@tangle-network/agent-eval 0.221.7 → 0.222.0

package/dist/{campaign-evidence-DunzcJGv.js → campaign-evidence-DN8rWuBj.js}

@@ -1,14 +1,15 @@
-import { t as AgentEvalError } from "./errors-BcOs5MKW.js";
-import { a as hashCanonical, i as compareCodeUnits, r as canonicalString } from "./canonical-Dq2hOQ8L.js";
-import { n as aggregatePairedHoldout, o as pairHoldout } from "./statistical-heldout-Bksv92Th.js";
-import { n as contentHash } from "./verdict-cache-CL_7tePD.js";
-import { r as projectCampaignCellQuality } from "./run-record-CysKLB93.js";
-import { t as FileLedgerJournal } from "./journal-C7om8Gum.js";
-import { t as defineEvaluationClaim } from "./claim-BPVuyHuy.js";
+import { t as AgentEvalError } from "./errors-DeEEV9Vx.js";
+import { a as hashCanonical, i as compareCodeUnits, r as canonicalString } from "./canonical-CicJlINA.js";
+import { n as aggregatePairedHoldout, o as pairHoldout } from "./statistical-heldout-TP6ndL0B.js";
+import { t as FileLedgerJournal } from "./journal-C1uLJi0w.js";
+import { n as contentHash } from "./verdict-cache-BVH-fl18.js";
+import { r as projectCampaignCellQuality } from "./run-record-CZy4-TnZ.js";
+import { l as surfaceHash, o as renderSurfaceDiff, s as surfaceContentHash, t as autoCloseTruncatedJson, u as surfaceHashMatches } from "./json-recovery-DK4MDm5r.js";
+import { t as defineEvaluationClaim } from "./claim-Dod_mjk3.js";
 import { createHash } from "node:crypto";
 import { z } from "zod";
 import { join } from "node:path";
-//#region src/campaign/coverage.ts
+// src/campaign/coverage.ts
 /** Reject campaign designs whose denominator cannot be identified exactly. */
 function assertCampaignDesign(scenarios, reps) {
 	if (!Number.isSafeInteger(reps) || reps < 1) throw new Error("campaign design requires reps to be a positive safe integer");
@@ -135,8 +136,7 @@ function assertCompleteCampaign(campaign, scenarios, reps, requireJudgeScore, la
 	const coverage = campaignCoverage(campaign.cells, scenarios, reps, requireJudgeScore);
 	if (!coverage.complete) throw new Error(`${label} is incomplete (${coverage.scorableCellIds.length}/${coverage.expectedCellIds.length} designed cells scorable) — ${formatCoverageFailures(coverage)}. Refusing to compare unequal results.`);
 }
-//#endregion
-//#region src/integrity/backend-integrity.ts
+// src/integrity/backend-integrity.ts
 /**
 * Error thrown when an integrity assertion fails. Caller can pattern-match
 * by `code === 'AGENT_EVAL_BACKEND_STUB'` to differentiate from other
@@ -252,158 +252,7 @@ function assertBackendReport(report, opts) {
 	if (!allowMixed && report.verdict === "mixed") throw new BackendIntegrityError(`backend-integrity: partial backend failure rejected — ${report.diagnosis}`, report);
 	return report;
 }
-//#endregion
-//#region src/campaign/surface-identity.ts
-const GIT_OBJECT_ID = /^(?:[a-f0-9]{40}|[a-f0-9]{64})$/;
-const SHA256 = /^sha256:[a-f0-9]{64}$/;
-/** Validate the immutable identity shape; the owning executor verifies the Git objects and patch. */
-function assertCodeSurfaceIdentity(surface) {
-	if (!surface || typeof surface !== "object") throw new TypeError("CodeSurface must be an object");
-	const candidate = surface;
-	if (candidate.kind !== "code") throw new TypeError("CodeSurface.kind must be \"code\"");
-	if (typeof candidate.worktreeRef !== "string" || candidate.worktreeRef.trim().length === 0) throw new TypeError("CodeSurface.worktreeRef must be a non-empty locator");
-	if (typeof candidate.baseRef !== "string" || candidate.baseRef.trim().length === 0) throw new TypeError("CodeSurface.baseRef must be a non-empty ref label");
-	for (const [field, value] of [
-		["baseCommit", candidate.baseCommit],
-		["baseTree", candidate.baseTree],
-		["candidateCommit", candidate.candidateCommit],
-		["candidateTree", candidate.candidateTree]
-	]) if (typeof value !== "string" || !GIT_OBJECT_ID.test(value)) throw new TypeError(`CodeSurface.${field} must be a full Git object id`);
-	const patch = candidate.patch;
-	if (!patch || typeof patch !== "object" || patch.format !== "git-diff-binary") throw new TypeError("CodeSurface.patch.format must be \"git-diff-binary\"");
