@sanity/ailf-studio 1.2.2 → 1.3.1

package/dist/index.d.ts

@@ -434,6 +434,14 @@ interface ScoreItem {
     /** Ceiling score — gold-standard docs injected */
     ceilingScore?: number;
 }
+/** Reference to an artifact stored in an external object store (GCS). */
+interface ArtifactRef {
+    store: "gcs";
+    bucket: string;
+    path: string;
+    bytes?: number;
+    entryCount?: number;
+}
 /** A single gap/recommendation from gap analysis */
 interface RecommendationGap {
     affectedTaskIds: string[];
@@ -451,6 +459,38 @@ interface RecommendationsData {
     generatedAt: string;
     totalPotentialLift: number;
 }
+/**
+ * Per-test result stored in reports for drill-down and audit.
+ * Mirrors StoredTestResult from @sanity/ailf-core.
+ *
+ * Per D0030, new reports omit `responseOutput` / `responseOutputTruncated`
+ * inline — the full output lives in the `testOutputs` GCS artifact and is
+ * fetched via `useArtifactCache`. Both fields remain optional so the
+ * reader path tolerates legacy reports that were published before W0045.
+ */
+interface StoredTestResultData {
+    area: string;
+    canonicalDocs?: DocumentRef[];
+    compositeScore?: number;
+    cost?: number;
+    dimensions: {
+        dimension: string;
+        reason: string;
+        score: number;
+    }[];
+    latencyMs?: number;
+    modelId: string;
+    outputFailure?: boolean;
+    responseOutput?: string;
+    responseOutputTruncated?: boolean;
+    taskId: string;
+    tokenUsage?: {
+        completion: number;
+        prompt: number;
+        total: number;
+    };
+    variant: "baseline" | "gold";
+}
 /** A single low-scoring grader judgment stored in reports */
 interface JudgmentData {
     /** Docs the task expected the model to use */
@@ -499,6 +539,13 @@ interface PerModelData {
 interface SummaryData {
     /** Per-feature agent behavior data (only present when agentic mode ran) */
     agentBehavior?: FeatureAgentBehaviorData[] | null;
+    /** External artifact references — present when pipeline uploads to GCS (D0030) */
+    artifacts?: {
+        testOutputs?: ArtifactRef;
+        renderedPrompts?: ArtifactRef;
+        rawResults?: ArtifactRef;
+        traces?: ArtifactRef;
+    };
     belowCritical: string[];
     /** All Sanity documents used across the entire evaluation */
     documentManifest?: DocumentRef[];
@@ -523,6 +570,8 @@ interface SummaryData {
     /** Gap analysis recommendations (when gap analysis was run) */
     recommendations: null | RecommendationsData;
     scores: ScoreItem[];
+    /** Per-test results with model output and metadata (D0029) */
+    testResults?: StoredTestResultData[] | null;
     timestamp: string;
 }
 /** Shape returned by scoreTimelineQuery */
@@ -734,7 +783,7 @@ declare const scoreTimelineQuery: string;
  *
  * Used by: ReportDetail view
  */
-declare const reportDetailQuery = "\n  *[_type == \"ailf.report\" && reportId == $reportId][0] {\n    _id,\n    reportId,\n    completedAt,\n    durationMs,\n    tag,\n    title,\n    provenance,\n    summary,\n    comparison\n  }\n";
+declare const reportDetailQuery = "\n  *[_type == \"ailf.report\" && reportId == $reportId][0] {\n    _id,\n    reportId,\n    completedAt,\n    durationMs,\n    tag,\n    title,\n    provenance,\n    summary,\n    \"comparison\": comparison {\n      areas,\n      deltas,\n      generatedAt,\n      improved,\n      mismatched,\n      noiseThreshold,\n      noiseThresholdEmpirical,\n      notEvaluated,\n      regressed,\n      unchanged\n    }\n  }\n";
 /**
  * Find all reports that evaluated a specific Sanity document or perspective.
  *

package/dist/index.js

@@ -9,6 +9,8 @@ import { useClient, useCurrentUser } from "sanity";
 
 // src/lib/constants.ts
 var API_VERSION = "2026-03-11";
+var isProduction = process.env.NODE_ENV === "production";
+var ARTIFACT_API_BASE_URL = isProduction ? "https://ailf-api.sanity.build/v1" : "http://localhost:3000/v1";
 
