@iowarp/clio-coder 0.3.7 → 0.3.8

package/src/domains/dispatch/write-boundary-enforcer.ts

@@ -8,6 +8,11 @@
  * a path that changed cannot be pinned on one of them, and the plan compiler
  * has already refused the case where more than one of them was allowed to
  * write. Every step in a violating window fails, and the verdict says why.
+ *
+ * The window also carries what its runs recorded writing, which is what keeps
+ * the checkout diff from being read as authorship. A change nobody in the
+ * window recorded belongs to whoever else has the checkout open, and the
+ * verdict reports it without touching it.
  */
 
 import { existsSync } from "node:fs";
@@ -16,9 +21,11 @@ import type { ExecutionPlan } from "./execution-plan.js";
 import type { ExecutionWriteBoundaryOutcome } from "./execution-scheduler.js";
 import {
 	assertWriteBoundaryInsideRoot,
+	assertWriteBoundaryVisibleToGit,
 	captureWorkspaceSnapshot,
 	enforceWriteBoundary,
 	type WorkspaceSnapshot,
+	type WriteBoundaryAttribution,
 	type WriteBoundaryVerdict,
 	writeWriteBoundaryVerdict,
 } from "./write-boundary.js";
@@ -30,6 +37,15 @@ export interface WriteBoundaryEnforcerInput {
 	rootId: string;
 	/** Declared allowlist for a step, or undefined when it declares no boundary. */
 	boundaryFor(stepId: string): ReadonlyArray<string> | undefined;
+	/**
+	 * What the step's own run recorded writing, or null when the step has no
+	 * closed write record: a code step running a registered command, a runtime
+	 * that publishes no tool telemetry, a run whose telemetry had holes, or a run
+	 * that shelled out or spawned a worker and could have written a path no
+	 * argument names. Absent entirely means the caller offers no record at all,
+	 * and every change outside the declaration is blamed on the window.
+	 */
+	recordedWritesFor?(stepId: string): ReadonlyArray<string> | null;
 	/** Called once per closed window with the sealed verdict and where it landed. */
 	onVerdict?(verdict: WriteBoundaryVerdict, path: string): void;
 }
@@ -54,6 +70,9 @@ export function preflightWriteBoundaries(plan: ExecutionPlan, root: string): voi
 		);
 	}
 	for (const step of declared) assertWriteBoundaryInsideRoot(root, step.writes ?? []);
+	// One question over the whole plan: an entry is refused for being invisible
+	// to git whichever step declared it, and the diagnostic names the path.
+	assertWriteBoundaryVisibleToGit(root, [...new Set(declared.flatMap((step) => [...(step.writes ?? [])]))]);
 }
 
 export function createWriteBoundaryEnforcer(input: WriteBoundaryEnforcerInput): WriteBoundaryEnforcer {
@@ -63,10 +82,34 @@ export function createWriteBoundaryEnforcer(input: WriteBoundaryEnforcerInput):
 		for (const stepId of stepIds) for (const entry of input.boundaryFor(stepId) ?? []) allow.add(entry);
 		return [...allow].sort();
 	};
+	/**
+	 * The window's write witness. One step without a record makes the whole
+	 * window's record incomplete: the steps share a checkout, so a change no
+	 * other step wrote could still be the unobserved one's.
+	 */
+	const attributionFor = (stepIds: ReadonlyArray<string>): WriteBoundaryAttribution | undefined => {
+		const recordedWritesFor = input.recordedWritesFor;
+		if (recordedWritesFor === undefined) return undefined;
+		const recorded = new Set<string>();
+		let complete = true;
+		for (const stepId of stepIds) {
+			const observed = recordedWritesFor(stepId);
+			if (observed === null) {
+				complete = false;
+				continue;
+			}
+			for (const target of observed) recorded.add(target);
+		}
+		return { recorded: [...recorded].sort(), complete };
+	};
 	return {
 		begin(window, stepIds) {
 			const allow = allowFor(stepIds);
 			assertWriteBoundaryInsideRoot(input.root, allow);
+			// Also checked whole-plan at preflight. Repeated here because a
+			// delegation plan splices steps into a running fleet, and their
+			// declarations reach this function without ever passing preflight.
+			assertWriteBoundaryVisibleToGit(input.root, allow);
 			open.set(window, { snapshot: captureWorkspaceSnapshot(input.root), allow });
 		},
 		async verify(window, stepIds) {
@@ -75,11 +118,13 @@ export function createWriteBoundaryEnforcer(input: WriteBoundaryEnforcerInput):
 				throw new Error(`write boundary: window '${window}' was verified without a snapshot`);
 			}
 			open.delete(window);
+			const attribution = attributionFor(stepIds);
 			const verdict = enforceWriteBoundary({
 				snapshot: pending.snapshot,
 				window,
 				stepIds,
 				allow: pending.allow,
+				...(attribution === undefined ? {} : { attribution }),
 			});
 			const path = writeWriteBoundaryVerdict(input.rootId, verdict);
 			input.onVerdict?.(verdict, path);

