fix(learning): preserve exact candidates and research evidence

.agent/critical-audit/2026-09-05-learning-system/runtime-inventory.json

@@ -0,0 +1,53 @@
+{
+  "schemaVersion": 1,
+  "auditId": "2026-09-05-learning-system",
+  "repository": "agent-runtime",
+  "initialCommit": "a16d8a3b91481b140cb552e373d5bde98b34af05",
+  "status": "historical_counts_preserved_original_list_unavailable",
+  "listUnavailable": true,
+  "paths": null,
+  "perFileHashes": null,
+  "coverageByFile": null,
+  "counts": {
+    "previouslyObservedUniquePaths": 55,
+    "production": 38,
+    "tests": 7,
+    "callers": 5,
+    "documentation": 5,
+    "productionFilesUnderSrcImprovementCompletelyRead": 24
+  },
+  "countConsistency": {
+    "sum": 55,
+    "expected": 55,
+    "consistent": true
+  },
+  "evidence": {
+    "kind": "Previously verified inventory summarized in the session continuation record.",
+    "originalPath": "/tmp/learning-system-audit-20260905/runtime-inventory.json",
+    "originalAvailable": false,
+    "loss": "The environment refresh removed the scratch file before artifact persistence.",
+    "durableScopeDescription": "https://github.com/tangle-network/agent-runtime/blob/7fe89641322dc1d0e60dc8c4d66472d786320558/docs/research/learning-system-audit-2026-09-05.md#L34",
+    "durableMainReport": {
+      "path": "docs/research/learning-system-audit-2026-09-05.md",
+      "sha256": "a093f4b696dc211b485174f0c8f29b9a0c22d255bee968b1219bcd1425494458",
+      "commit": "7fe89641322dc1d0e60dc8c4d66472d786320558",
+      "url": "https://github.com/tangle-network/agent-runtime/blob/7fe89641322dc1d0e60dc8c4d66472d786320558/docs/research/learning-system-audit-2026-09-05.md"
+    }
+  },
+  "limitations": [
+    "The original 55-path list and per-file hashes cannot be recovered from the retained evidence.",
+    "No list of the remaining 31 files has been reconstructed.",
+    "The 24-file complete-read count is a prior audit observation, not a new read performed while creating this artifact.",
+    "External caller inspections used workspace snapshots rather than fetched main."
+  ],
+  "latestIntegration": {
+    "mainCommit": "2707e2321e7b26b0e71efacb86ebdf3ddd23adac",
+    "integrationCommit": "7f05654a69441db1bd135e8545787af785cde4a2",
+    "changedFiles": [
+      "AGENTS.md",
+      "CLAUDE.md"
+    ],
+    "reportAndCohortCommit": "7fe89641322dc1d0e60dc8c4d66472d786320558",
+    "scope": "Instruction-only main integration; this metadata update does not add files to the original audit inventory."
+  }
+}

.agent/skill-runs.jsonl

@@ -0,0 +1,2 @@
+{"timestamp":"2026-09-06 05:05:12 UTC","auditId":"2026-09-05-learning-system","skill":"/critical-audit","target":"agent-runtime, agent-eval, agent-knowledge learning process; 22 source findings; Runtime prior inventory n=55 files","verdict":"APPROVE","verdictScope":"The 22 identified source defects and their repaired regression coverage.","artifactPath":".agent/critical-audit/2026-09-05-learning-system","releaseStatus":{"eval":"0.174.0 live; retained owner proof embedded in manifest","runtime":"pending parent proof","knowledge":"pending owner proof for new dependency cohort"},"evidenceLimits":["Prior local scratch logs and original Runtime file inventory are unavailable after environment refresh.","Source repair approval does not establish continuing learning, meta-learning, or SOTA performance."],"next":"/stop","nextOwner":"Parent integration and release work continues independently."}
+{"timestamp":"2026-09-06T05:25:08.806615+00:00","auditId":"2026-09-05-learning-system","skill":"/critical-audit","event":"release-proof-update","target":"Retain terminal Knowledge 14.0.0 proof and current Runtime integration/local package evidence","verdict":"APPROVE","verdictScope":"Preserves the existing 22-source-defect repair review; no new source review executed.","artifactPath":".agent/critical-audit/2026-09-05-learning-system","releaseStatus":{"eval":"0.174.0 live","knowledge":"14.0.0 live; owner registry proof and terminal CI retained","runtime":"local package/cohort checks passed; CI, merge, publication pending"},"durableReportCommit":"7fe89641322dc1d0e60dc8c4d66472d786320558","runtimeMainCommit":"2707e2321e7b26b0e71efacb86ebdf3ddd23adac","runtimeIntegrationCommit":"7f05654a69441db1bd135e8545787af785cde4a2","knowledgeRegressionAnchorsPinned":8,"evidenceLimits":["Original failed runs and unavailable scratch observations remain distinct.","Knowledge consumer Interface 2.0.0 differs from Runtime consumer Interface 2.3.0.","No new capability experiment or production incidence measurement."],"next":"/stop","nextOwner":"Parent Runtime integration and publication."}