 // src/actions/GraduateToNativeAction.tsx
 import { jsx, jsxs } from "react/jsx-runtime";
@@ -721,6 +723,22 @@ function documentRefSchema() {
     type: "object"
   };
 }
+function artifactRefSchema() {
+  return {
+    fields: [
+      defineField4({ name: "store", title: "Store", type: "string" }),
+      defineField4({ name: "bucket", title: "Bucket", type: "string" }),
+      defineField4({ name: "path", title: "Path", type: "string" }),
+      defineField4({ name: "bytes", title: "Size (bytes)", type: "number" }),
+      defineField4({
+        name: "entryCount",
+        title: "Entry Count",
+        type: "number"
+      })
+    ],
+    type: "object"
+  };
+}
 var reportSchema = defineType4({
   groups: [
     { name: "main", title: "Main", default: true },
@@ -1341,6 +1359,117 @@ var reportSchema = defineType4({
           title: "Low-Scoring Judgments",
           type: "array"
         }),
+        defineField4({
+          description: "Per-test results with model output, grader reasoning, and metadata. One entry per test \xD7 model combination. See D0029.",
+          name: "testResults",
+          of: [
+            {
+              fields: [
+                defineField4({
+                  name: "taskId",
+                  title: "Task",
+                  type: "string"
+                }),
+                defineField4({
+                  name: "modelId",
+                  title: "Model ID",
+                  type: "string"
+                }),
+                defineField4({
+                  name: "area",
+                  title: "Feature Area",
+                  type: "string"
+                }),
+                defineField4({
+                  name: "variant",
+                  options: { list: ["gold", "baseline"] },
+                  title: "Variant",
+                  type: "string"
+                }),
+                defineField4({
+                  name: "dimensions",
+                  of: [
+                    {
+                      fields: [
+                        defineField4({
+                          name: "dimension",
+                          title: "Dimension",
+                          type: "string"
+                        }),
+                        defineField4({
+                          name: "score",
+                          title: "Score (0\u2013100)",
+                          type: "number"
+                        }),
+                        defineField4({
+                          name: "reason",
+                          title: "Reason",
+                          type: "text"
+                        })
+                      ],
+                      type: "object"
+                    }
+                  ],
+                  title: "Dimensions",
+                  type: "array"
+                }),
+                defineField4({
+                  name: "compositeScore",
+                  title: "Composite Score",
+                  type: "number"
+                }),
+                defineField4({
+                  name: "latencyMs",
+                  title: "Latency (ms)",
+                  type: "number"
+                }),
+                defineField4({
+                  name: "tokenUsage",
+                  fields: [
+                    defineField4({
+                      name: "prompt",
+                      title: "Prompt",
+                      type: "number"
+                    }),
+                    defineField4({
+                      name: "completion",
+                      title: "Completion",
+                      type: "number"
+                    }),
+                    defineField4({
+                      name: "total",
+                      title: "Total",
+                      type: "number"
+                    })
+                  ],
+                  title: "Token Usage",
+                  type: "object"
+                }),
+                defineField4({
+                  name: "cost",
+                  title: "Cost",
+                  type: "number"
+                }),
+                defineField4({
+                  description: "True when the model failed to produce output (empty response, API error, or refusal).",
+                  name: "outputFailure",
+                  title: "Output Failure",
+                  type: "boolean"
+                }),
+                defineField4({
+                  description: "Documentation pages the task expected the model to use.",
+                  name: "canonicalDocs",
+                  of: [documentRefSchema()],
+                  title: "Canonical Docs",
+                  type: "array"
+                })
+              ],
+              type: "object"
+            }
+          ],
+          title: "Test Results",
+          type: "array"
+        }),
         defineField4({
           name: "perModel",
           of: [
@@ -1519,6 +1648,34 @@ var reportSchema = defineType4({
           ],
           title: "Agent Behavior",
           type: "array"
+        }),
+        defineField4({
+          description: "External artifact references \u2014 points to large data in GCS that was too big for inline storage (D0030).",
+          fields: [
+            defineField4({
+              ...artifactRefSchema(),
+              name: "testOutputs",
+              title: "Test Outputs"
+            }),
+            defineField4({
+              ...artifactRefSchema(),
+              name: "renderedPrompts",
+              title: "Rendered Prompts"
+            }),
+            defineField4({
+              ...artifactRefSchema(),
+              name: "rawResults",
+              title: "Raw Results"
+            }),
+            defineField4({
+              ...artifactRefSchema(),
+              name: "traces",
+              title: "Traces"
+            })
+          ],
+          name: "artifacts",
+          title: "Artifacts",
+          type: "object"
         })
       ],
       group: ["main", "all-fields"],
@@ -3089,7 +3246,7 @@ import {
   TabPanel as TabPanel2,
   Text as Text41
 } from "@sanity/ui";
-import { useCallback as useCallback27 } from "react";
+import { useCallback as useCallback28 } from "react";
 import { useRouter as useRouter3 } from "sanity/router";
 
 // src/lib/help-context.ts
@@ -3752,7 +3909,18 @@ var reportDetailQuery = (
     title,
     provenance,
     summary,
-    comparison
+    "comparison": comparison {
+      areas,
+      deltas,
+      generatedAt,
+      improved,
+      mismatched,
+      noiseThreshold,
+      noiseThresholdEmpirical,
+      notEvaluated,
+      regressed,
+      unchanged
+    }
   }
 `
 );
@@ -5713,13 +5881,78 @@ import {
   Tooltip as Tooltip8
 } from "@sanity/ui";
 import {
-  useCallback as useCallback25,
+  useCallback as useCallback26,
   useEffect as useEffect9,
   useMemo as useMemo9,
-  useState as useState19
+  useState as useState20
 } from "react";
 import { useClient as useClient10 } from "sanity";
 