package/src/domains/dispatch/write-boundary.ts

@@ -21,18 +21,27 @@
  * Everything is pinned to the baseline commit recorded at snapshot time, so a
  * commit step that moves HEAD without touching the working tree reads as "no
  * change" instead of as a wholesale rewrite.
+ *
+ * A window's diff answers "what changed", never "who changed it". An operator
+ * editing the checkout while a fleet runs produces changes inside the window
+ * that no step made, and rolling those back to the baseline destroys work
+ * nobody asked Clio to touch. So a change is blamed on the window only when it
+ * intersects what the window's own runs recorded writing; everything else is
+ * reported as an unattributed concurrent change and left exactly as it is.
  */
 
 import { execFileSync } from "node:child_process";
 import { createHash } from "node:crypto";
 import { existsSync, lstatSync, mkdirSync, readlinkSync, realpathSync, rmSync, writeFileSync } from "node:fs";
-import { dirname, join, relative, resolve, sep } from "node:path";
+import { dirname, isAbsolute, join, relative, resolve, sep } from "node:path";
+import { asDirectoryPathBoundary, pathBoundaryEntryCovers } from "../../core/path-boundary.js";
 import { clioStateDir } from "../../core/xdg.js";
 import { atomicWrite } from "../../engine/session.js";
 import { isGitRepository } from "../../tools/compete-worktrees.js";
 import {
 	describeWriteBoundary,
 	normalizeWriteBoundary,
+	normalizeWriteBoundaryEntry,
 	type WriteBoundary,
 	writeBoundaryCovers,
 } from "../agents/write-boundary.js";
@@ -71,10 +80,15 @@ function gitBytes(root: string, args: ReadonlyArray<string>): Buffer {
  *
  * Enforcement sees exactly what git sees, so an ignored path is outside it: a
  * repository decides what counts as its own content, and this is not the place
- * to overrule that. The one thing subtracted is Clio's own state directory when
- * an operator has placed it inside the workspace. Receipts, code-step logs, and
- * boundary verdicts are the orchestrator writing its journal while the step
- * runs, and blaming the step for them would make every window a violation.
+ * to overrule that. The consequence is that a boundary declared over an ignored
+ * path could never be verified, which is why `assertWriteBoundaryVisibleToGit`
+ * refuses that declaration before anything runs rather than letting this
+ * function certify an unobserved window as clean.
+ *
+ * The one thing subtracted is Clio's own state directory when an operator has
+ * placed it inside the workspace. Receipts, code-step logs, and boundary
+ * verdicts are the orchestrator writing its journal while the step runs, and
+ * blaming the step for them would make every window a violation.
  */
 function dirtyPaths(root: string): string[] {
 	const raw = git(root, ["status", "--porcelain=v1", "-z", "-uall", "--no-renames"]);
@@ -222,14 +236,55 @@ export interface WriteBoundaryUnrecoverable {
 }
 
