@openwop/openwop-conformance 2.46.26 → 2.46.27

package/CHANGELOG.md

@@ -1,3 +1,9 @@
+## [2.46.27] — 2026-10-25 — RFC 0244 never-gated and trial-isolation legs
+
+- **New leg `0244.efficiency-never-gated` in `v2-eval-trajectory-trials`**, on the new fixture `conformance-eval-efficiency-ungated`: tasks `lean` (one scripted call) and `busy` (eight) share an output and an expectation, so `busy` must pass with `lean`'s score and the suite must pass at `passScore` 1. `blocked` when the calls never ran or `lean` fails. Judge `judgeNeverGated` in `lib/eval-efficiency-witness.ts`.
+- **New leg `0244.trial-memory-isolated` (RFC 0244 G3)**, on the new fixture `conformance-eval-memory-isolation` (`trials.count: 2`, gated on `agents.evalSuite.trials`), through the conformance-reserved memory tool `conformance.eval:memory` (contract in `fixtures.md`). Each trial must read the `memorySeed` entry, and its first read of `trial-note` must find nothing before its own write. Judge in `lib/eval-memory-witness.ts`; a tool whose writes never read back records `blocked`.
+- **Host impact:** none until a host lists a new fixture in discovery `fixtures`. Listing the memory fixture opts the host in to the memory-tool contract.
+
 ## [2.46.26] — 2026-10-25 — RFC 0244 model-usage and cost legs
 
 - **New leg `0244.efficiency-model-usage` in `v2-eval-trajectory-trials`**, on the existing trajectory and trials runs. Per-task `modelCallCount`, `inputTokens` and `outputTokens`, times the trial count, must sum to the agent's `provider.usage` inside its `agent.invocation.*` brackets (judge and delegate calls excluded); without brackets they must not exceed the run's usage.

package/dist/spec-artifacts.lock.json

@@ -1,5 +1,5 @@
 {
   "package": "@openwop/spec-artifacts",
-  "version": "2.46.26",
-  "stampSha256": "e53058ead4546c46dc23498974219d081a05193fdb5dec16db93a221d023e0d9"
+  "version": "2.46.27",
+  "stampSha256": "c91a318698be6336ff3fc2da76a916ef9f6b4b5fd31988fc2c59863895c88916"
 }

package/fixtures/eval-suites/conformance-eval-efficiency-ungated.json

@@ -0,0 +1,97 @@
+{
+  "suiteId": "core.openwop.evals.conformance-eval-efficiency-ungated",
+  "version": "1.0.0",
+  "targetAgentId": "core.conformance.agents.scripted",
+  "modes": [
+    "golden"
+  ],
+  "thresholds": {
+    "passScore": 1
+  },
+  "tasks": [
+    {
+      "taskId": "lean",
+      "input": {
+        "script": [
+          {
+            "tool": "conformance.eval:lookup",
+            "arguments": {
+              "orderId": "o-1"
+            }
+          }
+        ],
+        "output": "done"
+      },
+      "expected": {
+        "kind": "golden",
+        "match": {
+          "strategy": "contains",
+          "value": "done"
+        }
+      }
+    },
+    {
+      "taskId": "busy",
+      "input": {
+        "script": [
+          {
+            "tool": "conformance.eval:lookup",
+            "arguments": {
+              "orderId": "o-1"
+            }
+          },
+          {
+            "tool": "conformance.eval:lookup",
+            "arguments": {
+              "orderId": "o-2"
+            }
+          },
+          {
+            "tool": "conformance.eval:lookup",
+            "arguments": {
+              "orderId": "o-3"
+            }
+          },
+          {
+            "tool": "conformance.eval:lookup",
+            "arguments": {
+              "orderId": "o-4"
+            }
+          },
+          {
+            "tool": "conformance.eval:lookup",
+            "arguments": {
+              "orderId": "o-5"
+            }
+          },
+          {
+            "tool": "conformance.eval:lookup",
+            "arguments": {
+              "orderId": "o-6"
+            }
+          },
+          {
+            "tool": "conformance.eval:lookup",
+            "arguments": {
+              "orderId": "o-7"
+            }
+          },
+          {
+            "tool": "conformance.eval:lookup",
+            "arguments": {
+              "orderId": "o-8"
+            }
+          }
+        ],
+        "output": "done"
+      },
+      "expected": {
+        "kind": "golden",
+        "match": {
+          "strategy": "contains",
+          "value": "done"
+        }
+      }
+    }
+  ]
+}