-	if (typeof patch.sha256 !== "string" || !SHA256.test(patch.sha256)) throw new TypeError("CodeSurface.patch.sha256 must be a sha256 digest");
-	if (!Number.isSafeInteger(patch.byteLength) || patch.byteLength < 0) throw new TypeError("CodeSurface.patch.byteLength must be a non-negative safe integer");
-}
-/** Assert that a value is a valid non-empty component surface. */
-function assertComponentSurface(surface) {
-	if (!surface || typeof surface !== "object") throw new TypeError("ComponentSurface must be an object");
-	const candidate = surface;
-	if (candidate.kind !== "components") throw new TypeError("ComponentSurface.kind must be \"components\"");
-	if (!candidate.components || typeof candidate.components !== "object" || Array.isArray(candidate.components)) throw new TypeError("ComponentSurface.components must be an object");
-	const entries = Object.entries(candidate.components);
-	if (entries.length === 0) throw new TypeError("ComponentSurface.components must not be empty");
-	for (const [name, content] of entries) {
-		if (!name.trim() || name.trim() !== name) throw new TypeError("ComponentSurface component names must be trimmed and non-empty");
-		if (typeof content !== "string") throw new TypeError(`ComponentSurface component '${name}' must be a string`);
-	}
-}
-/**
-* Deterministic identity material for a component surface.
-*
-* `canonicalString` orders keys by UTF-16 code unit (RFC 8785), which is a
-* property of the value alone. The previous material ordered them with
-* `localeCompare`, which reads the host's collation — so the same surface
-* could produce two different identities on two machines, and the stored
-* identity would stop matching a recomputation of the identical surface.
-*/
-function componentSurfaceIdentityMaterial(surface) {
-	assertComponentSurface(surface);
-	return canonicalString({
-		schema: "tangle.component-surface",
-		components: surface.components
-	});
-}
-/**
-* The retired material builder, kept PRIVATE and reachable only from
-* {@link surfaceHashMatches}.
-*
-* A surface identity recorded before this release was minted from these bytes.
-* The verify path tries the current material first and falls back to this one,
-* so a stored identity still matches its own surface; nothing mints from it.
-*
-* Every component surface's identity moves, not only one whose names sort
-* differently under the host's collation: RFC 8785 also orders the two
-* top-level keys, so `components` precedes `schema` where this builder emitted
-* them in literal order. The retention window therefore covers every stored
-* component-surface identity, which is why this builder is kept rather than
-* scoped to the mixed-case case.
-*/
-function retiredComponentSurfaceIdentityMaterial(surface) {
-	assertComponentSurface(surface);
-	return JSON.stringify({
-		schema: "tangle.component-surface",
-		components: Object.fromEntries(Object.entries(surface.components).sort(([left], [right]) => left.localeCompare(right)))
-	});
-}
-/** Canonical, location-independent identity of a finalized code candidate.
-*  Commit metadata is excluded: two commits with the same base, final tree,
-*  and patch bytes are the same executable candidate. */
-function codeSurfaceIdentityMaterial(surface) {
-	assertCodeSurfaceIdentity(surface);
-	return JSON.stringify({
-		schema: "tangle.code-surface",
-		baseCommit: surface.baseCommit,
-		baseTree: surface.baseTree,
-		candidateTree: surface.candidateTree,
-		patch: {
-			format: surface.patch.format,
-			sha256: surface.patch.sha256,
-			byteLength: surface.patch.byteLength
-		}
-	});
-}
-/** Full SHA-256 content identity for a prompt or finalized code surface. */
-function surfaceContentHash(surface) {
-	const material = typeof surface === "string" ? surface : surface.kind === "components" ? componentSurfaceIdentityMaterial(surface) : codeSurfaceIdentityMaterial(surface);
-	return `sha256:${createHash("sha256").update(material).digest("hex")}`;
-}
-/** Short loop key derived from the same content identity as provenance. */
-function surfaceHash(surface) {
-	return surfaceContentHash(surface).slice(7, 23);
-}
-/**
-* Whether `storedHash` is the loop key of `surface`, under the current identity
-* material or the retired one.