+// src/lib/useArtifactCache.ts
+import { useCallback as useCallback14, useRef as useRef5, useState as useState10 } from "react";
+function useArtifactCache(reportId, artifactRef) {
+  const cacheRef = useRef5(/* @__PURE__ */ new Map());
+  const [status, setStatus] = useState10("idle");
+  const [error, setError] = useState10(null);
+  const fetchingRef = useRef5(false);
+  const fetchArtifacts = useCallback14(async () => {
+    if (fetchingRef.current || cacheRef.current.size > 0) return;
+    if (!artifactRef) return;
+    fetchingRef.current = true;
+    setStatus("loading");
+    setError(null);
+    try {
+      const signingRes = await fetch(
+        `${ARTIFACT_API_BASE_URL}/artifacts/${encodeURIComponent(reportId)}?type=testOutputs`,
+        {
+          credentials: "omit",
+          headers: { Accept: "application/json" }
+        }
+      );
+      if (!signingRes.ok) {
+        const body = await signingRes.text().catch(() => "");
+        throw new Error(
+          `Artifact signing failed: ${signingRes.status} ${signingRes.statusText}${body ? ` \u2014 ${body.slice(0, 200)}` : ""}`
+        );
+      }
+      const envelope = await signingRes.json();
+      if (envelope.object === "error" || !envelope.url) {
+        throw new Error(
+          envelope.error?.message ?? "Invalid signing response \u2014 missing signed URL"
+        );
+      }
+      const artifactRes = await fetch(envelope.url, {
+        credentials: "omit"
+      });
+      if (!artifactRes.ok) {
+        throw new Error(
+          `GCS artifact fetch failed: ${artifactRes.status} ${artifactRes.statusText}`
+        );
+      }
+      const artifact = await artifactRes.json();
+      const map = /* @__PURE__ */ new Map();
+      for (const [key, entry] of Object.entries(artifact.entries)) {
+        map.set(key, entry);
+      }
+      cacheRef.current = map;
+      setStatus("ready");
+    } catch (err) {
+      const message = err instanceof Error ? err.message : String(err);
+      setError(message);
+      setStatus("error");
+    } finally {
+      fetchingRef.current = false;
+    }
+  }, [reportId, artifactRef]);
+  const getOutput = useCallback14(
+    (taskId, modelId) => {
+      return cacheRef.current.get(`${taskId}::${modelId}`) ?? null;
+    },
+    []
+  );
+  return { status, error, getOutput, fetchArtifacts };
+}
+
 // src/lib/thresholds.ts
 var SCORE_POSITIVE = 80;
 var SCORE_CAUTION = 70;
@@ -5761,7 +5994,7 @@ function negativeDocLiftSentiment(count) {
 }
 
 // src/components/report-detail/AgentActivitySection.tsx
-import { useMemo as useMemo5, useState as useState10 } from "react";
+import { useMemo as useMemo5, useState as useState11 } from "react";
 import { HelpCircleIcon as HelpCircleIcon6, SearchIcon as SearchIcon4 } from "@sanity/icons";
 import {
   Badge as Badge5,
@@ -5936,7 +6169,7 @@ function FeatureActivityCard({
   ] }) });
 }
 function SearchQueryList({ queries }) {
-  const [filter, setFilter] = useState10("");
+  const [filter, setFilter] = useState11("");
   const filtered = useMemo5(() => {
     if (!filter) return queries;
     const lower = filter.toLowerCase();
@@ -6582,7 +6815,7 @@ function Pill({
 // src/components/report-detail/LineageCard.tsx
 import { LinkIcon as LinkIcon2 } from "@sanity/icons";
 import { Badge as Badge6, Card as Card13, Flex as Flex15, Stack as Stack19, Text as Text24 } from "@sanity/ui";
-import { useCallback as useCallback14, useEffect as useEffect7, useState as useState11 } from "react";
+import { useCallback as useCallback15, useEffect as useEffect7, useState as useState12 } from "react";
 import { useClient as useClient5 } from "sanity";
 import { useRouter as useRouter2 } from "sanity/router";
 import { jsx as jsx26, jsxs as jsxs22 } from "react/jsx-runtime";
@@ -6602,7 +6835,7 @@ function LineageCard({ provenance, reportId }) {
   const { lineage } = provenance;
   const router = useRouter2();
   const client = useClient5({ apiVersion: API_VERSION });
-  const [spawned, setSpawned] = useState11([]);
+  const [spawned, setSpawned] = useState12([]);
   useEffect7(() => {
     let cancelled = false;
     client.fetch(SPAWNED_REPORTS_QUERY, {
@@ -6650,7 +6883,7 @@ function LineageLink({
   reportId,
   router
 }) {
-  const handleClick = useCallback14(() => {
+  const handleClick = useCallback15(() => {
     router.navigate({ reportId });
   }, [reportId, router]);
   return /* @__PURE__ */ jsxs22(Flex15, { align: "center", gap: 2, children: [
@@ -6679,7 +6912,7 @@ function SpawnedReportRow({
   report,
   router
 }) {
-  const handleClick = useCallback14(() => {
+  const handleClick = useCallback15(() => {
     router.navigate({ reportId: report.reportId });
   }, [report.reportId, router]);
   const dateLabel = formatShortDate(report.completedAt);
@@ -6714,7 +6947,7 @@ function formatShortDate(iso) {
 }
 
 // src/components/report-detail/JudgmentList.tsx
-import { useCallback as useCallback15, useEffect as useEffect8, useRef as useRef5, useState as useState12 } from "react";
+import React2, { useCallback as useCallback16, useEffect as useEffect8, useRef as useRef6, useState as useState13 } from "react";
 import { CopyIcon, ErrorOutlineIcon as ErrorOutlineIcon2, HelpCircleIcon as HelpCircleIcon7 } from "@sanity/icons";
 import { Box as Box16, Button as Button3, Flex as Flex16, Stack as Stack20, Text as Text25, Tooltip as Tooltip7, useToast as useToast2 } from "@sanity/ui";
 