package/fixtures/eval-suites/conformance-eval-memory-isolation.json

@@ -0,0 +1,71 @@
+{
+  "suiteId": "core.openwop.evals.conformance-eval-memory-isolation",
+  "version": "1.0.0",
+  "targetAgentId": "core.conformance.agents.scripted",
+  "modes": [
+    "golden"
+  ],
+  "trials": {
+    "count": 2,
+    "k": 1,
+    "passRule": "all"
+  },
+  "tasks": [
+    {
+      "taskId": "write-then-read",
+      "input": {
+        "script": [
+          {
+            "tool": "conformance.eval:memory",
+            "arguments": {
+              "op": "read",
+              "id": "seeded"
+            }
+          },
+          {
+            "tool": "conformance.eval:memory",
+            "arguments": {
+              "op": "read",
+              "id": "trial-note"
+            }
+          },
+          {
+            "tool": "conformance.eval:memory",
+            "arguments": {
+              "op": "write",
+              "id": "trial-note",
+              "content": "written"
+            }
+          },
+          {
+            "tool": "conformance.eval:memory",
+            "arguments": {
+              "op": "read",
+              "id": "trial-note"
+            }
+          }
+        ],
+        "output": "done"
+      },
+      "expected": {
+        "kind": "golden",
+        "match": {
+          "strategy": "contains",
+          "value": "done"
+        }
+      },
+      "fixtures": {
+        "memorySeed": [
+          {
+            "id": "seeded",
+            "content": "seed",
+            "tags": [
+              "conformance"
+            ],
+            "createdAt": "2026-01-01T00:00:00Z"
+          }
+        ]
+      }
+    }
+  ]
+}

package/fixtures.md

@@ -77,6 +77,8 @@ All fixtures MUST advertise:
 | Eval — Fixture Error | `conformance-eval-fixture-error` | RFC 0245 — `onMiss: "fail"`; one task calls the same tool twice, served first by an `isError: true` entry, then by a success (`v2-eval-trajectory-trials`) | eval run `completed`; the first call is recorded as a tool error | ≤ 30s |
 | Eval — Counter (live) | `conformance-eval-counter-live` | RFC 0244 §D, G2 — `onMiss: "live"`; one task makes four `conformance.eval:counter` calls that match no entry, so each runs the live counter (`v2-eval-trajectory-trials`) | eval run `completed`; four live `{key, count}` outcomes | ≤ 30s |
 | Eval — Counter (miss) | `conformance-eval-counter-miss` | RFC 0244 §D, G2 — `onMiss: "fail"`; three `inc S` calls against two exact entries, then an unmatched `inc F` (`v2-eval-trajectory-trials`) | eval run `completed`; `{served: 1}`, `{served: 2}`, then two tool errors | ≤ 30s |
+| Eval — Efficiency Ungated | `conformance-eval-efficiency-ungated` | RFC 0244 §C — two golden tasks with one output, `lean` (1 call) and `busy` (8 calls), at `passScore: 1` (`v2-eval-trajectory-trials`) | eval run `completed`; both tasks and the suite pass | ≤ 30s |
+| Eval — Memory Isolation | `conformance-eval-memory-isolation` | RFC 0244 §B, G3 — `trials.count: 2`; one task seeds `seeded`, then reads it, reads `trial-note`, writes it and reads it back through `conformance.eval:memory` (`v2-eval-trajectory-trials`) | eval run `completed`; each trial reads the seed and finds no `trial-note` before its own write | ≤ 30s |
 | Dispatch Fan-Out Waves | `conformance-dispatch-fanout-waves` | RFC 0243 §A.3 — a parallel `core.dispatch` (`maxConcurrency: 2`, `wait-all`/`collect`) whose supervisor's `mockDispatchPlan` names three `conformance-delay` children. At most `min(2, maxFanOut)` run at once and none is dropped. Gated on `agents.fanOut` admitting `wait-all` with `collect`. | `completed` | ≤ 30s |
 | Dispatch Per-Item Input | `conformance-dispatch-per-item` | RFC 0243 §A.4 — a sequential `core.dispatch` whose decision dispatches `conformance-dispatch-input-mapping-child` twice with `nextWorkerInputs` `item-a`, `item-b`. At `agents.fanOut.perItemInput: true` each child receives its item; otherwise the dispatch node fails `validation_error` and no child is dispatched. | `completed`, or `failed` without `perItemInput` | ≤ 30s |
 | Dispatch Per-Item Length Mismatch | `conformance-dispatch-per-item-mismatch` | `orchestrator-decision.schema.json` `nextWorkerInputs` — one item for two children. Every dispatch host fails the dispatch node `validation_error` and dispatches no child. | `failed` | ≤ 30s |
@@ -1048,7 +1050,7 @@ The `fixtures/eval-suites/` sub-directory holds `AgentEvalSuite` documents (`../
 Every task targets **the scripted agent `core.conformance.agents.scripted`**. Its contract:
 
