validatePairedBaselineEvidenceV1, validateVerifiedBenchmarkSubmissionV1, verifiedBenchmarkDigestV1, } from "./verified-benchmark-validation.js";
+/**
+ * Pure verification: no transport, clock, mutation, service, or provider is
+ * required. Structural validation failures are ineligible rather than scored.
+ */
+export function verifyVerifiedBenchmarkSubmissionV1(input) {
+    const evidenceVerifierId = verifierId(input.evidenceVerifier);
+    let submissionDigest = null;
+    let participantDescriptorDigest = null;
+    try {
+        const submission = validateVerifiedBenchmarkSubmissionV1(input.submission, input.environmentDescriptor);
+        submissionDigest = submission.submissionDigest;
+        participantDescriptorDigest =
+            submission.participantDescriptor.descriptorDigest;
+        const reason = eligibilityReason(submission, input.suiteManifestDigest, input.evidenceVerifier);
+        const metricProjection = reason === null
+            ? evaluatorMetricProjectionV2(submission.traceEvidence)
+            : null;
+        return verification({
+            eligible: reason === null,
+            evidenceVerifierId,
+            submissionDigest,
+            participantDescriptorDigest,
+            metricProjection,
+            reason,
+        });
+    }
+    catch {
+        return verification({
+            eligible: false,
+            evidenceVerifierId,
+            submissionDigest,
+            participantDescriptorDigest,
+            metricProjection: null,
+            reason: "submission_invalid",
+        });
+    }
+}
+export function buildBenchmarkLeaderboardV1(input) {
+    if (!Array.isArray(input.submissions) || input.submissions.length > 10_000)
+        throw new TypeError("verified_benchmark_leaderboard_submissions_invalid");
+    const evaluated = input.submissions.map((submission) => ({
+        submission,
+        verification: verifyVerifiedBenchmarkSubmissionV1({
+            submission,
+            environmentDescriptor: input.environmentDescriptor,
+            suiteManifestDigest: input.suiteManifestDigest,
+            evidenceVerifier: input.evidenceVerifier,
+        }),
+    }));
+    const valid = evaluated
+        .filter(({ verification }) => verification.eligible)
+        .map(({ submission, verification }) => ({
+        submission: validateVerifiedBenchmarkSubmissionV1(submission, input.environmentDescriptor),
+        verification,
+    }))
+        .sort((left, right) => left.submission.submissionDigest.localeCompare(right.submission.submissionDigest));
+    const unique = [];
+    const duplicateRecords = [];
+    const submissionDigests = new Set();
+    const replayDigests = new Set();
+    for (const candidate of valid) {
+        const duplicate = submissionDigests.has(candidate.submission.submissionDigest);
+        const replayed = replayDigests.has(candidate.submission.replayDigest);