CHANGELOG.md

@@ -1,5 +1,31 @@
 # Changelog
 
+## 0.195.0
+
+Learning results now use Eval 0.174 and Knowledge 14.
+Code improvement preserves Eval's native proposer result type.
+The [learning audit](docs/research/learning-system-audit-2026-09-05.md) records the findings and the continuing-learning design.
+
+### Exact strategy and code candidates
+
+Strategy checkpoints require `checkpoint.executionRef`, which identifies callbacks, transports, baseline implementations, and external state.
+Runtime also hashes profiles, settings, canonical JSON task payloads, and authored module bytes.
+Changed inputs reject resume before their saved results can be reused.
+Create a new checkpoint when these dependencies change.
+
+`reflectiveGenerator` now requires `createImprovementProposalSource(context)` instead of a pre-bound proposal source.
+Construct the proposer with `repoRoot: context.worktreePath` and use its signal and cost account for drafting.
+Draft failures, stale file contents, and failed patch batches now reject the candidate.
+The complete patch batch applies atomically.
+Tracked diagnosis-only edits no longer become code candidates.
+
+### Persistent lessons and research
+
+Observer responses must match the declared schema.
+Malformed responses and failed lesson writes remain errors in observation and harvest results.
+Corpus records are detached immutable values, and file appends serialize conflicting IDs across processes.
+Supervised knowledge updates execute the supplied profile without appending fixed research instructions.
+
 ## 0.194.0
 
 ### Profiles own recursive authority

api-surface.json

@@ -163,7 +163,7 @@
       "ImproveCandidateValidator": "type 2f7a88cfa0dd",
       "ImproveCodeBaseOptions": "type 8c4ff8a0f7b4",
       "ImproveCodeOptions": "type d3c28c32ba64",
-      "ImproveCodeResult": "type 20846088307b",
+      "ImproveCodeResult": "type e4fa798734e5",
       "ImproveCodeRunOptions": "type 0377eee40f14",
       "ImproveCost": "type 638b3b3b0c34",
       "ImproveCustomCodeGeneratorOptions": "type a15f465fdaa4",
@@ -172,7 +172,7 @@
       "ImproveMethodFactory": "type 50bf77048852",
       "ImproveMethodLineage": "type a2c01575d256",
       "ImproveMethodOptions": "type 6259a221feb5",
-      "ImproveMethodResult": "type e6193eeff9a1",
+      "ImproveMethodResult": "type fa46b216be21",
       "ImproveMethodSource": "type 316a33848162",
       "ImproveOptimizationRunOptions": "type 37126daf2710",
       "ImproveOptions": "type 69633089568b",
@@ -255,7 +255,7 @@
       "RawTraceDistillerOptions": "type a04b8b0d09e1",
       "ReadonlyAgentProfile": "type 416a5d04e179",
       "RecoverExpiredAgentCandidateOptions": "type c9854bda8f64",
-      "ReflectiveGeneratorOptions": "type 827a021107bd",
+      "ReflectiveGeneratorOptions": "type 5d7e3177ca32",
       "ResearchLoopResult": "type 81ea8dc20823",
       "ResearchLoopRunnerOptions": "type dff4c537bf69",
       "ResolvedAgentCandidateContainer": "type 7ff5ca832b4b",
@@ -1244,7 +1244,7 @@
       "EvolutionArchiveNode": "type 9044695f0cc0",
       "EvolutionAuthor": "type ff06ea1bcc70",
       "EvolutionBandInfo": "type f6e42c0bffc6",
-      "EvolutionCandidate": "type c82fd77ca514",
+      "EvolutionCandidate": "type 570d4ff1249f",
       "EvolutionGeneration": "type b6d6f9e8a114",
       "EvolutionReport": "type 7d793b69d7bd",
       "ExecCtx": "type 3ba322d20513",
@@ -1563,7 +1563,7 @@
       "Strategy": "type ab0f3faf7faa",
       "StrategyArtifacts": "type 1a6d725b4744",
       "StrategyCtx": "type 4bb9ec2bb400",
-      "StrategyEvolutionConfig": "type 0ad65ef5d83c",
+      "StrategyEvolutionConfig": "type 03298305c450",
       "StrategyMessage": "type b757f8750838",
       "StrategyResult": "type 10229c661c11",
       "StrategyShotResult": "type 0277d5595f17",
@@ -1932,7 +1932,7 @@
       "startRetainedRun": "value 0751c447e00f",
       "startRetainedRunInEnvironment": "value c50aa480330a",
       "stopSentinel": "value bcbc53b87d5d",
-      "strategyAuthorContract": "value 32c68ce6949c",
+      "strategyAuthorContract": "value 5ee41e231930",
       "strategyAuthorSystemPrompt": "value 373728f5643d",
       "streamAgentTurn": "value d21f23205538",
       "structuralRollout": "value fc852ab503f9",

bench/CHANGELOG.md

@@ -1,5 +1,9 @@
 # Changelog
 
+## 0.8.32
+
+The dependency ranges now require Runtime 0.195.0, Eval 0.174, and Knowledge 14.
+
 ## 0.8.31
 
 Manifest-only: the Sandbox dependency range now admits the published 0.37 cohort.

bench/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tangle-network/agent-bench",
-  "version": "0.8.31",
+  "version": "0.8.32",
   "type": "module",
   "description": "Benchmark adapters and execution for agent-runtime across coding, tool-use, RAG, memory, browser, and terminal tasks.",
   "repository": {

docs/api/index.md

@@ -6234,7 +6234,7 @@ Number of generations explored by Runtime's code path.
 