@@ -6768,14 +7001,29 @@ function groupByArea(judgments) {
 function dimensionLabel2(dim) {
   return dim.split("-").map((w) => w.charAt(0).toUpperCase() + w.slice(1)).join(" ");
 }
+function testResultKey(taskId, dimension, modelId) {
+  return `${taskId}::${dimension}::${modelId}`;
+}
 function JudgmentList({
+  artifactCache,
   focus,
   judgments,
-  onFocusChange
+  onFocusChange,
+  testResults
 }) {
   if (judgments.length === 0) return null;
   const pruned = pruneNoise(judgments);
   if (pruned.length === 0) return null;
+  const testResultMap = React2.useMemo(() => {
+    const map = /* @__PURE__ */ new Map();
+    if (!testResults) return map;
+    for (const tr of testResults) {
+      for (const dim of tr.dimensions) {
+        map.set(testResultKey(tr.taskId, dim.dimension, tr.modelId), tr);
+      }
+    }
+    return map;
+  }, [testResults]);
   const outputFailureCount = judgments.length - pruned.length;
   const grouped = groupByArea(pruned);
   return /* @__PURE__ */ jsxs23(Stack20, { space: 3, children: [
@@ -6868,13 +7116,18 @@ function JudgmentList({
           ] }),
           /* @__PURE__ */ jsx28(Stack20, { space: 2, children: areaJudgments.map((j) => {
             const slug = judgmentSlug(j);
+            const matchedResult = testResultMap.get(
+              testResultKey(j.taskId, j.dimension, j.modelId)
+            );
             return /* @__PURE__ */ jsx28(
               JudgmentCard,
               {
+                artifactCache,
                 focused: slug === focus,
                 judgment: j,
                 onFocusChange,
-                slug
+                slug,
+                testResult: matchedResult
               },
               slug
             );
@@ -6886,17 +7139,29 @@ function JudgmentList({
   ] });
 }
 function JudgmentCard({
+  artifactCache,
   focused,
   judgment,
   onFocusChange,
-  slug
+  slug,
+  testResult
 }) {
-  const [expanded, setExpanded] = useState12(focused);
-  const cardRef = useRef5(null);
+  const [expanded, setExpanded] = useState13(focused);
+  const [outputExpanded, setOutputExpanded] = useState13(false);
+  const cardRef = useRef6(null);
   const toast = useToast2();
   const dimLabel = dimensionLabel2(judgment.dimension);
   const sep = judgment.taskId.indexOf(" - ");
   const taskName = sep > 0 ? judgment.taskId.substring(sep + 3) : judgment.taskId;
+  const inlineOutput = testResult?.responseOutput;
+  const artifactEntry = artifactCache?.getOutput(
+    judgment.taskId,
+    judgment.modelId
+  );
+  const resolvedOutput = inlineOutput ?? artifactEntry?.responseOutput ?? null;
+  const resolvedTruncated = testResult?.responseOutputTruncated ?? artifactEntry?.responseOutputTruncated ?? false;
+  const canFetchArtifact = !inlineOutput && !artifactEntry && artifactCache != null && artifactCache.status !== "ready";
+  const hasOutputOrCanFetch = resolvedOutput != null || canFetchArtifact;
   useEffect8(() => {
     if (focused) {
       setExpanded(true);
@@ -6906,10 +7171,10 @@ function JudgmentCard({
       return () => clearTimeout(timer);
     }
   }, []);
-  const handleToggle = useCallback15(() => {
+  const handleToggle = useCallback16(() => {
     setExpanded((prev) => !prev);
   }, []);
-  const handleCopyLink = useCallback15(
+  const handleCopyLink = useCallback16(
     (e) => {
       e.stopPropagation();
       onFocusChange?.(slug);
@@ -6933,6 +7198,39 @@ function JudgmentCard({
     },
     [toast]
   );
+  const handleCopyOutput = useCallback16(
+    (e) => {
+      e.stopPropagation();
+      if (!resolvedOutput) return;
+      navigator.clipboard.writeText(resolvedOutput).then(
+        () => {
+          toast.push({
+            closable: true,
+            status: "success",
+            title: "Model output copied to clipboard"
+          });
+        },
+        () => {
+          toast.push({
+            closable: true,
+            status: "error",
+            title: "Failed to copy output"
+          });
+        }
+      );
+    },
+    [resolvedOutput, toast]
+  );
+  const handleToggleOutput = useCallback16(
+    (e) => {
+      e.stopPropagation();
+      if (!outputExpanded && !resolvedOutput && artifactCache) {
+        artifactCache.fetchArtifacts();
+      }
+      setOutputExpanded((prev) => !prev);
+    },
+    [outputExpanded, resolvedOutput, artifactCache]
+  );
   return /* @__PURE__ */ jsx28(
     Box16,
     {
@@ -7013,22 +7311,115 @@ function JudgmentCard({
               )
             }
           ),
+          hasOutputOrCanFetch && /* @__PURE__ */ jsxs23(Box16, { children: [
+            /* @__PURE__ */ jsxs23(Flex16, { align: "center", gap: 2, children: [
+              /* @__PURE__ */ jsxs23(
+                "button",
+                {
+                  onClick: handleToggleOutput,
+                  style: {
+                    alignItems: "center",
+                    background: "none",
+                    border: "none",
+                    color: "var(--card-muted-fg-color)",
+                    cursor: "pointer",
+                    display: "flex",
+                    fontSize: 13,
+                    gap: 4,
+                    padding: 0
+                  },
+                  children: [
+                    /* @__PURE__ */ jsx28("span", { children: outputExpanded ? "\u25BE" : "\u25B8" }),
+                    /* @__PURE__ */ jsx28("span", { style: { fontWeight: 500 }, children: "Model Output" })
+                  ]
+                }
+              ),
+              resolvedTruncated && /* @__PURE__ */ jsx28(
+                "span",
+                {
+                  style: {
+                    backgroundColor: "rgba(251,191,36,0.15)",
+                    borderRadius: 4,
+                    color: "#fbbf24",
+                    fontSize: 11,
+                    padding: "1px 6px"
+                  },
+                  children: "truncated"
+                }
+              ),
+              testResult?.latencyMs != null && /* @__PURE__ */ jsxs23(Text25, { muted: true, size: 0, children: [
+                (testResult.latencyMs / 1e3).toFixed(1),
+                "s"
+              ] }),
+              testResult?.cost != null && testResult.cost > 0 && /* @__PURE__ */ jsxs23(Text25, { muted: true, size: 0, children: [
+                "$",
+                testResult.cost.toFixed(4)
+              ] })
+            ] }),
+            outputExpanded && /* @__PURE__ */ jsxs23(
+              Box16,
+              {
+                marginTop: 2,
+                style: {
+                  backgroundColor: "var(--card-muted-bg-color, rgba(255,255,255,0.04))",
+                  borderRadius: 4,
+                  maxHeight: 400,
+                  overflowY: "auto",
+                  padding: 12
+                },
+                children: [
+                  !resolvedOutput && artifactCache?.status === "loading" && /* @__PURE__ */ jsx28(Text25, { muted: true, size: 1, children: "Fetching model output\u2026" }),
+                  !resolvedOutput && artifactCache?.status === "error" && /* @__PURE__ */ jsxs23(Text25, { muted: true, size: 1, style: { color: "#f87171" }, children: [
+                    "Failed to load model output",
+                    artifactCache.error ? `: ${artifactCache.error}` : ""
+                  ] }),
+                  resolvedOutput && /* @__PURE__ */ jsx28(
+                    "pre",
+                    {
+                      style: {
+                        fontFamily: "var(--font-code-family, 'SF Mono', 'Fira Code', monospace)",
+                        fontSize: 13,
+                        lineHeight: 1.5,
+                        margin: 0,
+                        whiteSpace: "pre-wrap",
+                        wordBreak: "break-word"
+                      },
+                      children: resolvedOutput
+                    }
+                  )
+                ]
+              }
+            )
+          ] }),
           /* @__PURE__ */ jsxs23(Flex16, { align: "center", gap: 2, wrap: "wrap", children: [
             judgment.canonicalDocs && judgment.canonicalDocs.length > 0 && /* @__PURE__ */ jsxs23(Fragment9, { children: [
               /* @__PURE__ */ jsx28(Text25, { muted: true, size: 1, children: "Docs:" }),
               judgment.canonicalDocs.map((doc) => /* @__PURE__ */ jsx28(DocBadge, { doc }, doc.slug))
             ] }),
-            /* @__PURE__ */ jsx28("span", { style: { marginLeft: "auto" }, children: /* @__PURE__ */ jsx28(
-              Button3,
-              {
-                fontSize: 1,
-                icon: CopyIcon,
-                mode: "ghost",
-                onClick: handleCopyLink,
-                padding: 2,
-                text: "Copy Link"
-              }
-            ) })
+            /* @__PURE__ */ jsxs23("span", { style: { display: "flex", gap: 4, marginLeft: "auto" }, children: [
+              resolvedOutput && /* @__PURE__ */ jsx28(
+                Button3,
+                {
+                  fontSize: 1,
+                  icon: CopyIcon,
+                  mode: "ghost",
+                  onClick: handleCopyOutput,
+                  padding: 2,
+                  text: "Copy Output"
+                }
+              ),
+              /* @__PURE__ */ jsx28(
+                Button3,
+                {
+                  fontSize: 1,
+                  icon: CopyIcon,
+                  mode: "ghost",
+                  onClick: handleCopyLink,
+                  padding: 2,
+                  text: "Copy Link"
+                }
+              )
+            ] })
           ] })
         ] })
       ] })
@@ -7038,7 +7429,7 @@ function JudgmentCard({
 function DocBadge({
   doc
 }) {
-  const [hovered, setHovered] = useState12(false);
+  const [hovered, setHovered] = useState13(false);
   const isLinked = Boolean(doc.documentId);
   const tooltipLabel = isLinked ? `Edit "${doc.title || doc.slug}"` : doc.title || doc.slug;
   const badge = /* @__PURE__ */ jsx28(
@@ -7204,20 +7595,20 @@ import {
   MenuDivider,
   useToast as useToast9
 } from "@sanity/ui";
-import { useCallback as useCallback21, useState as useState16 } from "react";
+import { useCallback as useCallback22, useState as useState17 } from "react";
 import { useClient as useClient9 } from "sanity";
 