+        if (duplicate || replayed) {
+            duplicateRecords.push(ineligibleRecord({
+                ...candidate.verification,
+                eligible: false,
+                metricProjection: null,
+                reason: duplicate ? "duplicate_submission" : "replayed_trace",
+            }));
+            continue;
+        }
+        submissionDigests.add(candidate.submission.submissionDigest);
+        replayDigests.add(candidate.submission.replayDigest);
+        unique.push(candidate);
+    }
+    unique.sort((left, right) => right.verification.metricProjection.compositeScoreBps -
+        left.verification.metricProjection.compositeScoreBps ||
+        left.verification.participantDescriptorDigest.localeCompare(right.verification.participantDescriptorDigest) ||
+        left.verification.submissionDigest.localeCompare(right.verification.submissionDigest));
+    let previousScore = null;
+    let rank = 0;
+    const entries = unique.map(({ submission, verification: entry }, index) => {
+        if (entry.metricProjection.compositeScoreBps !== previousScore)
+            rank = index + 1;
+        previousScore = entry.metricProjection.compositeScoreBps;
+        const body = {
+            schemaVersion: 1,
+            rank,
+            submissionId: submission.submissionId,
+            participantDescriptorDigest: entry.participantDescriptorDigest,
+            metricProjection: entry.metricProjection,
+        };
+        return Object.freeze({
+            ...body,
+            entryDigest: verifiedBenchmarkDigestV1("leaderboard-entry", body),
+        });
+    });
+    const ineligible = evaluated
+        .map(({ verification }) => verification)
+        .filter((entry) => !entry.eligible)
+        .map((entry) => ineligibleRecord(entry))
+        .concat(duplicateRecords)
+        .sort((left, right) => `${left.reason}:${left.submissionDigest ?? ""}`.localeCompare(`${right.reason}:${right.submissionDigest ?? ""}`));
+    const body = {
+        schemaVersion: 1,
+        entries: Object.freeze(entries),
+        ineligibleSubmissions: Object.freeze(ineligible),
+    };
+    return Object.freeze({
+        ...body,
+        leaderboardDigest: verifiedBenchmarkDigestV1("leaderboard", body),
+    });
+}
+function eligibilityReason(submission, suiteManifestDigest, evidenceVerifier) {
+    const participant = validateEvaluationParticipantDescriptorV1(submission.participantDescriptor);
+    const artifact = validateBenchmarkArtifactSourceLockV1(submission.artifactLock);
+    const trace = validateBenchmarkTraceEvidenceV1(submission.traceEvidence);
+    const suite = submission.suiteManifest;
+    if (suite.manifestDigest !== suiteManifestDigest)
+        return "suite_binding_mismatch";
+    if (trace.participantDescriptorDigest !== participant.descriptorDigest ||
+        trace.artifactLockDigest !== artifact.lockDigest ||
