@tangle-network/agent-eval 0.203.1 → 0.204.0

package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/typecheck.log

@@ -0,0 +1 @@
+$ tsc --noEmit

package/docs/proofs/campaign-attempt-retention-20261001/artifacts/current/verify-package.log

@@ -0,0 +1,340 @@
+$ pnpm check:analyst-benchmark && pnpm run check:skill && pnpm run check:model-ids && pnpm run check:canonical-json && pnpm run check:collation-ordering && publint && attw --pack --profile esm-only . && node scripts/verify-package-exports.mjs && pnpm run contract:finding:check && pnpm run evidence:check
+$ node scripts/check-analyst-benchmark-implementation.mjs
+public analyst benchmark digests valid: implementation 65afefb4ed822b6400865ba85817d45826a6351209b8509deb07b3cabb059d70 (111 files), dependency lock e57c71c31ecb3ec79d6cf08dc6f1603aa8109bafee4208e63f0e3a1234619d5a (4 files)
+$ node scripts/check-skill.mjs
+agent-eval skill valid: 89 description chars, 8957 bytes
+$ node scripts/check-model-id-requests.mjs
+model id request gate valid: every model id under src is either asserted against the served id or waived (6 waivers, all matched)
+$ node scripts/check-canonical-json.mjs
+canonical json gate valid: src/ledger-core/canonical.ts is the only encoder under src/
+$ node scripts/check-collation-ordering.mjs
+collation ordering gate valid: no canonicalizing function orders with localeCompare
+Running publint v0.3.24 for @tangle-network/agent-eval...
+Packing files with `pnpm pack`...
+Linting...
+All good!
+
+@tangle-network/agent-eval v0.203.1
+
+Build tools:
+- @arethetypeswrong/cli@^0.18.5
+- typescript@^7.0.2
+- esbuild@^0.28.2
+- tsdown@^0.23.0
+
+ (ignoring resolutions: 'node10', 'node16-cjs')
+
+(ignored per resolution) ⚠️ A require call resolved to an ESM JavaScript file, which is an error in Node and some bundlers. CommonJS consumers will need to use a dynamic import. https://github.com/arethetypeswrong/arethetypeswrong.github.io/blob/main/docs/problems/CJSResolvesToESM.md
+
+(ignored per resolution) 💀 Import failed to resolve to type declarations or JavaScript files. https://github.com/arethetypeswrong/arethetypeswrong.github.io/blob/main/docs/problems/NoResolution.md
+
+
+"@tangle-network/agent-eval"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 🟢 
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/jev"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/profile-cell"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/analyst"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/reporting"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/rl"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/fuzz"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/traces"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/trace-attributes"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/wire"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/benchmarks"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/pipelines"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/meta-eval"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/builder-eval"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/matrix"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/multishot"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/multishot/golden"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/campaign"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/search"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/ledger-core"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/authenticity"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/contract"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/rollout"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/supervisor-run"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/hosted"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/trace-repair"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/diagnosis"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/experiment"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/trajectory-replay"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/openapi.json"
+
+node16 (from ESM): 🟢 (JSON)
+bundler: 🟢 (JSON)
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) 🟢 (JSON)
+
+***********************************
+
+"@tangle-network/agent-eval/storyboard"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/adapters/http"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/evaluation"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+
+"@tangle-network/agent-eval/jev/protocol"
+
+node16 (from ESM): 🟢 (ESM)
+bundler: 🟢 
+node10: (ignored) 💀 Resolution failed
+node16 (from CJS): (ignored) ⚠️ ESM (dynamic import only)
+
+***********************************
+$ tsx scripts/emit-finding-contract.ts --check
+finding contract is fresh: version 1, 22 subject patterns
+$ tsx scripts/render-evidence-index.ts --check
+evidence index is fresh: 10 records, index matches

package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-legacy/capture-retention_0/attempts/33d57ccb-d4d8-4c23-92e3-c1ddf7db6058/attempt-1/artifacts/captured-response.json