-*
-* A stored key is 16 hex characters with no room for a scheme tag, so the
-* scheme cannot be read off the value the way an `agent-profile-cell` id names
-* its own. The verify path therefore tries both, which gives the same property:
-* a key minted by an earlier release still matches its own surface, and a
-* surface that was actually edited matches neither.
-*
-* Only a component surface can differ between the two; a prompt or code surface
-* produces identical material under both, so the second comparison is a no-op
-* for them.
-*/
-function surfaceHashMatches(surface, storedHash) {
-	if (surfaceHash(surface) === storedHash) return true;
-	if (typeof surface === "string" || surface.kind !== "components") return false;
-	return createHash("sha256").update(retiredComponentSurfaceIdentityMaterial(surface)).digest("hex").slice(0, 16) === storedHash;
-}
-/** Canonical customer-visible description of the exact before/after surfaces. */
-function renderSurfaceDiff(winnerSurface, baselineSurface) {
-	if (typeof winnerSurface === "string" && typeof baselineSurface === "string") return [
-		"--- baseline",
-		"+++ winner",
-		...baselineSurface.split("\n").map((line) => `- ${line}`),
-		...winnerSurface.split("\n").map((line) => `+ ${line}`)
-	].join("\n");
-	const describe = (surface) => {
-		if (typeof surface === "string") return "(prompt surface)";
-		if (surface.kind === "components") {
-			assertComponentSurface(surface);
-			return Object.entries(surface.components).sort(([left], [right]) => left.localeCompare(right)).map(([name, content]) => `[${name}]\n${content}`).join("\n\n");
-		}
-		assertCodeSurfaceIdentity(surface);
-		return [
-			`baseCommit=${surface.baseCommit}`,
-			`baseTree=${surface.baseTree}`,
-			`candidateCommit=${surface.candidateCommit}`,
-			`candidateTree=${surface.candidateTree}`,
-			`patch=${surface.patch.sha256}`,
-			`patchBytes=${surface.patch.byteLength}`,
-			...surface.summary ? [surface.summary] : []
-		].join("\n");
-	};
-	return `--- baseline\n${describe(baselineSurface)}\n+++ winner\n${describe(winnerSurface)}`;
-}
-/** Bind a campaign cache entry to the exact surface and caller-owned execution revision. */
-function surfaceDispatchRef(surface, executionRef = "anonymous") {
-	if (!executionRef.trim() || executionRef.trim() !== executionRef) throw new Error("surfaceDispatchRef: executionRef must be trimmed and non-empty");
-	return `surface:${executionRef}:${surfaceContentHash(surface)}`;
-}
-//#endregion
-//#region src/experiment/final-evidence.ts
+// src/experiment/final-evidence.ts
 const identity = z.string().min(1).refine((value) => value.trim() === value);
 const digest = z.string().regex(/^sha256:[a-f0-9]{64}$/);
 const reservationSchema = z.object({
@@ -579,8 +428,7 @@ function openFinalEvidenceLedger(options) {
 		read: () => outcome(() => journal.replay())
 	};
 }
-//#endregion
-//#region src/campaign/gates/power-preflight.ts
+// src/campaign/gates/power-preflight.ts
 /** Two-sided z for the common confidence levels; interpolation is overkill here. */
 function zFor(confidence) {
 	if (confidence >= .99) return 2.576;
@@ -620,151 +468,7 @@ function powerPreflight(opts) {
 		recommendation: sharedChannelCaveat ? `${recommendation} ${sharedChannelCaveat}` : recommendation
 	};
 }
-//#endregion
-//#region src/json-recovery.ts
-/**
-* Truncation-tolerant JSON recovery — shared by every parser that reads JSON
-* out of a model response (reflective-mutation proposals, judge scores, the
-* completion-correctness checker).
-*
-* LLMs routinely hit a max_tokens cap mid-emission, leaving a JSON prefix
-* with an unclosed string / object / array and often a dangling key or
-* trailing comma. Throwing on that prefix — and letting the throw fold into
-* a fabricated zero score downstream — is the bug class this module exists
-* to prevent (see `JudgeParseError`'s contract: a synthetic zero is
-* indistinguishable from a real low score). Recovering the complete prefix
-* turns a would-be fabricated zero into a real measurement.
-*/
-/**
-* Walk the input as JSON-aware (string vs not, escape-aware) and close
-* unclosed `{` / `[` in LIFO order at the tail. If the input was already
-* balanced returns it unchanged. If a string was open at end-of-input we
-* also close it with `"` first, since a truncated string-mid-value is the
-* most common LLM cap-hit failure mode and JSON.parse cannot proceed
-* without one.