 // src/components/report-detail/report-actions/CopyReportAction.tsx
 import { ClipboardIcon } from "@sanity/icons";
 import { MenuItem, useToast as useToast4 } from "@sanity/ui";
-import { useCallback as useCallback16, useState as useState13 } from "react";
+import { useCallback as useCallback17, useState as useState14 } from "react";
 import { useClient as useClient6 } from "sanity";
 import { jsx as jsx31 } from "react/jsx-runtime";
 function CopyReportAction({ documentId }) {
   const client = useClient6({ apiVersion: API_VERSION });
   const toast = useToast4();
-  const [copying, setCopying] = useState13(false);
-  const handleClick = useCallback16(async () => {
+  const [copying, setCopying] = useState14(false);
+  const handleClick = useCallback17(async () => {
     setCopying(true);
     try {
       const doc = await client.fetch(
@@ -7263,11 +7654,11 @@ function CopyReportAction({ documentId }) {
 // src/components/report-detail/report-actions/CopyReportIdAction.tsx
 import { CopyIcon as CopyIcon2 } from "@sanity/icons";
 import { MenuItem as MenuItem2, useToast as useToast5 } from "@sanity/ui";
-import { useCallback as useCallback17 } from "react";
+import { useCallback as useCallback18 } from "react";
 import { jsx as jsx32 } from "react/jsx-runtime";
 function CopyReportIdAction({ reportId }) {
   const toast = useToast5();
-  const handleClick = useCallback17(() => {
+  const handleClick = useCallback18(() => {
     navigator.clipboard.writeText(reportId).then(
       () => {
         toast.push({
@@ -7291,13 +7682,13 @@ function CopyReportIdAction({ reportId }) {
 // src/components/report-detail/report-actions/CopyVisionQueryAction.tsx
 import { SearchIcon as SearchIcon5 } from "@sanity/icons";
 import { MenuItem as MenuItem3, useToast as useToast6 } from "@sanity/ui";
-import { useCallback as useCallback18 } from "react";
+import { useCallback as useCallback19 } from "react";
 import { jsx as jsx33 } from "react/jsx-runtime";
 function CopyVisionQueryAction({
   reportId
 }) {
   const toast = useToast6();
-  const handleClick = useCallback18(() => {
+  const handleClick = useCallback19(() => {
     const query = `*[_type == "ailf.report" && reportId == "${reportId}"][0]`;
     navigator.clipboard.writeText(query).then(
       () => {
@@ -7400,7 +7791,7 @@ function DeleteReportAction({
 // src/components/report-detail/report-actions/DownloadReportAction.tsx
 import { DownloadIcon } from "@sanity/icons";
 import { MenuItem as MenuItem5, useToast as useToast7 } from "@sanity/ui";
-import { useCallback as useCallback19, useState as useState14 } from "react";
+import { useCallback as useCallback20, useState as useState15 } from "react";
 import { useClient as useClient7 } from "sanity";
 import { jsx as jsx36 } from "react/jsx-runtime";
 function DownloadReportAction({
@@ -7409,8 +7800,8 @@ function DownloadReportAction({
 }) {
   const client = useClient7({ apiVersion: API_VERSION });
   const toast = useToast7();
-  const [downloading, setDownloading] = useState14(false);
-  const handleClick = useCallback19(async () => {
+  const [downloading, setDownloading] = useState15(false);
+  const handleClick = useCallback20(async () => {
     setDownloading(true);
     try {
       const doc = await client.fetch(
@@ -7464,7 +7855,7 @@ function DownloadReportAction({
 // src/components/report-detail/report-actions/RerunEvaluationAction.tsx
 import { PlayIcon as PlayIcon2 } from "@sanity/icons";
 import { MenuItem as MenuItem6, useToast as useToast8 } from "@sanity/ui";
-import { useCallback as useCallback20, useState as useState15 } from "react";
+import { useCallback as useCallback21, useState as useState16 } from "react";
 import { useClient as useClient8, useCurrentUser as useCurrentUser3 } from "sanity";
 