+        trace.scenarioManifestDigest !== suite.scenarioManifest.manifestDigest)
+        return "trace_binding_mismatch";
+    if (trace.interactionCount >
+        suite.scenarioManifest.resourceBudget.maximumInteractions)
+        return "budget_exceeded";
+    if (!evidenceVerifier.verify({
+        role: "candidate",
+        traceEvidence: trace,
+        participantDescriptor: participant,
+        artifactLock: artifact,
+        suiteManifest: suite,
+    }))
+        return "trace_provenance_unverified";
+    if (submission.replayDigest !==
+        verifiedBenchmarkDigestV1("replay-binding", {
+            participantDescriptorDigest: participant.descriptorDigest,
+            scenarioManifestDigest: suite.scenarioManifest.manifestDigest,
+            artifactLockDigest: artifact.lockDigest,
+            traceDigest: trace.traceDigest,
+        }))
+        return "replay_binding_mismatch";
+    const baseline = validatePairedBaselineEvidenceV1(submission.pairedBaseline);
+    if (baseline.participantDescriptor.descriptorDigest !==
+        suite.baselineParticipant.descriptorDigest ||
+        baseline.artifactLock.lockDigest !== suite.baselineArtifactLock.lockDigest)
+        return "baseline_binding_mismatch";
+    if (baseline.traceEvidence.participantDescriptorDigest !==
+        suite.baselineParticipant.descriptorDigest ||
+        baseline.traceEvidence.artifactLockDigest !==
+            suite.baselineArtifactLock.lockDigest ||
+        baseline.traceEvidence.scenarioManifestDigest !==
+            suite.scenarioManifest.manifestDigest)
+        return "baseline_replay_mismatch";
+    if (baseline.traceEvidence.interactionCount >
+        suite.scenarioManifest.resourceBudget.maximumInteractions)
+        return "baseline_budget_exceeded";
+    if (!evidenceVerifier.verify({
+        role: "baseline",
+        traceEvidence: baseline.traceEvidence,
+        participantDescriptor: baseline.participantDescriptor,
+        artifactLock: baseline.artifactLock,
+        suiteManifest: suite,
+    }))
+        return "baseline_provenance_unverified";
+    return null;
+}
+function verification(input) {
+    const body = { schemaVersion: 1, ...input };
+    return Object.freeze({
+        ...body,
+        verificationDigest: verifiedBenchmarkDigestV1("verification", body),
+    });
+}
+function ineligibleRecord(entry) {
+    const body = {
+        schemaVersion: 1,
+        submissionDigest: entry.submissionDigest,
+        reason: entry.reason ?? "submission_invalid",
+    };
+    return Object.freeze({
+        ...body,
+        recordDigest: verifiedBenchmarkDigestV1("ineligible-submission", body),
+    });
+}
+function verifierId(value) {
+    if (!value ||
+        typeof value !== "object" ||
+        typeof value.verifierId !== "string" ||
+        value.verifierId.length < 1 ||
+        value.verifierId.length > 256 ||