-*
-* Returns null when the structure is unrecoverable (e.g. depth would go
-* negative — that's an *over*-closed prefix, not a truncation).
-*/
-function autoCloseTruncatedJson(raw) {
-	const stack = [];
-	let inString = false;
-	let escaped = false;
-	for (const c of raw) {
-		if (escaped) {
-			escaped = false;
-			continue;
-		}
-		if (inString) {
-			if (c === "\\") {
-				escaped = true;
-				continue;
-			}
-			if (c === "\"") {
-				inString = false;
-				continue;
-			}
-			continue;
-		}
-		if (c === "\"") {
-			inString = true;
-			continue;
-		}
-		if (c === "{" || c === "[") stack.push(c);
-		else if (c === "}") {
-			if (stack.pop() !== "{") return null;
-		} else if (c === "]") {
-			if (stack.pop() !== "[") return null;
-		}
-	}
-	if (stack.length === 0 && !inString) return raw;
-	let suffix = "";
-	if (escaped) suffix += "\\";
-	if (inString) suffix += "\"";
-	while (stack.length > 0) {
-		const opener = stack.pop();
-		suffix += opener === "{" ? "}" : "]";
-	}
-	return raw + suffix;
-}
-const UNPARSEABLE = Symbol("unparseable");
-function tryParse(candidate) {
-	try {
-		return JSON.parse(candidate);
-	} catch {
-		return UNPARSEABLE;
-	}
-}
-/** Index of the last `,` that sits outside any string literal, or -1. Cutting
-*  there discards a dangling key / half-emitted value at the tail while
-*  keeping every complete member before it. */
-function lastCommaOutsideString(s) {
-	let inString = false;
-	let escaped = false;
-	let last = -1;
-	for (let i = 0; i < s.length; i++) {
-		const c = s[i];
-		if (escaped) {
-			escaped = false;
-			continue;
-		}
-		if (inString) {
-			if (c === "\\") escaped = true;
-			else if (c === "\"") inString = false;
-			continue;
-		}
-		if (c === "\"") inString = true;
-		else if (c === ",") last = i;
-	}
-	return last;
-}
-/**
-* Best-effort parse of a possibly-truncated JSON payload embedded in model
-* output (prose and markdown fences tolerated). Tries each opener position
-* (earliest `{`/`[` first, then the other when the first yields nothing —
-* prose like `[note] {"score":3}` must not poison the slice), and from each
-* start, in order:
-*
-*   1. plain `JSON.parse` of the opener → last-closer slice,
-*   2. auto-closing unclosed structures at the tail
-*      (`autoCloseTruncatedJson`),
-*   3. trimming the tail back to the previous complete member boundary (the
-*      last comma outside a string) and auto-closing again, repeatedly.
-*
-* Recovers e.g. `{"correct": false, "` → `{ correct: false }` — the exact
-* cap-hit shape that has zeroed real eval rows. Returns the parsed value
-* (always an object or array, given the slice starts at an opener), or
-* `null` when nothing parseable can be recovered. Never throws.
-*/
-function recoverTruncatedJson(text) {
-	const starts = [text.indexOf("{"), text.indexOf("[")].filter((i) => i >= 0).sort((a, b) => a - b);
-	for (const start of starts) {
-		const recovered = recoverFrom(text.slice(start));
-		if (recovered !== UNPARSEABLE) return recovered;
-	}
-	return null;
-}
-function recoverFrom(slice) {
-	let candidate = slice;
-	const lastClose = Math.max(candidate.lastIndexOf("}"), candidate.lastIndexOf("]"));
-	if (lastClose > 0) {
-		const balanced = tryParse(candidate.slice(0, lastClose + 1));
-		if (balanced !== UNPARSEABLE) return balanced;
-	}
-	for (let i = 0; i < 64; i++) {
-		const closed = autoCloseTruncatedJson(candidate);
-		if (closed !== null) {
-			const parsed = tryParse(closed);
-			if (parsed !== UNPARSEABLE) return parsed;
-		}
-		const cut = lastCommaOutsideString(candidate);
-		if (cut <= 0) return UNPARSEABLE;
-		candidate = candidate.slice(0, cut);
-	}
-	return UNPARSEABLE;
-}
-//#endregion
-//#region src/reflective-mutation.ts
+// src/reflective-mutation.ts
 /**
 * Reflective mutation — primitives for trace-conditioned prompt rewriting.
