fix(improvement): enforce training provenance at held-out boundaries

CHANGELOG.md

@@ -1,5 +1,11 @@
 # Changelog
 
+## 0.244.0
+
+**Migration:** `harness.train(...)` now requires its own `executionRef` for the trainer, serving adapter, and their dependencies. The evaluation identity bound to the harness is no longer substituted into the training receipt. The bound parent profile and validator are unchanged.
+
+Profile optimization and authored/optimized profile promotion refuse final-task content digests already present in either measured profile's declared training receipts, including ancestors and trainer-visible validation data. Baseline exposure is refused before optimizer construction or analyst work; selected-candidate exposure is refused before final measurement. Unrelated task namespaces are not conflated. This checks known exact content digests; it does not certify undeclared or semantically duplicated data as fresh.
+
 ## 0.243.0
 
 Training, optimization, and bound harnesses now share candidate-validator admission and invocation. **Migration:** validators must return `undefined` synchronously or throw. Promises and other return values are rejected instead of silently accepting an unchecked candidate; use a block body for side effects. Composed optimizer leaves obey the same rule, while original callbacks remain bound into execution identity.

api-surface.json

@@ -148,7 +148,7 @@
       "PlannerError": "value 3b3330bf4164",
       "ProfileImprovementHarness": "type cde3e185f5f2",
       "ProfileImprovementHarnessRunOptions": "type 3d1ff1d78855",
-      "ProfileImprovementHarnessTrainOptions": "type f1982d2b20f7",
+      "ProfileImprovementHarnessTrainOptions": "type 3e00f5c6931c",
       "ProfileTrainer": "type 2907ac3022c3",
       "ProfileTrainerRequest": "type 542a10bc9b12",
       "RESEARCH_SUPERVISOR_SYSTEM_PROMPT": "value 373728f5643d",

bench/CHANGELOG.md

@@ -1,5 +1,9 @@
 # Changelog
 
+## 0.13.6
+
+Follow Runtime 0.244.0 through the workspace dependency. Benchmark APIs and grading are unchanged.
+
 ## 0.13.5
 
 Consume Runtime 0.243.0 through the existing workspace dependency. Benchmark APIs and grading behavior are unchanged.

bench/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tangle-network/agent-bench",
-  "version": "0.13.5",
+  "version": "0.13.6",
   "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

@@ -3599,6 +3599,8 @@ Exact execution identity bound at construction.
 
 > **train**(`options`): `Promise`\<[`ImproveTrainingResult`](#improvetrainingresult)\>
 
+Train the bound profile with a separately pinned trainer and serving execution identity.
+
 ###### Parameters
 
 ###### options
@@ -7792,7 +7794,7 @@ Official SkillOpt configuration plus bounded Runtime findings context.
 
 ### ProfileImprovementHarnessTrainOptions
 
-> **ProfileImprovementHarnessTrainOptions** = `Omit`\<[`ImproveTrainingOptions`](#improvetrainingoptions), `"mode"` \| `"executionRef"`\>
+> **ProfileImprovementHarnessTrainOptions** = `Omit`\<[`ImproveTrainingOptions`](#improvetrainingoptions), `"mode"`\>
 
 ***
 

docs/api/primitive-catalog.md

@@ -7,7 +7,7 @@
 
 # Primitive catalog — the never-stale anti-reinvention inventory
 
-> **GENERATED** from `@tangle-network/agent-runtime@0.243.0` and `@tangle-network/agent-eval@0.182.0` 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@0.244.0` and `@tangle-network/agent-eval@0.182.0` 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
 

docs/architecture.md

@@ -92,7 +92,10 @@ Parents and children resume together through Scope, including reconstructed nest
 Provider managers can recover their original invocation and reuse an accepted result.
 Recovery preserves unresolved work when the provider cannot prove its identity or completion.
 The file run lock coordinates one local owner; this does not implement distributed fencing.
-**This recursive execution tree IS the product.** The three things we own
+**This recursive execution tree is the execution substrate, not the success criterion.**
+The product succeeds when retained learning improves future outcomes under the resource and
+verification conditions in sections 0.5 and 9. More nodes or more machinery do not establish that.
+The three things we own
 are small: (1) the **MCP** the agents share (`spawn · observe · steer · stop` +
 `define_check · run_check`); (2) the **profiles** (markdown — the only customization;
 "Drew" is one); (3) the **orchestrator** (`src/runtime/supervise/` — `Scope` + the

docs/canonical-api.md

@@ -4,7 +4,7 @@
 Generated signatures and the complete export list live in docs/api/.
 Run pnpm docs:freshness after editing this file. -->
 
-> **Version 0.243.0.**
+> **Version 0.244.0.**
 > [`docs/api/primitive-catalog.md`](./api/primitive-catalog.md) lists every export and import path.
 > `agent-eval` must satisfy `>=0.182.0 <0.183.0`.
 > `sandbox` must satisfy `>=0.36.4 <0.42.0`.
@@ -285,6 +285,19 @@ Rule of thumb: `delegate` = "I don't care how"; `supervise` = "I authored the dr
 
 ### Checkpoint training is candidate construction, not promotion
 
+The bound harness's `executionRef` identifies evaluation, not the trainer or serving adapter.
+Supply a separate, explicit training `executionRef` to `harness.train(...)`; it is recorded
+unchanged in the training receipt. The harness still binds the parent profile and validator.
+Omitting the training identity is an admission failure, not a fallback to evaluation identity.
+
+Runtime's profile optimization and profile proposal paths reject final measurement when a
+held-out scenario digest occurs in either measured profile's declared training receipts,
+including ancestors and trainer-visible validation rows. Exporters should retain the same
+canonical scenario content digest used by evaluation. This is an exact-digest check, not
+proof against undeclared training, differently encoded duplicates, or semantic contamination.
+Task names in different benchmark namespaces are not treated as interchangeable identities.
+Diagnostic evaluation outside these promotion paths remains the evaluator's responsibility.
+
 Training mode consumes a byte-pinned dataset envelope containing existing Eval export rows
 and the identities of every exposed training and validation task. Payloads are not rewritten.
 The trainer writes one bounded checkpoint; the serving port must independently verify its

docs/research/README.md

@@ -18,6 +18,7 @@ The research files below are source history and focused design inputs, not compe
 
 | Doc | What it holds |
 |-----|---------------|
+| [training-purpose-audit-2026-09-19.md](./training-purpose-audit-2026-09-19.md) | Training provenance defects, integration requirements, existing-owner reuse, and finite engineering/value completion criteria. |
 | [simplification-map-2026-09-07.md](./simplification-map-2026-09-07.md) | Twelve-reader map of the 0.196.0 tree: consumer-less subpaths, dead mechanisms, duplication, and the three files that hold the churn, with an ordered cut list. Raw findings in `.agent/simplification-map-2026-09-07/`. |
 | [learning-system-audit-2026-09-05.md](./learning-system-audit-2026-09-05.md) | Current-source audit of learning across Runtime, Eval, and Knowledge, with reproduced failures and a unification design. |
 | [rsi-atom-masterplan.md](./rsi-atom-masterplan.md) | Historical self-designing-atom plan. Distributed execution work is superseded by `agent-managed-compute/`. |