@@ -0,0 +1,4 @@
+{
+  "continuation": true,
+  "runAttemptId": "33d57ccb-d4d8-4c23-92e3-c1ddf7db6058"
+}

package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-legacy/capture-retention_0/attempts/33d57ccb-d4d8-4c23-92e3-c1ddf7db6058/attempt-1/identity.json

@@ -0,0 +1,10 @@
+{
+  "runAttemptId": "33d57ccb-d4d8-4c23-92e3-c1ddf7db6058",
+  "number": 1,
+  "manifestHash": "6d4d16fdd6cbaa1119fc301e329b128e6485461e9bc013bdbeaea9ec0270b19c",
+  "cellId": "capture-retention:0",
+  "scenarioId": "capture-retention",
+  "rep": 0,
+  "seed": 42,
+  "startedAt": "2026-10-01T17:09:52.253Z"
+}

package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-legacy/capture-retention_0/attempts/33d57ccb-d4d8-4c23-92e3-c1ddf7db6058/attempt-1/result.json

@@ -0,0 +1 @@
+{"manifestHash":"6d4d16fdd6cbaa1119fc301e329b128e6485461e9bc013bdbeaea9ec0270b19c","attempt":{"runAttemptId":"33d57ccb-d4d8-4c23-92e3-c1ddf7db6058","number":1},"cellId":"capture-retention:0","scenarioId":"capture-retention","rep":0,"artifact":{"continuation":true},"judgeScores":{},"costUsd":0,"costProvenance":{"kind":"observed","usd":0},"costCallIds":[],"tokenUsage":{"input":0,"output":0},"durationMs":0,"seed":42,"cached":false}

package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-legacy/capture-retention_0/attempts/33d57ccb-d4d8-4c23-92e3-c1ddf7db6058/attempt-1/trace/spans.jsonl

@@ -0,0 +1 @@
+{"name":"legacy.continuation","cellId":"capture-retention:0","startMs":1790874592253,"durationMs":0}

package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-legacy/capture-retention_0/cached-result.json

@@ -0,0 +1 @@
+{"manifestHash":"6e492a4dc01bf06d2f7e169cdde455d0b0416e34bd4396c1ec7f72f795d57fa7","cellId":"capture-retention:0","scenarioId":"capture-retention","rep":0,"artifact":{"cellId":"capture-retention:0","runAttemptId":"cc7546fa-7b18-4a50-8444-204439c84bae","invocation":2,"content":"Actual filesystem capture from invocation 2"},"judgeScores":{},"costUsd":0,"costProvenance":{"kind":"observed","usd":0},"costCallIds":[],"tokenUsage":{"input":0,"output":0},"durationMs":256,"seed":42,"cached":false}

package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-legacy/capture-retention_0/captured-response.json

@@ -0,0 +1,6 @@
+{
+  "cellId": "capture-retention:0",
+  "runAttemptId": "cc7546fa-7b18-4a50-8444-204439c84bae",
+  "invocation": 2,
+  "content": "Actual filesystem capture from invocation 2"
+}

package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-legacy/capture-retention_0/failure-receipt.json