 ##### raw
 
-> **raw**: `SelfImproveResult`\<`TScenario`, `TArtifact`\>
+> **raw**: `SelfImproveProposerResult`\<`TScenario`, `TArtifact`\>
 
 #### Methods
 
@@ -6626,11 +6626,66 @@ Findings to fall back to when the generation had NO failing cells, so a
 
 ### ReflectiveGeneratorOptions
 
-#### Properties
+#### Methods
+
+##### createImprovementProposalSource()
+
+> **createImprovementProposalSource**(`context`): [`ImprovementProposalSource`](analyst-loop.md#improvementproposalsource)\<[`SurfaceImprovementEdit`](agent.md#surfaceimprovementedit)\>
+
+Bind proposal reads and paid calls to this candidate's worktree and account.
+
+###### Parameters
+
+###### context
+
+###### worktreePath
+
+`string`
+
+The candidate worktree — a clean checkout of the current incumbent.
+
+###### findings
+
+readonly `ProposalFinding`[]
+
+Search or production findings explicitly admitted for proposal use.
+
+###### maxShots
+
+`number`
+
+DEPTH: max iterations the generator may take (agentic uses this; the
+ reflective generator ignores it).
 
-##### improvementProposalSource
+###### signal
+
+`AbortSignal`
+
+###### generation?
+
+`number`
+
+Generation coordinates supplied by Runtime's internal code candidate driver.
+
+###### candidateIndex?
+
+`number`
+
+###### costLedger?
+
+`CostLedgerHandle`
+
+Shared run-wide paid-call account supplied by agent-eval 0.117+.
+
+###### costPhase?
+
+`string`
+
+Receipt attribution phase supplied alongside `costLedger`.
+
+###### Returns
 
-> **improvementProposalSource**: [`ImprovementProposalSource`](analyst-loop.md#improvementproposalsource)\<[`SurfaceImprovementEdit`](agent.md#surfaceimprovementedit)\>
+[`ImprovementProposalSource`](analyst-loop.md#improvementproposalsource)\<[`SurfaceImprovementEdit`](agent.md#surfaceimprovementedit)\>
 
 ***
 

docs/api/primitive-catalog.md

@@ -7,7 +7,7 @@
 
 # Primitive catalog — the never-stale anti-reinvention inventory
 
-> **GENERATED** from `@tangle-network/agent-runtime@0.194.0` and `@tangle-network/agent-eval@0.173.0` by `scripts/gen-primitive-catalog.mjs`. Do NOT hand-edit — run `pnpm run docs:api`. This is the mechanical companion to the JUDGMENT in `canonical-api.md` (§2 decision table + §1.5 AgentProfile law): that doc says WHICH primitive to reach for and what NOT to build; this catalog proves WHAT exists. Per-symbol signatures + `file:line` live in the per-module pages under `docs/api/`.
+> **GENERATED** from `@tangle-network/agent-runtime@0.195.0` and `@tangle-network/agent-eval@0.174.0` by `scripts/gen-primitive-catalog.mjs`. Do NOT hand-edit — run `pnpm run docs:api`. This is the mechanical companion to the JUDGMENT in `canonical-api.md` (§2 decision table + §1.5 AgentProfile law): that doc says WHICH primitive to reach for and what NOT to build; this catalog proves WHAT exists. Per-symbol signatures + `file:line` live in the per-module pages under `docs/api/`.
 
 ## 1. agent-runtime — own public surface
 
@@ -1949,7 +1949,7 @@ Import from `@tangle-network/agent-eval` — 58 exports.
 
 ### CAMPAIGN — profile matrix, gates, improvement loop
 
-Import from `@tangle-network/agent-eval/campaign` — 408 exports.
+Import from `@tangle-network/agent-eval/campaign` — 409 exports.
 
 | Symbol | Kind | Summary |
 |---|---|---|
@@ -2050,6 +2050,7 @@ Import from `@tangle-network/agent-eval/campaign` — 408 exports.
 | `sequentialPairedGate` | function | Anytime-valid sequential paired gate. Conforms to the existing `Gate` |
 | `skillOptOptimizationMethod` | function | Run Microsoft's SkillOpt trainer as a complete optimization method. |
 | `surfaceContentHash` | function | Full SHA-256 content identity for a prompt or finalized code surface. |
+| `surfaceDispatchRef` | function | Bind a campaign cache entry to the exact surface and caller-owned execution revision. |
 | `surfaceHash` | function | Short loop key derived from the same content identity as provenance. |
 | `tangleTracesRoot` | function | The shared, out-of-repo root for campaign/benchmark run bundles. Keeping run |
 | `traceAnalystQualityJudge` | function | _(no summary — add a TSDoc line at the declaration)_ |