 // src/lib/eval-scope.ts
@@ -7519,8 +7910,8 @@ function RerunEvaluationAction({
   const client = useClient8({ apiVersion: API_VERSION });
   const currentUser = useCurrentUser3();
   const toast = useToast8();
-  const [requesting, setRequesting] = useState15(false);
-  const handleClick = useCallback20(async () => {
+  const [requesting, setRequesting] = useState16(false);
+  const handleClick = useCallback21(async () => {
     setRequesting(true);
     try {
       const scope = extractEvalScope(provenance);
@@ -7568,7 +7959,7 @@ function ReportActions({
 }) {
   const client = useClient9({ apiVersion: API_VERSION });
   const toast = useToast9();
-  const handleCopyId = useCallback21(() => {
+  const handleCopyId = useCallback22(() => {
     navigator.clipboard.writeText(reportId).then(
       () => {
         toast.push({
@@ -7586,15 +7977,15 @@ function ReportActions({
       }
     );
   }, [reportId, toast]);
-  const [deleteDialogOpen, setDeleteDialogOpen] = useState16(false);
-  const [deleting, setDeleting] = useState16(false);
-  const handleRequestDelete = useCallback21(() => {
+  const [deleteDialogOpen, setDeleteDialogOpen] = useState17(false);
+  const [deleting, setDeleting] = useState17(false);
+  const handleRequestDelete = useCallback22(() => {
     setDeleteDialogOpen(true);
   }, []);
-  const handleDeleteClose = useCallback21(() => {
+  const handleDeleteClose = useCallback22(() => {
     if (!deleting) setDeleteDialogOpen(false);
   }, [deleting]);
-  const handleDeleteConfirm = useCallback21(async () => {
+  const handleDeleteConfirm = useCallback22(async () => {
     setDeleting(true);
     try {
       await client.delete(documentId);
@@ -7714,9 +8105,9 @@ import { Box as Box19, Flex as Flex23, Stack as Stack25, Text as Text31 } from "
 
 // src/components/report-detail/AreaScoresGrid.tsx
 import React3, {
-  useCallback as useCallback22,
+  useCallback as useCallback23,
   useMemo as useMemo6,
-  useState as useState17
+  useState as useState18
 } from "react";
 import { WarningOutlineIcon as WarningOutlineIcon2 } from "@sanity/icons";
 import { Box as Box18, Flex as Flex21, Stack as Stack24, Text as Text29 } from "@sanity/ui";
@@ -7766,9 +8157,9 @@ function AreaScoresGrid({
   );
   const showLift = isLiteracyMode(mode);
   const dimKeys = useMemo6(() => collectDimensionKeys(scores), [scores]);
-  const [sortField, setSortField] = useState17("score");
-  const [sortDir, setSortDir] = useState17("desc");
-  const handleSort = useCallback22(
+  const [sortField, setSortField] = useState18("score");
+  const [sortDir, setSortDir] = useState18("desc");
+  const handleSort = useCallback23(
     (field) => {
       if (field === sortField) {
         setSortDir((d) => d === "asc" ? "desc" : "asc");
@@ -8251,7 +8642,7 @@ function ColHeader2({
   onClick,
   tooltip
 }) {
-  const handleKeyDown = useCallback22(
+  const handleKeyDown = useCallback23(
     (e) => {
       if (onClick && (e.key === "Enter" || e.key === " ")) {
         e.preventDefault();
@@ -8296,7 +8687,7 @@ function ColHeader2({
 }
 