@@ -0,0 +1,58 @@
+{
+  "schemaVersion": 1,
+  "runAttemptId": "08005dbd-35ff-488f-8942-22732fc5f036",
+  "recordedAt": "2026-10-01T16:47:43.723Z",
+  "failure": {
+    "stage": "dispatch",
+    "error": {
+      "name": "Error",
+      "message": "Operational interruption after native artifact capture",
+      "stack": "Error: Operational interruption after native artifact capture\n    at Object.dispatch (file:///home/drew/webb/_wt/lab-autoroute-campaign-20261001/[eval1]:16:22)\n    at async executeCell (file:///home/drew/webb/_wt/lab-evidence-unified/node_modules/.pnpm/@tangle-network+agent-eval@0.203.0/node_modules/@tangle-network/agent-eval/dist/llm-judge-C-0P1WiT.js:459:19)\n    at async file:///home/drew/webb/_wt/lab-evidence-unified/node_modules/.pnpm/@tangle-network+agent-eval@0.203.0/node_modules/@tangle-network/agent-eval/dist/llm-judge-C-0P1WiT.js:888:15\n    at async Promise.allSettled (index 0)\n    at async runCampaign (file:///home/drew/webb/_wt/lab-evidence-unified/node_modules/.pnpm/@tangle-network+agent-eval@0.203.0/node_modules/@tangle-network/agent-eval/dist/llm-judge-C-0P1WiT.js:935:22)\n    at async file:///home/drew/webb/_wt/lab-autoroute-campaign-20261001/[eval1]:19:17"
+    }
+  },
+  "cell": {
+    "manifestHash": "6e492a4dc01bf06d2f7e169cdde455d0b0416e34bd4396c1ec7f72f795d57fa7",
+    "cellId": "capture-retention:0",
+    "scenarioId": "capture-retention",
+    "rep": 0,
+    "artifact": null,
+    "judgeScores": {},
+    "costUsd": 0,
+    "costProvenance": {
+      "kind": "observed",
+      "usd": 0
+    },
+    "costCallIds": [],
+    "tokenUsage": {
+      "input": 0,
+      "output": 0
+    },
+    "durationMs": 1,
+    "seed": 42,
+    "cached": false,
+    "errorStage": "dispatch",
+    "error": "Operational interruption after native artifact capture"
+  },
+  "cost": {
+    "totalCalls": 0,
+    "pendingCalls": 0,
+    "unresolvedCalls": 0,
+    "reservedCostUsd": 0,
+    "inputTokens": 0,
+    "outputTokens": 0,
+    "reasoningTokens": 0,
+    "cachedTokens": 0,
+    "cacheWriteTokens": 0,
+    "totalCostUsd": 0,
+    "costProvenance": {
+      "kind": "observed",
+      "usd": 0
+    },
+    "byChannel": [],
+    "unpricedModels": [],
+    "fullyPriced": true,
+    "usageComplete": true,
+    "accountingComplete": true,
+    "incompleteReasons": []
+  }
+}

package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-legacy/capture-retention_0/latest-attempt.json

@@ -0,0 +1 @@
+{"runAttemptId":"33d57ccb-d4d8-4c23-92e3-c1ddf7db6058","number":1}

package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-legacy/capture-retention_0/spans.jsonl

@@ -0,0 +1 @@
+{"name":"capture-invocation","cellId":"capture-retention:0","startMs":1790873263979,"invocation":2,"durationMs":0}

package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-optimization/baseline/capture-retention_0/attempts/6b25a1d2-5324-445b-b59d-bdc7381cd4b4/attempt-1/artifacts/captured-response.json

@@ -0,0 +1,4 @@
+{
+  "surface": "baseline",
+  "runAttemptId": "6b25a1d2-5324-445b-b59d-bdc7381cd4b4"
+}

package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-optimization/baseline/capture-retention_0/attempts/6b25a1d2-5324-445b-b59d-bdc7381cd4b4/attempt-1/identity.json

@@ -0,0 +1,10 @@
+{
+  "runAttemptId": "6b25a1d2-5324-445b-b59d-bdc7381cd4b4",
+  "number": 1,
+  "manifestHash": "2797bc683f6a3ccb516b5e73664a0e57e15e99cf45de01ab85b897950d51aeb3",
+  "cellId": "capture-retention:0",
+  "scenarioId": "capture-retention",
+  "rep": 0,
+  "seed": 42,
+  "startedAt": "2026-10-01T17:09:53.561Z"
+}

package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-optimization/baseline/capture-retention_0/attempts/6b25a1d2-5324-445b-b59d-bdc7381cd4b4/attempt-1/result.json

@@ -0,0 +1 @@
+{"manifestHash":"2797bc683f6a3ccb516b5e73664a0e57e15e99cf45de01ab85b897950d51aeb3","attempt":{"runAttemptId":"6b25a1d2-5324-445b-b59d-bdc7381cd4b4","number":1},"cellId":"capture-retention:0","scenarioId":"capture-retention","rep":0,"artifact":{"surface":"baseline"},"judgeScores":{},"costUsd":0,"costProvenance":{"kind":"observed","usd":0},"costCallIds":[],"tokenUsage":{"input":0,"output":0},"durationMs":0,"seed":42,"cached":false}

package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-optimization/baseline/capture-retention_0/attempts/6b25a1d2-5324-445b-b59d-bdc7381cd4b4/attempt-1/trace/spans.jsonl

