fix: omit unmeasured receipt latency

README.md

@@ -148,6 +148,9 @@ const artifact = await createEvaluationReceiptFromNativeReport(
 );
 ```
 
+Latency percentiles are optional in a receipt projection. Omit measurements the native
+framework did not collect; do not use zero as a stand-in for unknown evidence.
+
 The artifact contains aggregate metrics and content digests, not raw prompts, expected
 outputs, or failure transcripts. Its SHA-256 digest can also be placed in a signed Trust
 Bundle as `eval:<name>`, allowing the same eval evidence to be independently verified and

packages/evals/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@stackbilt/evals",
-  "version": "0.2.1",
+  "version": "0.2.2",
   "description": "Edge-native evaluation framework for AI agents and classifiers. Zero runtime dependencies beyond Zod.",
   "type": "module",
   "main": "./dist/index.js",

packages/evals/src/index.ts

@@ -20,6 +20,7 @@ export {
 export {
   DigestSchema,
   EvaluationGateSchema,
+  EvaluationReceiptSummarySchema,
   EvaluationReportBindingSchema,
   EvaluationReceiptSchema,
   EvaluationReceiptArtifactSchema,
@@ -32,6 +33,7 @@ export {
 } from './receipt.js';
 export type {
   EvaluationGate,
+  EvaluationReceiptSummary,
   EvaluationReportBinding,
   EvaluationReceipt,
   EvaluationReceiptArtifact,

packages/evals/src/receipt.test.ts

@@ -98,6 +98,24 @@ describe('evaluation receipts', () => {
     expect(changed.receipt.report.digest).not.toBe(artifact.receipt.report.digest);
   });
 
+  it('omits latency rather than fabricating unmeasured percentiles', async () => {
+    const { p50_latency_ms, p90_latency_ms, p95_latency_ms, ...summaryWithoutLatency } = report().summary;
+    const artifact = await createEvaluationReceiptFromNativeReport(
+      { framework: 'native', result: 'pass' },
+      {
+        runId: report().run_id,
+        runner: 'native-framework@1.0.0',
+        startedAt: report().started_at,
+        finishedAt: report().finished_at,
+        summary: summaryWithoutLatency,
+      },
+      options,
+    );
+
+    expect(artifact.receipt.report.summary).not.toHaveProperty('p50_latency_ms');
+    expect(p50_latency_ms + p90_latency_ms + p95_latency_ms).toBeGreaterThan(0);
+  });
+
   it('keeps the deployment gate explicit rather than deriving it from pass rate', async () => {
     const blocked = await createEvaluationReceipt(report(), {
       ...options,

packages/evals/src/receipt.ts

@@ -13,13 +13,23 @@ export const EvaluationGateSchema = z.object({
   reasons: z.array(z.string().min(1)).default([]),
 });
 
+/**
+ * Public aggregate shared across eval frameworks. Latency percentiles are
+ * optional because a producer must omit measurements it did not collect.
+ */
+export const EvaluationReceiptSummarySchema = EvalRunReportSchema.shape.summary.extend({
+  p50_latency_ms: z.number().nonnegative().optional(),
+  p90_latency_ms: z.number().nonnegative().optional(),
+  p95_latency_ms: z.number().nonnegative().optional(),
+});
+
 /** Public projection required to bind a framework-native private report. */
 export const EvaluationReportBindingSchema = z.object({
   runId: z.string().uuid(),
   runner: z.string().min(1),
   startedAt: z.string().datetime(),
   finishedAt: z.string().datetime(),
-  summary: EvalRunReportSchema.shape.summary,
+  summary: EvaluationReceiptSummarySchema,
 });
 
 export const EvaluationReceiptSchema = z.object({
@@ -42,7 +52,7 @@ export const EvaluationReceiptSchema = z.object({
     digest: DigestSchema,
     started_at: z.string().datetime(),
     finished_at: z.string().datetime(),
-    summary: EvalRunReportSchema.shape.summary,
+    summary: EvaluationReceiptSummarySchema,
   }),
   runner: z.string().min(1),
   gate: EvaluationGateSchema,
@@ -59,6 +69,7 @@ export const EvaluationReceiptArtifactSchema = z.object({
 });
 
 export type EvaluationGate = z.infer<typeof EvaluationGateSchema>;
+export type EvaluationReceiptSummary = z.infer<typeof EvaluationReceiptSummarySchema>;
 export type EvaluationReportBinding = z.infer<typeof EvaluationReportBindingSchema>;
 export type EvaluationReceipt = z.infer<typeof EvaluationReceiptSchema>;
 export type EvaluationReceiptArtifact = z.infer<typeof EvaluationReceiptArtifactSchema>;