 // src/components/report-detail/ModelSelector.tsx
-import { useCallback as useCallback23 } from "react";
+import { useCallback as useCallback24 } from "react";
 import { Flex as Flex22, Text as Text30 } from "@sanity/ui";
 import { jsx as jsx41, jsxs as jsxs29 } from "react/jsx-runtime";
 var pillBase = {
@@ -8373,7 +8764,7 @@ function Pill2({
   label,
   onClick
 }) {
-  const handleKeyDown = useCallback23(
+  const handleKeyDown = useCallback24(
     (e) => {
       if (e.key === "Enter" || e.key === " ") {
         e.preventDefault();
@@ -8406,13 +8797,13 @@ function Pill2({
 }
 
 // src/components/report-detail/useModelSelection.ts
-import { useCallback as useCallback24, useMemo as useMemo7, useState as useState18 } from "react";
+import { useCallback as useCallback25, useMemo as useMemo7, useState as useState19 } from "react";
 function useModelSelection({
   scores,
   perModel
 }) {
-  const [selection, setSelection] = useState18(null);
-  const onSelectionChange = useCallback24((next) => {
+  const [selection, setSelection] = useState19(null);
+  const onSelectionChange = useCallback25((next) => {
     setSelection(next);
   }, []);
   const isExpanded = selection === "expanded";
@@ -8915,8 +9306,8 @@ function ReportDetail({
   subTab
 }) {
   const client = useClient10({ apiVersion: API_VERSION });
-  const [loading, setLoading] = useState19(true);
-  const [report, setReport] = useState19(null);
+  const [loading, setLoading] = useState20(true);
+  const [report, setReport] = useState20(null);
   useEffect9(() => {
     let cancelled = false;
     setLoading(true);
@@ -8932,6 +9323,10 @@ function ReportDetail({
       cancelled = true;
     };
   }, [client, reportId]);
+  const artifactCache = useArtifactCache(
+    reportId,
+    report?.summary?.artifacts?.testOutputs
+  );
   const { summary } = report ?? {};
   const hasWeaknesses = Boolean(
     summary?.lowScoringJudgments && summary.lowScoringJudgments.length > 0
@@ -8957,7 +9352,7 @@ function ReportDetail({
     if (disabledTabs.has(parsed)) return "overview";
     return tabs.some((t) => t.id === parsed) ? parsed : "overview";
   }, [activeTab, disabledTabs, tabs]);
-  const handleTabClick = useCallback25(
+  const handleTabClick = useCallback26(
     (tabId) => {
       onTabChange(tabId === "overview" ? null : tabId, null, null);
     },
@@ -9064,6 +9459,7 @@ function ReportDetail({
     currentTab === "diagnostics" && hasDiagnostics && /* @__PURE__ */ jsx44(
       DiagnosticsPanel,
       {
+        artifactCache,
         comparison,
         focus,
         judgments: summary.lowScoringJudgments,
@@ -9071,7 +9467,8 @@ function ReportDetail({
         onNavigate: (newSubTab, newFocus) => onTabChange("diagnostics", newSubTab, newFocus),
         perModel: summary.perModel,
         scores: summary.scores,
-        subTab
+        subTab,
+        testResults: summary.testResults
       }
     ),
     currentTab === "activity" && hasAgentActivity && /* @__PURE__ */ jsx44(
@@ -9100,6 +9497,7 @@ var DIAG_TABS = [
   { id: "issues", label: "Issues" }
 ];
 function DiagnosticsPanel({
+  artifactCache,
   comparison,
   focus,
   judgments,
@@ -9107,7 +9505,8 @@ function DiagnosticsPanel({
   onNavigate,
   perModel,
   scores,
-  subTab: subTabParam
+  subTab: subTabParam,
+  testResults
 }) {
   const subTab = parseDiagSubTab(subTabParam);
   const issueCount = scores.filter((s) => s.totalScore < SCORE_CAUTION).length + scores.filter((s) => s.negativeDocLift).length + scores.filter(
@@ -9186,9 +9585,11 @@ function DiagnosticsPanel({
       judgments && judgments.length > 0 && /* @__PURE__ */ jsx44(
         JudgmentList,
         {
+          artifactCache,
           focus,
           judgments,
-          onFocusChange: (slug) => onNavigate("issues", slug)
+          onFocusChange: (slug) => onNavigate("issues", slug),
+          testResults
         }
       )
     ] })
@@ -9283,7 +9684,7 @@ import { jsx as jsx51, jsxs as jsxs38 } from "react/jsx-runtime";
 