@@ -0,0 +1 @@
+{"name":"native.optimization.capture","cellId":"capture-retention:0","startMs":1790874593561,"surface":"baseline","durationMs":0}

package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-optimization/baseline/capture-retention_0/cached-result.json

@@ -0,0 +1 @@
+{"manifestHash":"2797bc683f6a3ccb516b5e73664a0e57e15e99cf45de01ab85b897950d51aeb3","attempt":{"runAttemptId":"6b25a1d2-5324-445b-b59d-bdc7381cd4b4","number":1},"cellId":"capture-retention:0","scenarioId":"capture-retention","rep":0,"artifact":{"surface":"baseline"},"judgeScores":{},"costUsd":0,"costProvenance":{"kind":"observed","usd":0},"costCallIds":[],"tokenUsage":{"input":0,"output":0},"durationMs":0,"seed":42,"cached":false}

package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-optimization/baseline/capture-retention_0/latest-attempt.json

@@ -0,0 +1 @@
+{"runAttemptId":"6b25a1d2-5324-445b-b59d-bdc7381cd4b4","number":1}

package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-optimization/first-interruption.json

@@ -0,0 +1,5 @@
+{
+  "firstError": "campaignMeanComposite: campaign has no complete cell-quality scores",
+  "proposals": 0,
+  "histories": []
+}

package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-optimization/search/blobs/5de45d632937e75ff90fec9f1f2ec425904d3632da71f6b26f83b1e1eb8cde5a.json

@@ -0,0 +1 @@
+{"kind":"mutable-surface","surface":"baseline"}

package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-optimization/search/ledger.jsonl