 /**
- * `clean` means nothing outside the allowlist changed. `rolled-back` means
- * something did and the repository now looks as it did before. Anything else is
- * `rollback-incomplete`: the working tree is left exactly as the step made it
- * and handed to the operator with the list, because a rollback that guesses at
- * content it never recorded destroys work.
+ * `clean` means no change outside the allowlist is attributable to this
+ * window's steps. `rolled-back` means one was and the repository now looks as
+ * it did before. Anything else is `rollback-incomplete`: the working tree is
+ * left exactly as the step made it and handed to the operator with the list,
+ * because a rollback that guesses at content it never recorded destroys work.
+ *
+ * A `clean` verdict can still carry `unattributed` paths. Those changed inside
+ * the window and no run in it recorded writing them, so they are the
+ * operator's, another process's, or a build artifact's, and nothing here may
+ * touch them.
  */
 export type WriteBoundaryStatus = "clean" | "rolled-back" | "rollback-incomplete";
 
+/**
+ * What the window's own runs recorded aiming a successful mutation at.
+ *
+ * This is the write-side witness the diff cannot be: the checkout says a path
+ * changed, and only the run's own tool-call record says whether the run is what
+ * changed it. The record comes from the same fold that grounds a sealed
+ * mutation report, so the boundary and the result contract judge one set of
+ * facts.
+ *
+ * A run can write through a channel no tool event enumerates, and then its path
+ * set is a lower bound rather than the whole list. That is not silently
+ * tolerated: the recorder knows which tools can do it (bash, verify, dispatch,
+ * steer, and any tool this process has no schema for) and reports the record
+ * open, which arrives here as `complete: false`. So the set is only ever read
+ * as a closed list for a run that could not have written outside it, and
+ * enforcement over a run that could stays exactly as strong as it was before
+ * attribution existed.
+ */
+export interface WriteBoundaryAttribution {
+	/**
+	 * Paths the window's runs aimed a successful mutation at. Absolute, or
+	 * relative to the workspace root; anything resolving outside the workspace is
+	 * dropped, because git could not have reported it as a change either.
+	 */
+	recorded: ReadonlyArray<string>;
+	/**
+	 * False when at least one step in the window has no closed write record: a
+	 * code step running a registered command, a runtime that publishes no tool
+	 * telemetry, a run whose telemetry came back with holes, or a run that
+	 * called a tool able to write a path its arguments do not name. An open
+	 * record cannot clear anybody, so the window falls back to blaming every
+	 * change outside the allowlist, and the verdict says it did.
+	 */
+	complete: boolean;
+}
+
 export interface WriteBoundaryVerdict {
 	version: 1;
 	/** Which scheduling window this covers, e.g. `wave-2` or `revalidate-suite.check.1`. */
@@ -243,12 +298,25 @@ export interface WriteBoundaryVerdict {
 	checkedAt: string;
 	changedPaths: ReadonlyArray<string>;
 	violations: ReadonlyArray<string>;
+	/**
+	 * Paths outside the allowlist that changed inside the window and that no run
+	 * in it recorded writing. Recorded so the operator can see them, never rolled
+	 * back: the window did not make them, so the baseline commit is not their
+	 * prior content.
+	 */
+	unattributed: ReadonlyArray<string>;
+	/** Whether every step in the window offered an enumerable write record. */
+	attributionComplete: boolean;
 	rolledBack: ReadonlyArray<WriteBoundaryRollback>;
 	unrecoverable: ReadonlyArray<WriteBoundaryUnrecoverable>;
 	status: WriteBoundaryStatus;
 	/** Typed failure reason, or null when the window stayed inside its boundary. */
 	reason: typeof WRITE_BOUNDARY_VIOLATION_REASON | null;
-	/** Operator-facing message naming the offending paths and the declaration. */
+	/**
+	 * Operator-facing message naming the offending paths and the declaration, or
+	 * naming the concurrent changes the window saw but did not cause. Null only
+	 * when neither happened.
+	 */
 	detail: string | null;
 	digest: string;
 }
@@ -258,6 +326,12 @@ export interface EnforceWriteBoundaryInput {
 	window: string;
 	stepIds: ReadonlyArray<string>;
 	allow: WriteBoundary;
+	/**
+	 * What the window's runs recorded writing. Absent means no caller offered a
+	 * record, which reads the same as an incomplete one: every change outside the
+	 * allowlist is blamed on the window.
+	 */
+	attribution?: WriteBoundaryAttribution;
 }
 
 function canonicalVerdict(verdict: Omit<WriteBoundaryVerdict, "digest">): string {
@@ -317,6 +391,22 @@ function rollbackPath(
 	return { rolledBack: { path: change.path, action: "restored", restoredFrom: `${snapshot.head}:${change.path}` } };
 }
 