+        typeof value.verify !== "function")
+        return "invalid";
+    return value.verifierId;
+}
+//# sourceMappingURL=verified-benchmark-runtime.js.map

package/dist/verified-benchmark-runtime.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"verified-benchmark-runtime.js","sourceRoot":"","sources":["../src/verified-benchmark-runtime.ts"],"names":[],"mappings":"AAWA,OAAO,EACL,2BAA2B,EAC3B,qCAAqC,EACrC,gCAAgC,EAChC,yCAAyC,EACzC,gCAAgC,EAChC,qCAAqC,EACrC,yBAAyB,GAC1B,MAAM,oCAAoC,CAAC;AAE5C;;;GAGG;AACH,MAAM,UAAU,mCAAmC,CAAC,KAKnD;IACC,MAAM,kBAAkB,GAAG,UAAU,CAAC,KAAK,CAAC,gBAAgB,CAAC,CAAC;IAC9D,IAAI,gBAAgB,GAA4B,IAAI,CAAC;IACrD,IAAI,2BAA2B,GAA4B,IAAI,CAAC;IAChE,IAAI,CAAC;QACH,MAAM,UAAU,GAAG,qCAAqC,CACtD,KAAK,CAAC,UAAU,EAChB,KAAK,CAAC,qBAAqB,CAC5B,CAAC;QACF,gBAAgB,GAAG,UAAU,CAAC,gBAAgB,CAAC;QAC/C,2BAA2B;YACzB,UAAU,CAAC,qBAAqB,CAAC,gBAAgB,CAAC;QACpD,MAAM,MAAM,GAAG,iBAAiB,CAC9B,UAAU,EACV,KAAK,CAAC,mBAAmB,EACzB,KAAK,CAAC,gBAAgB,CACvB,CAAC;QACF,MAAM,gBAAgB,GACpB,MAAM,KAAK,IAAI;YACb,CAAC,CAAC,2BAA2B,CAAC,UAAU,CAAC,aAAa,CAAC;YACvD,CAAC,CAAC,IAAI,CAAC;QACX,OAAO,YAAY,CAAC;YAClB,QAAQ,EAAE,MAAM,KAAK,IAAI;YACzB,kBAAkB;YAClB,gBAAgB;YAChB,2BAA2B;YAC3B,gBAAgB;YAChB,MAAM;SACP,CAAC,CAAC;IACL,CAAC;IAAC,MAAM,CAAC;QACP,OAAO,YAAY,CAAC;YAClB,QAAQ,EAAE,KAAK;YACf,kBAAkB;YAClB,gBAAgB;YAChB,2BAA2B;YAC3B,gBAAgB,EAAE,IAAI;YACtB,MAAM,EAAE,oBAAoB;SAC7B,CAAC,CAAC;IACL,CAAC;AACH,CAAC;AAED,MAAM,UAAU,2BAA2B,CAAC,KAK3C;IACC,IAAI,CAAC,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,WAAW,CAAC,IAAI,KAAK,CAAC,WAAW,CAAC,MAAM,GAAG,MAAM;QACxE,MAAM,IAAI,SAAS,CAAC,oDAAoD,CAAC,CAAC;IAC5E,MAAM,SAAS,GAAG,KAAK,CAAC,WAAW,CAAC,GAAG,CAAC,CAAC,UAAU,EAAE,EAAE,CAAC,CAAC;QACvD,UAAU;QACV,YAAY,EAAE,mCAAmC,CAAC;YAChD,UAAU;YACV,qBAAqB,EAAE,KAAK,CAAC,qBAAqB;YAClD,mBAAmB,EAAE,KAAK,CAAC,mBAAmB;YAC9C,gBAAgB,EAAE,KAAK,CAAC,gBAAgB;SACzC,CAAC;KACH,CAAC,CAAC,CAAC;IACJ,MAAM,KAAK,GAAG,SAAS;SACpB,MAAM,CAAC,CAAC,EAAE,YAAY,EAAE,EAAE,EAAE,CAAC,YAAY,CAAC,QAAQ,CAAC;SACnD,GAAG,CAAC,CAAC,EAAE,UAAU,EAAE,YAAY,EAAE,EAAE,EAAE,CAAC,CAAC;QACtC,UAAU,EAAE,qCAAqC,CAC/C,UAAU,EACV,KAAK,CAAC,qBAAqB,CAC5B;QACD,YAAY;KACb,CAAC,CAAC;SACF,IAAI,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE,EAAE,CACpB,IAAI,CAAC,UAAU,CAAC,gBAAgB,CAAC,aAAa,CAC5C,KAAK,CAAC,UAAU,CAAC,gBAAgB,CAClC,CACF,CAAC;IACJ,MAAM,MAAM,GAAiB,EAAE,CAAC;IAChC,MAAM,gBAAgB,GAA8C,EAAE,CAAC;IACvE,MAAM,iBAAiB,GAAG,IAAI,GAAG,EAAU,CAAC;IAC5C,MAAM,aAAa,GAAG,IAAI,GAAG,EAAU,CAAC;IACxC,KAAK,MAAM,SAAS,IAAI,KAAK,EAAE,CAAC;QAC9B,MAAM,SAAS,GAAG,iBAAiB,CAAC,GAAG,CACrC,SAAS,CAAC,UAAU,CAAC,gBAAgB,CACtC,CAAC;QACF,MAAM,QAAQ,GAAG,aAAa,CAAC,GAAG,CAAC,SAAS,CAAC,UAAU,CAAC,YAAY,CAAC,CAAC;QACtE,IAAI,SAAS,IAAI,QAAQ,EAAE,CAAC;YAC1B,gBAAgB,CAAC,IAAI,CACnB,gBAAgB,CAAC;gBACf,GAAG,SAAS,CAAC,YAAY;gBACzB,QAAQ,EAAE,KAAK;gBACf,gBAAgB,EAAE,IAAI;gBACtB,MAAM,EAAE,SAAS,CAAC,CAAC,CAAC,sBAAsB,CAAC,CAAC,CAAC,gBAAgB;aAC9D,CAAC,CACH,CAAC;YACF,SAAS;QACX,CAAC;QACD,iBAAiB,CAAC,GAAG,CAAC,SAAS,CAAC,UAAU,CAAC,gBAAgB,CAAC,CAAC;QAC7D,aAAa,CAAC,GAAG,CAAC,SAAS,CAAC,UAAU,CAAC,YAAY,CAAC,CAAC;QACrD,MAAM,CAAC,IAAI,CAAC,SAAS,CAAC,CAAC;IACzB,CAAC;IACD,MAAM,CAAC,IAAI,CACT,CAAC,IAAI,EAAE,KAAK,EAAE,EAAE,CACd,KAAK,CAAC,YAAY,CAAC,gBAAiB,CAAC,iBAAiB;QACpD,IAAI,CAAC,YAAY,CAAC,gBAAiB,CAAC,iBAAiB;QACvD,IAAI,CAAC,YAAY,CAAC,2BAA4B,CAAC,aAAa,CAC1D,KAAK,CAAC,YAAY,CAAC,2BAA4B,CAChD;QACD,IAAI,CAAC,YAAY,CAAC,gBAAiB,CAAC,aAAa,CAC/C,KAAK,CAAC,YAAY,CAAC,gBAAiB,CACrC,CACJ,CAAC;IACF,IAAI,aAAa,GAAkB,IAAI,CAAC;IACxC,IAAI,IAAI,GAAG,CAAC,CAAC;IACb,MAAM,OAAO,GAAG,MAAM,CAAC,GAAG,CAAC,CAAC,EAAE,UAAU,EAAE,YAAY,EAAE,KAAK,EAAE,EAAE,KAAK,EAAE,EAAE;QACxE,IAAI,KAAK,CAAC,gBAAiB,CAAC,iBAAiB,KAAK,aAAa;YAC7D,IAAI,GAAG,KAAK,GAAG,CAAC,CAAC;QACnB,aAAa,GAAG,KAAK,CAAC,gBAAiB,CAAC,iBAAiB,CAAC;QAC1D,MAAM,IAAI,GAAG;YACX,aAAa,EAAE,CAAU;YACzB,IAAI;YACJ,YAAY,EAAE,UAAU,CAAC,YAAY;YACrC,2BAA2B,EAAE,KAAK,CAAC,2BAA4B;YAC/D,gBAAgB,EAAE,KAAK,CAAC,gBAAiB;SAC1C,CAAC;QACF,OAAO,MAAM,CAAC,MAAM,CAAC;YACnB,GAAG,IAAI;YACP,WAAW,EAAE,yBAAyB,CAAC,mBAAmB,EAAE,IAAI,CAAC;SAClE,CAAwC,CAAC;IAC5C,CAAC,CAAC,CAAC;IACH,MAAM,UAAU,GAAG,SAAS;SACzB,GAAG,CAAC,CAAC,EAAE,YAAY,EAAE,EAAE,EAAE,CAAC,YAAY,CAAC;SACvC,MAAM,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,CAAC,KAAK,CAAC,QAAQ,CAAC;SAClC,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,gBAAgB,CAAC,KAAK,CAAC,CAAC;SACvC,MAAM,CAAC,gBAAgB,CAAC;SACxB,IAAI,CAAC,CAAC,IAAI,EAAE,KAAK,EAAE,EAAE,CACpB,GAAG,IAAI,CAAC,MAAM,IAAI,IAAI,CAAC,gBAAgB,IAAI,EAAE,EAAE,CAAC,aAAa,CAC3D,GAAG,KAAK,CAAC,MAAM,IAAI,KAAK,CAAC,gBAAgB,IAAI,EAAE,EAAE,CAClD,CACF,CAAC;IACJ,MAAM,IAAI,GAAG;QACX,aAAa,EAAE,CAAU;QACzB,OAAO,EAAE,MAAM,CAAC,MAAM,CAAC,OAAO,CAAC;QAC/B,qBAAqB,EAAE,MAAM,CAAC,MAAM,CAAC,UAAU,CAAC;KACjD,CAAC;IACF,OAAO,MAAM,CAAC,MAAM,CAAC;QACnB,GAAG,IAAI;QACP,iBAAiB,EAAE,yBAAyB,CAAC,aAAa,EAAE,IAAI,CAAC;KAClE,CAAC,CAAC;AACL,CAAC;AAED,SAAS,iBAAiB,CACxB,UAAyC,EACzC,mBAAqC,EACrC,gBAAkD;IAElD,MAAM,WAAW,GAAG,yCAAyC,CAC3D,UAAU,CAAC,qBAAqB,CACjC,CAAC;IACF,MAAM,QAAQ,GAAG,qCAAqC,CACpD,UAAU,CAAC,YAAY,CACxB,CAAC;IACF,MAAM,KAAK,GAAG,gCAAgC,CAAC,UAAU,CAAC,aAAa,CAAC,CAAC;IACzE,MAAM,KAAK,GAAG,UAAU,CAAC,aAAa,CAAC;IACvC,IAAI,KAAK,CAAC,cAAc,KAAK,mBAAmB;QAC9C,OAAO,wBAAwB,CAAC;IAClC,IACE,KAAK,CAAC,2BAA2B,KAAK,WAAW,CAAC,gBAAgB;QAClE,KAAK,CAAC,kBAAkB,KAAK,QAAQ,CAAC,UAAU;QAChD,KAAK,CAAC,sBAAsB,KAAK,KAAK,CAAC,gBAAgB,CAAC,cAAc;QAEtE,OAAO,wBAAwB,CAAC;IAClC,IACE,KAAK,CAAC,gBAAgB;QACtB,KAAK,CAAC,gBAAgB,CAAC,cAAc,CAAC,mBAAmB;QAEzD,OAAO,iBAAiB,CAAC;IAC3B,IACE,CAAC,gBAAgB,CAAC,MAAM,CAAC;QACvB,IAAI,EAAE,WAAW;QACjB,aAAa,EAAE,KAAK;QACpB,qBAAqB,EAAE,WAAW;QAClC,YAAY,EAAE,QAAQ;QACtB,aAAa,EAAE,KAAK;KACrB,CAAC;QAEF,OAAO,6BAA6B,CAAC;IACvC,IACE,UAAU,CAAC,YAAY;QACvB,yBAAyB,CAAC,gBAAgB,EAAE;YAC1C,2BAA2B,EAAE,WAAW,CAAC,gBAAgB;YACzD,sBAAsB,EAAE,KAAK,CAAC,gBAAgB,CAAC,cAAc;YAC7D,kBAAkB,EAAE,QAAQ,CAAC,UAAU;YACvC,WAAW,EAAE,KAAK,CAAC,WAAW;SAC/B,CAAC;QAEF,OAAO,yBAAyB,CAAC;IACnC,MAAM,QAAQ,GAAG,gCAAgC,CAAC,UAAU,CAAC,cAAc,CAAC,CAAC;IAC7E,IACE,QAAQ,CAAC,qBAAqB,CAAC,gBAAgB;QAC7C,KAAK,CAAC,mBAAmB,CAAC,gBAAgB;QAC5C,QAAQ,CAAC,YAAY,CAAC,UAAU,KAAK,KAAK,CAAC,oBAAoB,CAAC,UAAU;QAE1E,OAAO,2BAA2B,CAAC;IACrC,IACE,QAAQ,CAAC,aAAa,CAAC,2BAA2B;QAChD,KAAK,CAAC,mBAAmB,CAAC,gBAAgB;QAC5C,QAAQ,CAAC,aAAa,CAAC,kBAAkB;YACvC,KAAK,CAAC,oBAAoB,CAAC,UAAU;QACvC,QAAQ,CAAC,aAAa,CAAC,sBAAsB;YAC3C,KAAK,CAAC,gBAAgB,CAAC,cAAc;QAEvC,OAAO,0BAA0B,CAAC;IACpC,IACE,QAAQ,CAAC,aAAa,CAAC,gBAAgB;QACvC,KAAK,CAAC,gBAAgB,CAAC,cAAc,CAAC,mBAAmB;QAEzD,OAAO,0BAA0B,CAAC;IACpC,IACE,CAAC,gBAAgB,CAAC,MAAM,CAAC;QACvB,IAAI,EAAE,UAAU;QAChB,aAAa,EAAE,QAAQ,CAAC,aAAa;QACrC,qBAAqB,EAAE,QAAQ,CAAC,qBAAqB;QACrD,YAAY,EAAE,QAAQ,CAAC,YAAY;QACnC,aAAa,EAAE,KAAK;KACrB,CAAC;QAEF,OAAO,gCAAgC,CAAC;IAC1C,OAAO,IAAI,CAAC;AACd,CAAC;AAED,SAAS,YAAY,CACnB,KAGC;IAED,MAAM,IAAI,GAAG,EAAE,aAAa,EAAE,CAAU,EAAE,GAAG,KAAK,EAAE,CAAC;IACrD,OAAO,MAAM,CAAC,MAAM,CAAC;QACnB,GAAG,IAAI;QACP,kBAAkB,EAAE,yBAAyB,CAAC,cAAc,EAAE,IAAI,CAAC;KACpE,CAAC,CAAC;AACL,CAAC;AACD,SAAS,gBAAgB,CACvB,KAAsC;IAEtC,MAAM,IAAI,GAAG;QACX,aAAa,EAAE,CAAU;QACzB,gBAAgB,EAAE,KAAK,CAAC,gBAAgB;QACxC,MAAM,EAAE,KAAK,CAAC,MAAM,IAAI,oBAAoB;KAC7C,CAAC;IACF,OAAO,MAAM,CAAC,MAAM,CAAC;QACnB,GAAG,IAAI;QACP,YAAY,EAAE,yBAAyB,CAAC,uBAAuB,EAAE,IAAI,CAAC;KACvE,CAAC,CAAC;AACL,CAAC;AACD,SAAS,UAAU,CAAC,KAAuC;IACzD,IACE,CAAC,KAAK;QACN,OAAO,KAAK,KAAK,QAAQ;QACzB,OAAO,KAAK,CAAC,UAAU,KAAK,QAAQ;QACpC,KAAK,CAAC,UAAU,CAAC,MAAM,GAAG,CAAC;QAC3B,KAAK,CAAC,UAAU,CAAC,MAAM,GAAG,GAAG;QAC7B,OAAO,KAAK,CAAC,MAAM,KAAK,UAAU;QAElC,OAAO,SAAS,CAAC;IACnB,OAAO,KAAK,CAAC,UAAU,CAAC;AAC1B,CAAC"}

package/dist/verified-benchmark-validation.d.ts

@@ -0,0 +1,64 @@
+import { type PlanningDigestV1 } from "@agentplat/collective-planning";
+import type { MultiDomainEnvironmentDescriptorV1, MultiDomainScenarioManifestV1 } from "./multi-domain-environment-contracts.js";
+import type { BenchmarkArtifactSourceLockV1, BenchmarkSuiteManifestV1, BenchmarkTraceEvidenceV1, EvaluationParticipantDescriptorV1, MetricProjectionV2, PairedBaselineEvidenceV1, VerifiedBenchmarkSubmissionV1 } from "./verified-benchmark-contracts.js";
+export declare function verifiedBenchmarkDigestV1(domain: string, value: unknown): PlanningDigestV1;
+export declare function createEvaluationParticipantDescriptorV1(input: {
+    readonly adapterId: string;
+    readonly adapterVersion: number;
+    readonly coordinationModel: "centralized" | "collective";
+    readonly implementationDigest: PlanningDigestV1;
+    readonly capabilities: readonly string[];
+}): EvaluationParticipantDescriptorV1;
+export declare function validateEvaluationParticipantDescriptorV1(input: unknown): EvaluationParticipantDescriptorV1;
+export declare function createBenchmarkArtifactSourceLockV1(input: {
+    readonly sourceId: string;
+    readonly sourceDigest: PlanningDigestV1;
+    readonly artifactId: string;
+    readonly artifactDigest: PlanningDigestV1;
+    readonly buildDigest: PlanningDigestV1;
+}): BenchmarkArtifactSourceLockV1;