@@ -0,0 +1,7 @@
+{"entryHash":"sha256:76319ccbd5c372d04f6b18142621b692b774142b1e9e5a65e4c9420da28e629a","event":{"artifactKind":"prompt","artifacts":[],"budget":{"deadline":null,"maxCells":null,"maxConcurrency":null,"maxNodes":3,"maxUsd":null,"reservedClaimUsd":0},"containment":null,"derivedFrom":null,"eventId":"search-opened","identity":{"agent":{"revision":"sha256:e5e9fb9a702bdbda8979b9618aec28f5d4952695f702d8aab98d1716fa55b966","uri":"dispatch-ref:operational-attempt-retention-v1"},"benchmark":{"revision":"sha256:23ea149e81be5435966bd1d185776c7d2e119739830f98a47d4738bdbd46ade6","uri":"campaign://scenarios"},"model":{"alias":"unspecified","provider":"unspecified","unknown":"runOptimization was not told which model the agent runs"}},"kind":"search-opened","objective":{"claim":{"generalization":"fixed-roster","independentUnit":"taskId","population":{"description":"the tasks this search was given","id":"sha256:23ea149e81be5435966bd1d185776c7d2e119739830f98a47d4738bdbd46ade6"},"samplingFrame":"the caller-supplied task list","use":"development"},"direction":"maximize","judge":{"unknown":"runOptimization receives its judges as functions without a pinned source"},"metric":"composite"},"occurredAt":"2026-10-01T17:09:52.342Z","policy":{"allocation":"uniform","expansion":"incumbent","seed":42},"process":{"executionRef":{"revision":"sha256:50e5c663f35dcf1e181f9add0ac7483881ac9f3a99197986ef973fe259d57072","uri":"npm:@tangle-network/agent-eval#runSearch"},"name":"runOptimization"},"splits":{"heldOutUnits":true,"selection":{"taskSetDigest":"sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945","tasks":[]},"test":{"taskSetDigest":"sha256:4f53cda18c2baa0c0354bb5f9a3ecbe5ed12ab4d8e11ba873c2f11161202b945","tasks":[]},"train":{"taskSetDigest":"sha256:7c4dc4726ad6422de11b95ea1bb36feb66db38af317d018ae2f2070eca98640a","tasks":[{"source":{"revision":"sha256:70a52f79806918ce9e4833c4ed81360245988fced43a4605ab703d7afacb20a3","uri":"scenario://capture-retention"},"taskId":"capture-retention","unitId":"capture-retention"}]}},"subject":"operational-retention-qualification"},"previousHash":null,"schema":"tangle.search-ledger.2026-09","searchId":"optimization-4064971491bb5e3a","sequence":0}
+{"entryHash":"sha256:e838b065367b2bdbceedd76d922ef754986e871130f81980c1450343847ddfc2","event":{"artifact":{"byteLength":47,"role":"surface","sha256":"sha256:5de45d632937e75ff90fec9f1f2ec425904d3632da71f6b26f83b1e1eb8cde5a","uri":"file:///mnt/traces/eval-attempt-artifacts-20261001/after-optimization/search/blobs/5de45d632937e75ff90fec9f1f2ec425904d3632da71f6b26f83b1e1eb8cde5a.json"},"artifactDigest":"sha256:8ba8496a2525ae171ffd104d632dede6ef418d9b95962a9d88e2fcdbc8d48d24","artifacts":[],"eventId":"node:node_81d791a36cc9a9bea6c2c497b1a0540f","kind":"node-registered","nodeId":"node_81d791a36cc9a9bea6c2c497b1a0540f","occurredAt":"2026-10-01T17:09:52.661Z","surfaces":[{"artifact":{"byteLength":47,"role":"surface","sha256":"sha256:5de45d632937e75ff90fec9f1f2ec425904d3632da71f6b26f83b1e1eb8cde5a","uri":"file:///mnt/traces/eval-attempt-artifacts-20261001/after-optimization/search/blobs/5de45d632937e75ff90fec9f1f2ec425904d3632da71f6b26f83b1e1eb8cde5a.json"},"kind":"prompt","surfaceId":"prompt"}]},"previousHash":"sha256:76319ccbd5c372d04f6b18142621b692b774142b1e9e5a65e4c9420da28e629a","schema":"tangle.search-ledger.2026-09","searchId":"optimization-4064971491bb5e3a","sequence":1}
+{"entryHash":"sha256:4ab8c743a015c4af3df0e3fff532b9d2bcb90e6ad5a592b5cef1507a15ebacae","event":{"artifacts":[],"attribution":"explicit","childNodeId":"node_81d791a36cc9a9bea6c2c497b1a0540f","diffs":[],"edgeId":"edge_ce3ed2991a4ac0d9f298b75cd4174071","eventId":"edge:edge_ce3ed2991a4ac0d9f298b75cd4174071","kind":"edge-recorded","label":"root","occurredAt":"2026-10-01T17:09:52.909Z","operator":"seed","parents":[],"proposer":null,"rationale":{"unknown":"the root is the caller-supplied starting artifact"},"selection":null},"previousHash":"sha256:e838b065367b2bdbceedd76d922ef754986e871130f81980c1450343847ddfc2","schema":"tangle.search-ledger.2026-09","searchId":"optimization-4064971491bb5e3a","sequence":2}