+/**
+ * Whether one changed path falls inside what a run recorded writing.
+ *
+ * A recorded target is usually the exact file a write or edit named, but a
+ * command's delete target can be a directory, so a change under a recorded
+ * prefix counts as the same write. Matching is on segment boundaries: `work`
+ * does not cover `workspace/a.txt`.
+ */
+function withinRecordedWrite(recorded: ReadonlySet<string>, path: string): boolean {
+	if (recorded.has(path)) return true;
+	for (const target of recorded) {
+		if (path.startsWith(`${target}/`)) return true;
+	}
+	return false;
+}
+
 /**
  * Compare, roll back, and return the verdict. Never throws for a violation: a
  * violation is evidence the caller records and acts on, not an exception that
@@ -325,7 +415,21 @@ function rollbackPath(
 export function enforceWriteBoundary(input: EnforceWriteBoundaryInput): WriteBoundaryVerdict {
 	const allow = normalizeWriteBoundary(input.allow);
 	const changes = diffWorkspace(input.snapshot);
-	const violations = changes.filter((change) => !writeBoundaryCovers(allow, change.path));
+	const outside = changes.filter((change) => !writeBoundaryCovers(allow, change.path));
+	const attributionComplete = input.attribution?.complete === true;
+	const recorded = new Set<string>();
+	for (const target of input.attribution?.recorded ?? []) {
+		const relativeTarget = repoRelative(input.snapshot.root, target);
+		if (relativeTarget !== null) recorded.add(relativeTarget);
+	}
+	// Without a complete record nothing can be cleared, so the window keeps the
+	// whole diff. With one, only what the runs actually wrote is theirs.
+	const violations = attributionComplete
+		? outside.filter((change) => withinRecordedWrite(recorded, change.path))
+		: outside;
+	const unattributed = attributionComplete
+		? outside.filter((change) => !withinRecordedWrite(recorded, change.path))
+		: [];
 	const rolledBack: WriteBoundaryRollback[] = [];
 	const unrecoverable: WriteBoundaryUnrecoverable[] = [];
 	for (const change of violations) {
@@ -335,15 +439,15 @@ export function enforceWriteBoundary(input: EnforceWriteBoundaryInput): WriteBou
 	}
 	const status: WriteBoundaryStatus =
 		violations.length === 0 ? "clean" : unrecoverable.length === 0 ? "rolled-back" : "rollback-incomplete";
-	const attribution =
+	const who =
 		input.stepIds.length === 1
 			? `step '${input.stepIds[0]}'`
 			: `steps ${input.stepIds.map((id) => `'${id}'`).join(", ")} ran concurrently in one checkout, so the write cannot be attributed to one of them`;
-	const detail =
+	const violationDetail =
 		violations.length === 0
-			? null
+			? []
 			: [
-					`${attribution} wrote outside its declared boundary.`,
+					`${who} wrote outside its declared boundary.`,
 					`Unauthorized paths: ${violations.map((change) => change.path).join(", ")}.`,
 					`Declared writes: ${describeWriteBoundary(allow)}.`,
 					status === "rolled-back"
@@ -351,8 +455,21 @@ export function enforceWriteBoundary(input: EnforceWriteBoundaryInput): WriteBou
 						: `Rollback is incomplete; the working tree is left as the step made it. Unrestorable: ${unrecoverable
 								.map((entry) => `${entry.path} (${entry.reason})`)
 								.join("; ")}.`,
+					attributionComplete
+						? "Each was blamed on this window because a run in it recorded writing that path."
+						: "This window had no complete write record, so every change outside the declaration was blamed on it.",
 					"If the change was legitimate, widen the step's `writes:` declaration to cover it.",
-				].join(" ");
+				];
+	const concurrentDetail =
+		unattributed.length === 0
+			? []
+			: [
+					`Concurrent changes were seen in this window that no run in it recorded writing: ${unattributed
+						.map((change) => change.path)
+						.join(", ")}.`,
+					"They were left untouched, because a path this window did not write is not this window's to restore.",
+				];
+	const detail = [...violationDetail, ...concurrentDetail];
 	const body: Omit<WriteBoundaryVerdict, "digest"> = {
 		version: 1,
 		window: input.window,
@@ -363,11 +480,13 @@ export function enforceWriteBoundary(input: EnforceWriteBoundaryInput): WriteBou
 		checkedAt: new Date().toISOString(),
 		changedPaths: changes.map((change) => change.path),
 		violations: violations.map((change) => change.path),
+		unattributed: unattributed.map((change) => change.path),
+		attributionComplete,
 		rolledBack,
 		unrecoverable,
 		status,
 		reason: violations.length === 0 ? null : WRITE_BOUNDARY_VIOLATION_REASON,
-		detail,
+		detail: detail.length === 0 ? null : detail.join(" "),
 	};
 	return { ...body, digest: createHash("sha256").update(canonicalVerdict(body), "utf8").digest("hex") };
 }
@@ -387,16 +506,76 @@ export function assertWriteBoundaryInsideRoot(root: string, allow: WriteBoundary
 		// path a step is about to create.
 		while (probe !== realRoot && !existsSync(probe)) probe = dirname(probe);
 		const resolved = realpathSync(probe);
-		if (resolved !== realRoot && !resolved.startsWith(realRoot + sep)) {
+		// The root is inside itself. The prefix walk above stops at the root when
+		// a declared directory does not exist yet, and `pathBoundaryEntryCovers`
+		// answers a different question: it asks whether a changed path falls under
+		// a declaration, where the directory entry itself is not a change git ever
+		// reports. Asking it about the root would refuse every boundary naming a
+		// path the step is about to create.
+		if (resolved !== realRoot && !pathBoundaryEntryCovers(asDirectoryPathBoundary(realRoot), resolved)) {
 			throw new Error(`write boundary: entry '${entry}' resolves to ${resolved}, outside the workspace at ${realRoot}`);
 		}
 	}
 }
 