 *
@@ -917,8 +621,7 @@ function parseReflectionResponse(raw, maxProposals) {
 	}
 	return out;
 }
-//#endregion
-//#region src/campaign/score-utils.ts
+// src/campaign/score-utils.ts
 /**
 * Shared campaign-score reductions used by every optimizer preset
 * (`runOptimization`, external optimization methods, `compareOptimizationMethods`).
@@ -1024,8 +727,7 @@ function campaignBreakdown(campaign) {
 		})
 	};
 }
-//#endregion
-//#region src/campaign/provenance.ts
+// src/campaign/provenance.ts
 /**
 * Loop provenance — the durable, queryable record of WHAT a self-improvement
 * loop did and WHY, plus the OTel spans that let an OTLP collector pivot from
@@ -1497,8 +1199,7 @@ async function emitLoopProvenance(args) {
 		spansPath
 	};
 }
-//#endregion
-//#region src/attestation.ts
+// src/attestation.ts
 /**
 * Reproducibility attestation for any serializable report object.
 *
@@ -1589,8 +1290,7 @@ function verifyAttestation(report, attested) {
 	};
 	return { valid: true };
 }
-//#endregion
-//#region src/experiment/evidence-receipt.ts
+// src/experiment/evidence-receipt.ts
 /**
 * Evidence receipts are the join between Runtime execution and Eval promotion.
 *
@@ -1703,8 +1403,7 @@ function requiredIdentity(value, field) {
 	if (normalized.length === 0) throw new TypeError(`evidence receipt: ${field} must be non-empty`);
 	return normalized;
 }
-//#endregion
-//#region src/experiment/campaign-evidence.ts
+// src/experiment/campaign-evidence.ts
 /** Bind a complete measured campaign to its executed surface and actual outputs. */
 function createCampaignEvidenceReceipt(input) {
 	const { campaign, surface, context } = input;
@@ -1727,7 +1426,6 @@ function createCampaignEvidenceReceipt(input) {
 		inputsHash: campaign.splitDigest
 	});
 }
-//#endregion
-export { surfaceHashMatches as A, campaignSplitDigestFromIdentities as B, assertCodeSurfaceIdentity as C, surfaceContentHash as D, renderSurfaceDiff as E, assertCampaignSplitIdentity as F, assertCompleteCampaign as I, campaignCoverage as L, assertRealBackend as M, summarizeBackendIntegrity as N, surfaceDispatchRef as O, assertCampaignDesign as P, campaignScenarioIdentity as R, openFinalEvidenceLedger as S, componentSurfaceIdentityMaterial as T, formatCoverageFailures as V, parseReflectionResponse as _, ATTESTATION_ALGORITHM as a, FinalEvidenceConflictError as b, campaignMeasurementDigest as c, loopProvenanceArgsFromResult as d, campaignBreakdown as f, buildReflectionPrompt as g, pairedCampaignComposites as h, verifyEvidenceReceipt as i, BackendIntegrityError as j, surfaceHash as k, canonicalDigest as l, campaignMeanCompositeOrNull as m, createEvidenceReceipt as n, attest as o, campaignMeanComposite as p, isIndependentEvidence as r, verifyAttestation as s, createCampaignEvidenceReceipt as t, emitLoopProvenance as u, recoverTruncatedJson as v, codeSurfaceIdentityMaterial as w, FinalEvidenceError as x, powerPreflight as y, campaignSplitDigest as z };
+export { campaignSplitDigest as A, assertRealBackend as C, assertCompleteCampaign as D, assertCampaignSplitIdentity as E, formatCoverageFailures as M, campaignCoverage as O, BackendIntegrityError as S, assertCampaignDesign as T, parseReflectionResponse as _, ATTESTATION_ALGORITHM as a, FinalEvidenceError as b, campaignMeasurementDigest as c, loopProvenanceArgsFromResult as d, campaignBreakdown as f, buildReflectionPrompt as g, pairedCampaignComposites as h, verifyEvidenceReceipt as i, campaignSplitDigestFromIdentities as j, campaignScenarioIdentity as k, canonicalDigest as l, campaignMeanCompositeOrNull as m, createEvidenceReceipt as n, attest as o, campaignMeanComposite as p, isIndependentEvidence as r, verifyAttestation as s, createCampaignEvidenceReceipt as t, emitLoopProvenance as u, powerPreflight as v, summarizeBackendIntegrity as w, openFinalEvidenceLedger as x, FinalEvidenceConflictError as y };
 
-//# sourceMappingURL=campaign-evidence-DunzcJGv.js.map
+//# sourceMappingURL=campaign-evidence-DN8rWuBj.js.map