+{"entryHash":"sha256:cc294fa828b5e2ccde590f70dd69e54ac0062a0d63b3d96b851efacad1eac93e","event":{"artifacts":[],"cellId":"cell_5083c7fe2d187178a34b95db92434caf","eventId":"cell:cell_5083c7fe2d187178a34b95db92434caf","kind":"cell-allocated","lane":"campaign","nodeId":"node_81d791a36cc9a9bea6c2c497b1a0540f","occurredAt":"2026-10-01T17:09:53.215Z","rep":0,"reservation":{"kind":"estimate","usd":0},"split":"train","stage":"root","taskId":"capture-retention","unitId":"capture-retention"},"previousHash":"sha256:4ab8c743a015c4af3df0e3fff532b9d2bcb90e6ad5a592b5cef1507a15ebacae","schema":"tangle.search-ledger.2026-09","searchId":"optimization-4064971491bb5e3a","sequence":3}
+{"entryHash":"sha256:3f61619e85e3a03f4859fb6a2b887c207c09a60b1acc03acfa910546cc5a2522","event":{"accounting":{"cost":{"source":"provider","status":"known","usd":0},"tokens":{"cachedTokens":0,"inputTokens":0,"outputTokens":0,"status":"known"}},"artifacts":[],"attempt":1,"boxMinutes":null,"cellId":"cell_5083c7fe2d187178a34b95db92434caf","eventId":"settled:cell_5083c7fe2d187178a34b95db92434caf:1","identity":{"agent":{"revision":"sha256:e5e9fb9a702bdbda8979b9618aec28f5d4952695f702d8aab98d1716fa55b966","uri":"dispatch-ref:operational-attempt-retention-v1"},"benchmark":{"revision":"sha256:23ea149e81be5435966bd1d185776c7d2e119739830f98a47d4738bdbd46ade6","uri":"campaign://scenarios"},"model":{"alias":"unspecified","provider":"unspecified","unknown":"runOptimization was not told which model the agent runs"}},"kind":"cell-settled","occurredAt":"2026-10-01T17:09:53.820Z","outcome":{"error":{"code":"unscored","message":"the cell produced no complete judge score","retryable":false},"metrics":{},"status":"errored"},"placement":{"boxId":null,"lane":"campaign"},"queueMs":null,"runId":"cell_5083c7fe2d187178a34b95db92434caf:1","surfaceEvidence":[],"traceRef":{"unknown":"runCampaign reports no trace id per cell"},"wallMs":0},"previousHash":"sha256:cc294fa828b5e2ccde590f70dd69e54ac0062a0d63b3d96b851efacad1eac93e","schema":"tangle.search-ledger.2026-09","searchId":"optimization-4064971491bb5e3a","sequence":4}
+{"entryHash":"sha256:1c6d05179a3d10baa0e7e6f87ac849fd7d1e54420b75a545125a0cdb68f89c91","event":{"artifacts":[],"eventId":"operation-started:expand-0","kind":"operation-started","occurredAt":"2026-10-01T17:09:57.204Z","operationId":"expand-0","operationKind":"candidate-generation","reservation":{"kind":"estimate","usd":0}},"previousHash":"sha256:3f61619e85e3a03f4859fb6a2b887c207c09a60b1acc03acfa910546cc5a2522","schema":"tangle.search-ledger.2026-09","searchId":"optimization-4064971491bb5e3a","sequence":5}
+{"entryHash":"sha256:ceb6f122da3a83352b61534b6dafd3dd09d63d9b71a60f8553f2c900d03a88c1","event":{"accounting":{"cost":{"knownLowerBoundUsd":0,"reason":"the result of this work was lost, so its spend is unknown","status":"unknown"},"tokens":{"reason":"the result of this work was lost","status":"unknown"}},"artifacts":[],"eventId":"operation-recorded:expand-0","execution":{"kind":"deterministic","source":{"revision":"sha256:5cc7609eb1e31694d3c9f51c8be6ee9ed3fd61fce45b5da8311b00e2159c45d5","uri":"proposer:operational-retention-qualification"}},"kind":"operation-recorded","occurredAt":"2026-10-01T17:10:01.843Z","operationId":"expand-0","operationKind":"candidate-generation","outcome":{"failure":{"code":"proposer-error","message":"campaignMeanComposite: campaign has no complete cell-quality scores"},"status":"failed"}},"previousHash":"sha256:1c6d05179a3d10baa0e7e6f87ac849fd7d1e54420b75a545125a0cdb68f89c91","schema":"tangle.search-ledger.2026-09","searchId":"optimization-4064971491bb5e3a","sequence":6}