-/** Repo-relative form of an absolute path, or null when it is outside. */
+/**
+ * Refuse a declaration git will never report on.
+ *
+ * `git status` does not list an ignored path, so a boundary entry the
+ * repository ignores names a region enforcement is blind to: every window over
+ * it is certified clean without a single observation, which is a stronger claim
+ * than "nothing was checked" and reads as a passing one. That is the same
+ * defect `assertWriteBoundaryInsideRoot` refuses for an entry that leaves the
+ * workspace through a symlink, so it is refused in the same place and at the
+ * same time, before anything runs.
+ *
+ * A tracked path is reportable whatever the ignore rules say, and git's own
+ * index-aware check is what decides that, so this never refuses a path the
+ * repository is actually following.
+ */
+export function assertWriteBoundaryVisibleToGit(root: string, allow: WriteBoundary): void {
+	// Entry by entry rather than through `normalizeWriteBoundary`, whose 32-entry
+	// cap is a per-step declaration limit. A caller may legitimately ask this
+	// question about every entry a whole plan declared.
+	const entries = [...new Set(allow.map(normalizeWriteBoundaryEntry))].sort();
+	if (entries.length === 0) return;
+	const absolute = resolve(root);
+	let raw: string;
+	try {
+		raw = git(absolute, ["check-ignore", "-v", "-z", "--stdin"], `${entries.join("\0")}\0`);
+	} catch (error) {
+		// `check-ignore` exits 1 when nothing matched, which is the common answer
+		// and not a failure. Any other exit is a real inability to decide, and
+		// enforcement fails closed rather than assuming visibility.
+		const status = (error as { status?: number }).status;
+		if (status === 1) return;
+		throw new Error(
+			`write boundary: could not determine whether ${describeWriteBoundary(entries)} is ignored by git in ${absolute}: ${(error as Error).message}`,
+		);
+	}
+	const fields = raw.split("\0");
+	const refusals: string[] = [];
+	// `-v -z` emits source, line number, pattern, and pathname per match.
+	for (let index = 0; index + 3 < fields.length; index += 4) {
+		const [source, line, pattern, path] = fields.slice(index, index + 4);
+		if (path === undefined || path.length === 0) continue;
+		refusals.push(`'${path}' is ignored by ${source}:${line}:${pattern}`);
+	}
+	if (refusals.length === 0) return;
+	throw new Error(
+		`write boundary: ${refusals.join(", ")}. git status never reports an ignored path, so a boundary declared over one is verified against nothing and every window would be certified clean without an observation. Declare a path the repository tracks, or stop ignoring this one.`,
+	);
+}
+
+/**
+ * Workspace-relative form of a path, or null when it resolves outside. A
+ * relative input is taken as already workspace-relative, which is how a
+ * recorded write target and an absolute state directory both normalize here.
+ */
 function repoRelative(root: string, path: string): string | null {
-	const rel = relative(resolve(root), resolve(path));
-	if (rel.length === 0 || rel.startsWith("..") || rel.startsWith(`..${sep}`)) return null;
+	const base = resolve(root);
+	const rel = relative(base, resolve(base, path));
+	if (rel.length === 0 || rel === ".." || rel.startsWith(`..${sep}`) || isAbsolute(rel)) return null;
 	return rel.split(sep).join("/");
 }
 