 - `input.script` is an ordered list of `{ tool, arguments? }`. The agent makes exactly those tool calls, in order, each emitting `agent.tool-called` with `agentId: "core.conformance.agents.scripted"`, `toolName: tool` and `inputs: arguments`.
-- Each call goes through the host's eval tool boundary. In a task with `fixtures.toolResponses`, the RFC 0244 §D selection and `onMiss` rules apply. Otherwise the host answers the call with its own `arguments`. No scripted call reaches a live tool or an external effect, except a miss on the counter tool under `onMiss: "live"` (below).
+- Each call goes through the host's eval tool boundary. In a task with `fixtures.toolResponses`, the RFC 0244 §D selection and `onMiss` rules apply. Otherwise the host answers the call with its own `arguments`. No scripted call reaches a live tool or an external effect, except a miss on the counter tool under `onMiss: "live"` and a call to the memory tool (both below).
 - After the last call the agent decides, and its output is `input.output`.
 
 | Fixture | Facet | What the suite asserts |
@@ -1061,6 +1063,8 @@ Every task targets **the scripted agent `core.conformance.agents.scripted`**. It
 | `conformance-eval-fixture-error` | `agents.evalSuite.fixtureErrors` + `agents.evalSuite.fixtureMatching` | RFC 0245: two `agent.tool-called` and two `agent.tool-returned` for `conformance.eval:lookup`, paired by `callId`; the first return carries `error` with non-empty `code` and `message` and no `outcome` (`status: "error"` under `toolHooks.prePostEvents`), the second carries no `error` (`outcome` is optional; not `status: "error"` under tool hooks); `recovers` reports `fixtureMissCount: 0`. Without `fixtureErrors`: `400 validation_error`. That the agent saw an error and no live tool ran is not observable here (RFC 0245 G1, RFC 0244 G2) |
 | `conformance-eval-counter-live` | `agents.evalSuite.fixtureMatching` | run twice, before and after `conformance-eval-counter-miss`: `fixtureMissCount` > 0; the two live `inc L` counts are consecutive; the second run's first `L` is the first run's last plus one (else `blocked`), and its `F` and `S` reads equal the first run's |
 | `conformance-eval-counter-miss` | `agents.evalSuite.fixtureMatching` | the three `inc S` calls return `{served: 1}`, `{served: 2}`, then a tool error; `inc F` returns a tool error. Neither miss moves the live counter |
+| `conformance-eval-efficiency-ungated` | `agents.evalSuite` | RFC 0244 §C never-gated: with all nine scripted calls in the log and `lean` passing, `busy` passes with `lean`'s score and the suite passes. A host MUST NOT fail either on `efficiency`, reported or not |
+| `conformance-eval-memory-isolation` | `agents.evalSuite.trials` | RFC 0244 §B isolation: both reads of `seeded` return `seed`; of the four `trial-note` reads, two return `written` (each trial's read-back) and two return `null` (each trial's first read). A later trial reading `written` first saw an earlier trial's write. Fewer than two read-backs record `blocked` |
 
 **The counter tool `conformance.eval:counter` is conformance-RESERVED** (the pattern of `core.conformance.side-effect`). A host that lists either counter fixture in discovery `fixtures` MUST, at its eval tool boundary:
 
@@ -1070,6 +1074,14 @@ Every task targets **the scripted agent `core.conformance.agents.scripted`**. It
 
 The counter has no external effect, so serving it is safe on a production tenant. Two concurrent cuts on one tenant make the persistence check fail; the legs then record `blocked`, not a failure.
 
+**The memory tool `conformance.eval:memory` is conformance-RESERVED.** A host that lists `conformance-eval-memory-isolation` in discovery `fixtures` MUST, at its eval tool boundary:
+
+- answer every call to it from the agent's memory for the current trial: the store the task's `fixtures.memorySeed` seeds (RFC 0004 `MemoryAdapter` entries) and the agent's own memory writes land in. It is never served from `toolResponses` and never counted as a fixture miss.
+- answer `{op: "write", id, content}` by storing the entry and returning `{id}`, and `{op: "read", id}` by returning `{id, content}`, with `content: null` when the trial's memory holds no entry `id`;
+- record `outcome` on `agent.tool-returned` for every memory call.
+
+The tool reaches only the trial's eval memory, so it has no external effect. A host that answers it from a store other than the one its trials really use proves nothing about §B isolation, and the contract forbids it.
+
 ## OAuth provider fixtures
 
 RFC 0199 adds `provider.issuer`, `provider.pkce` and the two issuer-less providers (`issuerlessProviders[]`) the distinct-redirect-URI leg needs; the endpoints a host actually talks to are supplied at run time by the `oauth/authorize-start` seam, which points them at the suite's authorization-server double (`src/lib/oauth-as-double.ts`).