+export declare function validateBenchmarkArtifactSourceLockV1(input: unknown): BenchmarkArtifactSourceLockV1;
+export declare function createBenchmarkSuiteManifestV1(input: {
+    readonly suiteId: string;
+    readonly suiteVersion: number;
+    readonly environmentDescriptor: MultiDomainEnvironmentDescriptorV1;
+    readonly scenarioManifest: MultiDomainScenarioManifestV1;
+    readonly baselineParticipant: EvaluationParticipantDescriptorV1;
+    readonly baselineArtifactLock: BenchmarkArtifactSourceLockV1;
+}): BenchmarkSuiteManifestV1;
+export declare function validateBenchmarkSuiteManifestV1(input: unknown, environmentDescriptor: MultiDomainEnvironmentDescriptorV1): BenchmarkSuiteManifestV1;
+export declare function createBenchmarkTraceEvidenceV1(input: {
+    readonly traceId: string;
+    readonly participantDescriptor: EvaluationParticipantDescriptorV1;
+    readonly scenarioManifest: MultiDomainScenarioManifestV1;
+    readonly artifactLock: BenchmarkArtifactSourceLockV1;
+    readonly interactionCount: number;
+    readonly successfulObjectiveCount: number;
+    readonly recoveryToBaselineCount: number;
+    readonly roleCoherenceLength: number;
+}): BenchmarkTraceEvidenceV1;
+export declare function validateBenchmarkTraceEvidenceV1(input: unknown): BenchmarkTraceEvidenceV1;
+export declare function evaluatorMetricProjectionV2(traceEvidence: BenchmarkTraceEvidenceV1): MetricProjectionV2;
+export declare function benchmarkReplayDigestV1(input: {
+    readonly participantDescriptorDigest: PlanningDigestV1;
+    readonly scenarioManifestDigest: PlanningDigestV1;
+    readonly artifactLockDigest: PlanningDigestV1;
+    readonly traceDigest: PlanningDigestV1;
+}): PlanningDigestV1;
+export declare function createVerifiedBenchmarkSubmissionV1(input: {
+    readonly submissionId: string;
+    readonly participantDescriptor: EvaluationParticipantDescriptorV1;
+    readonly artifactLock: BenchmarkArtifactSourceLockV1;
+    readonly suiteManifest: BenchmarkSuiteManifestV1;
+    readonly traceEvidence: BenchmarkTraceEvidenceV1;
+    readonly pairedBaseline: PairedBaselineEvidenceV1;
+}): VerifiedBenchmarkSubmissionV1;
+export declare function validateVerifiedBenchmarkSubmissionV1(input: unknown, environmentDescriptor: MultiDomainEnvironmentDescriptorV1): VerifiedBenchmarkSubmissionV1;
+export declare function createPairedBaselineEvidenceV1(input: {
+    readonly participantDescriptor: EvaluationParticipantDescriptorV1;
+    readonly artifactLock: BenchmarkArtifactSourceLockV1;
+    readonly scenarioManifest: MultiDomainScenarioManifestV1;
+    readonly traceEvidence: BenchmarkTraceEvidenceV1;
+}): PairedBaselineEvidenceV1;
+export declare function validatePairedBaselineEvidenceV1(input: unknown): PairedBaselineEvidenceV1;
+//# sourceMappingURL=verified-benchmark-validation.d.ts.map