package/src/domains/eval/metrics/evidence.ts

@@ -6,7 +6,35 @@
  * than silently pass.
  */
 
-import type { RunReceipt } from "../../dispatch/types.js";
+import { readFileSync } from "node:fs";
+import { join } from "node:path";
+import type { RunEnvelope, RunReceipt } from "../../dispatch/types.js";
+import { formatTrustSummary, summarizeTrustStatus } from "../../evidence/trust-projection.js";
+import {
+	adaptRunReceiptTrustStatus,
+	inspectRunReceiptTrustStatus,
+	TRUST_STATUS_AXES,
+} from "../../evidence/trust-status.js";
+
+function dispatchScopeMetrics(receipt: RunReceipt): Record<string, string | boolean | number> {
+	const scope = receipt.pathScope;
+	if (scope === undefined) return {};
+	const entries = [...scope.workingContextPaths, ...scope.writeBoundaries];
+	const evidence = entries.flatMap((entry) => entry.evidence);
+	const count = (source: string): number => evidence.filter((entry) => entry.source === source).length;
+	return {
+		"dispatch.scope.mode": scope.mode,
+		"dispatch.scope.inferredPathCount": entries.filter((entry) =>
+			entry.evidence.some((item) => item.provenance === "inferred"),
+		).length,
+		"dispatch.scope.derivedPathCount": entries.filter((entry) =>
+			entry.evidence.some((item) => item.provenance === "derived"),
+		).length,
+		"dispatch.scope.source.task": count("task"),
+		"dispatch.scope.source.briefing": count("briefing"),
+		"dispatch.scope.source.writeRoots": count("writeRoots"),
+	};
+}
 