 // src/components/ScoreTimeline.tsx
 import { Card as Card20, Flex as Flex30, Select as Select2, Stack as Stack33, Text as Text40 } from "@sanity/ui";
-import { useCallback as useCallback26, useEffect as useEffect10, useMemo as useMemo10, useState as useState20 } from "react";
+import { useCallback as useCallback27, useEffect as useEffect10, useMemo as useMemo10, useState as useState21 } from "react";
 import { useClient as useClient11 } from "sanity";
 import { jsx as jsx52, jsxs as jsxs39 } from "react/jsx-runtime";
 var CHART_HEIGHT = 220;
@@ -9320,10 +9721,10 @@ function scoreForPoint(point, area) {
 }
 function ScoreTimeline({ mode = null, source = null }) {
   const client = useClient11({ apiVersion: API_VERSION });
-  const [dataPoints, setDataPoints] = useState20([]);
-  const [loading, setLoading] = useState20(true);
-  const [rangeDays, setRangeDays] = useState20(30);
-  const [selectedArea, setSelectedArea] = useState20(null);
+  const [dataPoints, setDataPoints] = useState21([]);
+  const [loading, setLoading] = useState21(true);
+  const [rangeDays, setRangeDays] = useState21(30);
+  const [selectedArea, setSelectedArea] = useState21(null);
   const areaNames = useMemo10(() => {
     const names = /* @__PURE__ */ new Set();
     for (const dp of dataPoints) {
@@ -9333,7 +9734,7 @@ function ScoreTimeline({ mode = null, source = null }) {
     }
     return Array.from(names).sort();
   }, [dataPoints]);
-  const fetchData = useCallback26(async () => {
+  const fetchData = useCallback27(async () => {
     setLoading(true);
     try {
       const startDate = rangeDays ? daysAgo(rangeDays) : "1970-01-01T00:00:00Z";
@@ -9369,14 +9770,14 @@ function ScoreTimeline({ mode = null, source = null }) {
     if (chartPoints.length === 0) return 0;
     return chartPoints.reduce((sum, p) => sum + p.score, 0) / chartPoints.length;
   }, [chartPoints]);
-  const handleRangeChange = useCallback26(
+  const handleRangeChange = useCallback27(
     (e) => {
       const val = e.currentTarget.value;
       setRangeDays(val === "all" ? null : Number(val));
     },
     []
   );
-  const handleAreaChange = useCallback26(
+  const handleAreaChange = useCallback27(
     (e) => {
       const val = e.currentTarget.value;
       setSelectedArea(val || null);
@@ -9522,7 +9923,7 @@ function DashboardContent() {
   const isDetail = reportId !== null;
   const activeTab = isDetail ? "latest" : VIEW_PARAM_MAP[routerState.view ?? ""] ?? "latest";
   const defaultTopic = deriveHelpTopic(routerState);
-  const navigateToTab = useCallback27(
+  const navigateToTab = useCallback28(
     (tab) => {
       if (tab === "latest") {
         router.navigate({});
@@ -9532,13 +9933,13 @@ function DashboardContent() {
     },
     [router]
   );
-  const handleSelectReport = useCallback27(
+  const handleSelectReport = useCallback28(
     (id) => {
       router.navigate({ reportId: id });
     },
     [router]
   );
-  const handleTabChange = useCallback27(
+  const handleTabChange = useCallback28(
     (tab, subTab, focus) => {
       if (!routerState.reportId) return;
       const state = {
@@ -9551,10 +9952,10 @@ function DashboardContent() {
     },
     [router, routerState.reportId]
   );
-  const handleBack = useCallback27(() => {
+  const handleBack = useCallback28(() => {
     router.navigate({});
   }, [router]);
-  const handleOpenHelp = useCallback27(() => {
+  const handleOpenHelp = useCallback28(() => {
     openHelp(defaultTopic);
   }, [openHelp, defaultTopic]);
   return /* @__PURE__ */ jsx53(Container, { width: 4, children: /* @__PURE__ */ jsxs40(Stack34, { padding: 4, space: 4, children: [
@@ -9645,7 +10046,7 @@ function ailfTool(options = {}) {
 // src/actions/RunEvaluationAction.tsx
 import { BarChartIcon as BarChartIcon2 } from "@sanity/icons";
 import { useToast as useToast10 } from "@sanity/ui";
-import { useCallback as useCallback28, useEffect as useEffect11, useRef as useRef6, useState as useState21 } from "react";
+import { useCallback as useCallback29, useEffect as useEffect11, useRef as useRef7, useState as useState22 } from "react";
 import {
   getReleaseIdFromReleaseDocumentId as getReleaseIdFromReleaseDocumentId3,
   useClient as useClient12,
@@ -9676,8 +10077,8 @@ function createRunEvaluationAction(options = {}) {
     const projectId = useProjectId2();
     const currentUser = useCurrentUser4();
     const toast = useToast10();
-    const [state, setState] = useState21({ status: "loading" });
-    const requestedAtRef = useRef6(null);
+    const [state, setState] = useState22({ status: "loading" });
+    const requestedAtRef = useRef7(null);
     const perspectiveId = getReleaseIdFromReleaseDocumentId3(release._id);
     useEffect11(() => {
       let cancelled = false;
@@ -9772,7 +10173,7 @@ function createRunEvaluationAction(options = {}) {
       }, 15e3);
       return () => clearTimeout(timer);
     }, [client, perspectiveId, state]);
-    const handleRequest = useCallback28(async () => {
+    const handleRequest = useCallback29(async () => {
       const releaseTitle = release.metadata?.title ?? perspectiveId ?? "release";
       const tag = `release-${slugify3(releaseTitle)}-${dateStamp3()}`;
       const now = Date.now();

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@sanity/ailf-studio",
-  "version": "1.2.2",
+  "version": "1.3.1",
   "description": "AI Literacy Framework — Sanity Studio dashboard plugin",
   "type": "module",
   "license": "MIT",