package/dist/verified-benchmark-validation.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"verified-benchmark-validation.d.ts","sourceRoot":"","sources":["../src/verified-benchmark-validation.ts"],"names":[],"mappings":"AAAA,OAAO,EAEL,KAAK,gBAAgB,EAEtB,MAAM,gCAAgC,CAAC;AAExC,OAAO,KAAK,EACV,kCAAkC,EAClC,6BAA6B,EAC9B,MAAM,yCAAyC,CAAC;AAUjD,OAAO,KAAK,EACV,6BAA6B,EAC7B,wBAAwB,EACxB,wBAAwB,EACxB,iCAAiC,EACjC,kBAAkB,EAClB,wBAAwB,EACxB,6BAA6B,EAC9B,MAAM,mCAAmC,CAAC;AAQ3C,wBAAgB,yBAAyB,CACvC,MAAM,EAAE,MAAM,EACd,KAAK,EAAE,OAAO,GACb,gBAAgB,CAElB;AAED,wBAAgB,uCAAuC,CAAC,KAAK,EAAE;IAC7D,QAAQ,CAAC,SAAS,EAAE,MAAM,CAAC;IAC3B,QAAQ,CAAC,cAAc,EAAE,MAAM,CAAC;IAChC,QAAQ,CAAC,iBAAiB,EAAE,aAAa,GAAG,YAAY,CAAC;IACzD,QAAQ,CAAC,oBAAoB,EAAE,gBAAgB,CAAC;IAChD,QAAQ,CAAC,YAAY,EAAE,SAAS,MAAM,EAAE,CAAC;CAC1C,GAAG,iCAAiC,CAyCpC;AAED,wBAAgB,yCAAyC,CACvD,KAAK,EAAE,OAAO,GACb,iCAAiC,CA4BnC;AAED,wBAAgB,mCAAmC,CAAC,KAAK,EAAE;IACzD,QAAQ,CAAC,QAAQ,EAAE,MAAM,CAAC;IAC1B,QAAQ,CAAC,YAAY,EAAE,gBAAgB,CAAC;IACxC,QAAQ,CAAC,UAAU,EAAE,MAAM,CAAC;IAC5B,QAAQ,CAAC,cAAc,EAAE,gBAAgB,CAAC;IAC1C,QAAQ,CAAC,WAAW,EAAE,gBAAgB,CAAC;CACxC,GAAG,6BAA6B,CA2BhC;AAED,wBAAgB,qCAAqC,CACnD,KAAK,EAAE,OAAO,GACb,6BAA6B,CA4B/B;AAED,wBAAgB,8BAA8B,CAAC,KAAK,EAAE;IACpD,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAC;IACzB,QAAQ,CAAC,YAAY,EAAE,MAAM,CAAC;IAC9B,QAAQ,CAAC,qBAAqB,EAAE,kCAAkC,CAAC;IACnE,QAAQ,CAAC,gBAAgB,EAAE,6BAA6B,CAAC;IACzD,QAAQ,CAAC,mBAAmB,EAAE,iCAAiC,CAAC;IAChE,QAAQ,CAAC,oBAAoB,EAAE,6BAA6B,CAAC;CAC9D,GAAG,wBAAwB,CAqD3B;AAED,wBAAgB,gCAAgC,CAC9C,KAAK,EAAE,OAAO,EACd,qBAAqB,EAAE,kCAAkC,GACxD,wBAAwB,CAsC1B;AAED,wBAAgB,8BAA8B,CAAC,KAAK,EAAE;IACpD,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAC;IACzB,QAAQ,CAAC,qBAAqB,EAAE,iCAAiC,CAAC;IAClE,QAAQ,CAAC,gBAAgB,EAAE,6BAA6B,CAAC;IACzD,QAAQ,CAAC,YAAY,EAAE,6BAA6B,CAAC;IACrD,QAAQ,CAAC,gBAAgB,EAAE,MAAM,CAAC;IAClC,QAAQ,CAAC,wBAAwB,EAAE,MAAM,CAAC;IAC1C,QAAQ,CAAC,uBAAuB,EAAE,MAAM,CAAC;IACzC,QAAQ,CAAC,mBAAmB,EAAE,MAAM,CAAC;CACtC,GAAG,wBAAwB,CAqD3B;AAED,wBAAgB,gCAAgC,CAC9C,KAAK,EAAE,OAAO,GACb,wBAAwB,CAoD1B;AAED,wBAAgB,2BAA2B,CACzC,aAAa,EAAE,wBAAwB,GACtC,kBAAkB,CA8BpB;AAED,wBAAgB,uBAAuB,CAAC,KAAK,EAAE;IAC7C,QAAQ,CAAC,2BAA2B,EAAE,gBAAgB,CAAC;IACvD,QAAQ,CAAC,sBAAsB,EAAE,gBAAgB,CAAC;IAClD,QAAQ,CAAC,kBAAkB,EAAE,gBAAgB,CAAC;IAC9C,QAAQ,CAAC,WAAW,EAAE,gBAAgB,CAAC;CACxC,GAAG,gBAAgB,CAenB;AAED,wBAAgB,mCAAmC,CAAC,KAAK,EAAE;IACzD,QAAQ,CAAC,YAAY,EAAE,MAAM,CAAC;IAC9B,QAAQ,CAAC,qBAAqB,EAAE,iCAAiC,CAAC;IAClE,QAAQ,CAAC,YAAY,EAAE,6BAA6B,CAAC;IACrD,QAAQ,CAAC,aAAa,EAAE,wBAAwB,CAAC;IACjD,QAAQ,CAAC,aAAa,EAAE,wBAAwB,CAAC;IACjD,QAAQ,CAAC,cAAc,EAAE,wBAAwB,CAAC;CACnD,GAAG,6BAA6B,CAgDhC;AAED,wBAAgB,qCAAqC,CACnD,KAAK,EAAE,OAAO,EACd,qBAAqB,EAAE,kCAAkC,GACxD,6BAA6B,CAmC/B;AAED,wBAAgB,8BAA8B,CAAC,KAAK,EAAE;IACpD,QAAQ,CAAC,qBAAqB,EAAE,iCAAiC,CAAC;IAClE,QAAQ,CAAC,YAAY,EAAE,6BAA6B,CAAC;IACrD,QAAQ,CAAC,gBAAgB,EAAE,6BAA6B,CAAC;IACzD,QAAQ,CAAC,aAAa,EAAE,wBAAwB,CAAC;CAClD,GAAG,wBAAwB,CAsC3B;AAED,wBAAgB,gCAAgC,CAC9C,KAAK,EAAE,OAAO,GACb,wBAAwB,CAqC1B"}