 /**
  * Extract the sealed RunReceipt that `clio-coder run --agent … --json` prints after
@@ -53,9 +81,14 @@ function isReceiptShaped(value: unknown): value is RunReceipt {
  * mirrors the durable findings summary and is omitted when that summary is
  * absent, so a gate on the metric fails closed instead of inventing a value.
  */
-export function evidenceMetricsFromReceipt(receipt: RunReceipt): Record<string, string | boolean | number> {
+export function evidenceMetricsFromReceipt(
+	receipt: RunReceipt,
+	options: { envelope?: RunEnvelope | null } = {},
+): Record<string, string | boolean | number> {
 	return {
 		"evidence.verification": receipt.verification.state,
+		...dispatchScopeMetrics(receipt),
+		...evidenceTrustMetrics(receipt, options.envelope ?? null),
 		...(receipt.findingsSummary === undefined
 			? {}
 			: { "evidence.firstPassSuccess": receipt.findingsSummary.firstPassSuccess === true }),
@@ -64,3 +97,47 @@ export function evidenceMetricsFromReceipt(receipt: RunReceipt): Record<string,
 		...(typeof receipt.costUsd === "number" && Number.isFinite(receipt.costUsd) ? { "cost.usd": receipt.costUsd } : {}),
 	};
 }
+
+/**
+ * The canonical trust status as metrics, so a benchmark table reads the same
+ * verdict as every operator surface instead of the raw receipt marker. The
+ * receipt is authenticated against its ledger row when the caller can supply
+ * one; without a row the canonical model reports the seal unchecked and the
+ * receipt-owned axes unobserved, which is the honest answer for a receipt
+ * that only ever arrived through stdout.
+ */
+function evidenceTrustMetrics(
+	receipt: RunReceipt,
+	envelope: RunEnvelope | null,
+): Record<string, string | boolean | number> {
+	// A missing row means the seal was never checked, not that it failed:
+	// only a row that exists and disagrees can break a seal.
+	const status =
+		envelope === null ? adaptRunReceiptTrustStatus(receipt) : inspectRunReceiptTrustStatus(receipt, envelope).status;
+	const summary = summarizeTrustStatus(status);
+	return {
+		"evidence.trust.version": summary.version,
+		"evidence.trust.verdict": summary.verdict,
+		"evidence.trust.summary": formatTrustSummary(status),
+		...Object.fromEntries(TRUST_STATUS_AXES.map((axis) => [`evidence.trust.${axis}`, status[axis].state])),
+	};
+}
+
+/**
+ * The ledger row a stdout receipt was sealed from, read from the state dir
+ * the run wrote to. Null when the ledger or the row cannot be read; the
+ * caller then reports the seal unchecked rather than guessing.
+ */
+export function readRunEnvelopeForReceipt(receipt: RunReceipt, stateDir: string): RunEnvelope | null {
+	try {
+		const parsed: unknown = JSON.parse(readFileSync(join(stateDir, "runs.json"), "utf8"));
+		if (!Array.isArray(parsed)) return null;
+		const row = parsed.find(
+			(entry): entry is RunEnvelope =>
+				typeof entry === "object" && entry !== null && (entry as { id?: unknown }).id === receipt.runId,
+		);
+		return row ?? null;
+	} catch {
+		return null;
+	}
+}

package/src/domains/eval/runners/clio-run.ts

@@ -1,5 +1,10 @@
 import { shellQuote } from "../../../core/shell-quote.js";
-import { evidenceMetricsFromReceipt, receiptFromRunJsonStdout } from "../metrics/evidence.js";
+import { clioStateDir } from "../../../core/xdg.js";
+import {
+	evidenceMetricsFromReceipt,
+	readRunEnvelopeForReceipt,
+	receiptFromRunJsonStdout,
+} from "../metrics/evidence.js";
 import { streamInvariantMetrics } from "../metrics/invariants.js";
 import { tokenMetricEntries } from "../metrics/token-stream.js";
 import type { EvalRunnerV2, EvalSuiteTargetV2 } from "../schema/suite.js";
@@ -35,6 +40,11 @@ export async function runClioRunRunner(
 	// prints; a runner without a receipt leaves them absent so any gate on
 	// them fails closed instead of reading prose labels.
 	const receipt = receiptFromRunJsonStdout(result.stdout);
+	// The canonical trust metrics authenticate the receipt against the ledger
+	// row the child wrote; a stdout receipt on its own only ever reads as an
+	// unchecked seal.
+	const envelope =
+		receipt === null ? null : readRunEnvelopeForReceipt(receipt, env?.CLIO_CODER_STATE_DIR ?? clioStateDir());
 	return {
 		assignmentId: receipt === null ? null : (receipt.lineage?.rootRunId ?? receipt.runId),
 		terminalReceiptDigest: receipt?.integrity.digest ?? null,
@@ -52,7 +62,7 @@ export async function runClioRunRunner(
 			"tools.failed": tools.failed,
 			"tools.blocked": tools.blocked,
 			"verifier.exitCode": result.exitCode,
-			...(receipt === null ? {} : evidenceMetricsFromReceipt(receipt)),
+			...(receipt === null ? {} : evidenceMetricsFromReceipt(receipt, { envelope })),
 			...(receipt === null
 				? {}
 				: { "evidence.qualityLabel": receipt.quality.typedValidations.length > 0 ? "measured" : "unmeasured" }),