package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-optimization/search/ledger.jsonl.head

@@ -0,0 +1 @@
+{"entryHash":"sha256:ceb6f122da3a83352b61534b6dafd3dd09d63d9b71a60f8553f2c900d03a88c1","sequence":6}

package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-path-guard/capture-retention_0/attempts/723fa9c6-099e-437c-bddf-a4cbd28a8508/attempt-1/artifacts/capture.json

@@ -0,0 +1,3 @@
+{
+  "own": true
+}

package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-path-guard/capture-retention_0/attempts/723fa9c6-099e-437c-bddf-a4cbd28a8508/attempt-1/identity.json

@@ -0,0 +1,10 @@
+{
+  "runAttemptId": "723fa9c6-099e-437c-bddf-a4cbd28a8508",
+  "number": 1,
+  "manifestHash": "6d4d16fdd6cbaa1119fc301e329b128e6485461e9bc013bdbeaea9ec0270b19c",
+  "cellId": "capture-retention:0",
+  "scenarioId": "capture-retention",
+  "rep": 0,
+  "seed": 42,
+  "startedAt": "2026-10-01T17:09:52.254Z"
+}

package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-path-guard/capture-retention_0/attempts/723fa9c6-099e-437c-bddf-a4cbd28a8508/attempt-1/result.json

@@ -0,0 +1 @@
+{"manifestHash":"6d4d16fdd6cbaa1119fc301e329b128e6485461e9bc013bdbeaea9ec0270b19c","attempt":{"runAttemptId":"723fa9c6-099e-437c-bddf-a4cbd28a8508","number":1},"cellId":"capture-retention:0","scenarioId":"capture-retention","rep":0,"artifact":{"rejection":"Artifact path must remain inside this execution attempt","identityBefore":"436843a42ed9523fa6827cac58c18e5eb2eb5ce8364888ee60db96d58efbb106","identityAfter":"436843a42ed9523fa6827cac58c18e5eb2eb5ce8364888ee60db96d58efbb106"},"judgeScores":{},"costUsd":0,"costProvenance":{"kind":"observed","usd":0},"costCallIds":[],"tokenUsage":{"input":0,"output":0},"durationMs":1,"seed":42,"cached":false}

package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-path-guard/capture-retention_0/cached-result.json

@@ -0,0 +1 @@
+{"manifestHash":"6d4d16fdd6cbaa1119fc301e329b128e6485461e9bc013bdbeaea9ec0270b19c","attempt":{"runAttemptId":"723fa9c6-099e-437c-bddf-a4cbd28a8508","number":1},"cellId":"capture-retention:0","scenarioId":"capture-retention","rep":0,"artifact":{"rejection":"Artifact path must remain inside this execution attempt","identityBefore":"436843a42ed9523fa6827cac58c18e5eb2eb5ce8364888ee60db96d58efbb106","identityAfter":"436843a42ed9523fa6827cac58c18e5eb2eb5ce8364888ee60db96d58efbb106"},"judgeScores":{},"costUsd":0,"costProvenance":{"kind":"observed","usd":0},"costCallIds":[],"tokenUsage":{"input":0,"output":0},"durationMs":1,"seed":42,"cached":false}

package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-path-guard/capture-retention_0/latest-attempt.json

@@ -0,0 +1 @@
+{"runAttemptId":"723fa9c6-099e-437c-bddf-a4cbd28a8508","number":1}

package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-resume/capture-retention_0/attempts/06db3429-c5c6-42c1-8623-08e2da1f563e/attempt-1/artifacts/captured-response.json

@@ -0,0 +1,6 @@
+{
+  "cellId": "capture-retention:0",
+  "runAttemptId": "06db3429-c5c6-42c1-8623-08e2da1f563e",
+  "invocation": 1,
+  "content": "Actual filesystem capture from invocation 1"
+}

package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-resume/capture-retention_0/attempts/06db3429-c5c6-42c1-8623-08e2da1f563e/attempt-1/failure-receipt.json

@@ -0,0 +1,62 @@
+{
+  "schemaVersion": 1,
+  "runAttemptId": "06db3429-c5c6-42c1-8623-08e2da1f563e",
+  "recordedAt": "2026-10-01T17:09:52.222Z",
+  "failure": {
+    "stage": "dispatch",
+    "error": {
+      "name": "Error",
+      "message": "Operational interruption after native artifact capture",
+      "stack": "Error: Operational interruption after native artifact capture\n    at Object.dispatch (file:///mnt/traces/eval-attempt-artifacts-20261001/qualify.mjs:28:33)\n    at async executeCell (file:///home/drew/code/_wt/eval-router-cache-usage-20261001/dist/llm-judge-BGJ-FcTy.js:487:19)\n    at async file:///home/drew/code/_wt/eval-router-cache-usage-20261001/dist/llm-judge-BGJ-FcTy.js:919:15\n    at async Promise.allSettled (index 0)\n    at async runCampaign (file:///home/drew/code/_wt/eval-router-cache-usage-20261001/dist/llm-judge-BGJ-FcTy.js:966:22)\n    at async qualify (file:///mnt/traces/eval-attempt-artifacts-20261001/qualify.mjs:34:18)\n    at async file:///mnt/traces/eval-attempt-artifacts-20261001/qualify.mjs:70:128"
+    }
+  },
+  "cell": {
+    "manifestHash": "6d4d16fdd6cbaa1119fc301e329b128e6485461e9bc013bdbeaea9ec0270b19c",
+    "attempt": {
+      "runAttemptId": "06db3429-c5c6-42c1-8623-08e2da1f563e",
+      "number": 1
+    },
+    "cellId": "capture-retention:0",
+    "scenarioId": "capture-retention",
+    "rep": 0,
+    "artifact": null,
+    "judgeScores": {},
+    "costUsd": 0,
+    "costProvenance": {
+      "kind": "observed",
+      "usd": 0
+    },
+    "costCallIds": [],
+    "tokenUsage": {
+      "input": 0,
+      "output": 0
+    },
+    "durationMs": 0,
+    "seed": 42,
+    "cached": false,
+    "errorStage": "dispatch",
+    "error": "Operational interruption after native artifact capture"
+  },
+  "cost": {
+    "totalCalls": 0,
+    "pendingCalls": 0,
+    "unresolvedCalls": 0,
+    "reservedCostUsd": 0,
+    "inputTokens": 0,
+    "outputTokens": 0,
+    "reasoningTokens": 0,
+    "cachedTokens": 0,
+    "cacheWriteTokens": 0,
+    "totalCostUsd": 0,
+    "costProvenance": {
+      "kind": "observed",
+      "usd": 0
+    },
+    "byChannel": [],
+    "unpricedModels": [],
+    "fullyPriced": true,
+    "usageComplete": true,
+    "accountingComplete": true,
+    "incompleteReasons": []
+  }
+}

package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-resume/capture-retention_0/attempts/06db3429-c5c6-42c1-8623-08e2da1f563e/attempt-1/identity.json

@@ -0,0 +1,10 @@
+{
+  "runAttemptId": "06db3429-c5c6-42c1-8623-08e2da1f563e",
+  "number": 1,
+  "manifestHash": "6d4d16fdd6cbaa1119fc301e329b128e6485461e9bc013bdbeaea9ec0270b19c",
+  "cellId": "capture-retention:0",
+  "scenarioId": "capture-retention",
+  "rep": 0,
+  "seed": 42,
+  "startedAt": "2026-10-01T17:09:52.222Z"
+}

package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-resume/capture-retention_0/attempts/06db3429-c5c6-42c1-8623-08e2da1f563e/attempt-1/result.json

@@ -0,0 +1 @@
+{"manifestHash":"6d4d16fdd6cbaa1119fc301e329b128e6485461e9bc013bdbeaea9ec0270b19c","attempt":{"runAttemptId":"06db3429-c5c6-42c1-8623-08e2da1f563e","number":1},"cellId":"capture-retention:0","scenarioId":"capture-retention","rep":0,"artifact":null,"judgeScores":{},"costUsd":0,"costProvenance":{"kind":"observed","usd":0},"costCallIds":[],"tokenUsage":{"input":0,"output":0},"durationMs":0,"seed":42,"cached":false,"errorStage":"dispatch","error":"Operational interruption after native artifact capture"}

package/docs/proofs/campaign-attempt-retention-20261001/artifacts/predecessor/after-resume/capture-retention_0/attempts/06db3429-c5c6-42c1-8623-08e2da1f563e/attempt-1/trace/spans.jsonl

@@ -0,0 +1 @@
+{"name":"native.callback.capture","cellId":"capture-retention:0","startMs":1790874592222,"invocation":1,"runAttemptId":"06db3429-c5c6-42c1-8623-08e2da1f563e","durationMs":0}