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@openwop/openwop-conformance",
-  "version": "2.46.26",
+  "version": "2.46.27",
   "description": "Black-box conformance suite for OpenWOP hosts (v2 current; v1 scenarios retained for the frozen v1 tree).",
   "repository": {
     "type": "git",
@@ -56,6 +56,6 @@
     "@openwop/spec-artifacts": "file:../spec-artifacts"
   },
   "peerDependencies": {
-    "@openwop/spec-artifacts": "2.46.26"
+    "@openwop/spec-artifacts": "2.46.27"
   }
 }

package/requirements.json

@@ -3,10 +3,10 @@
   "generatedFrom": "src/scenarios/*.test.ts",
   "counts": {
     "files": 749,
-    "tests": 2846,
-    "withStableId": 2846,
+    "tests": 2848,
+    "withStableId": 2848,
     "interpolatedTitles": 0,
-    "explicitIds": 2719
+    "explicitIds": 2721
   },
   "records": [
     {
@@ -31092,7 +31092,7 @@
     {
       "id": "openwop.it.v2-eval-trajectory-trials.the-rfc-0244-suite-summary-and-event-fields-validate-and-the-negatives-are-refus",
       "file": "v2-eval-trajectory-trials.test.ts",
-      "line": 196,
+      "line": 211,
       "title": "the RFC 0244 suite, summary and event fields validate and the negatives are refused",
       "explicitId": "openwop.requirement.0244.wire-shape",
       "citations": [
@@ -31131,7 +31131,7 @@
     {
       "id": "openwop.it.v2-eval-trajectory-trials.the-rfc-0244-additions-are-content-free-eval-summary-no-content-leak",
       "file": "v2-eval-trajectory-trials.test.ts",
-      "line": 219,
+      "line": 234,
       "title": "the RFC 0244 additions are content-free (eval-summary-no-content-leak)",
       "explicitId": "openwop.requirement.0244.content-free",
       "citations": [
@@ -31148,7 +31148,7 @@
     {
       "id": "openwop.it.v2-eval-trajectory-trials.the-eval-suite-fixtures-validate-against-the-v2-suite-schema",
       "file": "v2-eval-trajectory-trials.test.ts",
-      "line": 226,
+      "line": 241,
       "title": "the eval-suite fixtures validate against the v2 suite schema",
       "explicitId": "openwop.requirement.0244.fixtures-valid",
       "citations": [
@@ -31162,7 +31162,7 @@
     {
       "id": "openwop.it.v2-eval-trajectory-trials.the-rfc-0246-gating-fields-validate-and-stay-in-0-1",
       "file": "v2-eval-trajectory-trials.test.ts",
-      "line": 233,
+      "line": 248,
       "title": "the RFC 0246 gating fields validate and stay in 0–1",
       "explicitId": "openwop.requirement.0246.gated-fields-shape",
       "citations": [
@@ -31184,7 +31184,7 @@
     {
       "id": "openwop.it.v2-eval-trajectory-trials.a-fixture-entry-may-be-marked-iserror-and-only-as-a-boolean-rfc-0245-a",
       "file": "v2-eval-trajectory-trials.test.ts",
-      "line": 241,
+      "line": 256,
       "title": "a fixture entry may be marked isError, and only as a boolean (RFC 0245 §A)",
       "explicitId": "openwop.requirement.0245.fixture-error-shape",
       "citations": [
@@ -31206,7 +31206,7 @@
     {
       "id": "openwop.it.v2-eval-trajectory-trials.a-trajectory-suite-is-graded-per-the-reference-grader",
       "file": "v2-eval-trajectory-trials.test.ts",
-      "line": 251,
+      "line": 266,
       "title": "a trajectory suite is graded per the reference grader",
       "explicitId": "openwop.requirement.0244.trajectory-graded",
       "citations": [
@@ -31240,7 +31240,7 @@
     {
       "id": "openwop.it.v2-eval-trajectory-trials.an-eval-run-brackets-its-scores-with-one-eval-started-and-one-eval-completed-rfc",
       "file": "v2-eval-trajectory-trials.test.ts",
-      "line": 285,
+      "line": 300,
       "title": "an eval run brackets its scores with one eval.started and one eval.completed (RFC 0081 §C)",
       "explicitId": "openwop.requirement.0081.eval-envelope",
       "citations": [
@@ -31254,7 +31254,7 @@
     {
       "id": "openwop.it.v2-eval-trajectory-trials.the-served-evalsummary-validates-and-its-counts-match-its-rows-rfc-0081-c",
       "file": "v2-eval-trajectory-trials.test.ts",
-      "line": 293,
+      "line": 308,
       "title": "the served EvalSummary validates and its counts match its rows (RFC 0081 §C)",
       "explicitId": "openwop.requirement.0081.summary-served",
       "citations": [
@@ -31268,7 +31268,7 @@
     {
       "id": "openwop.it.v2-eval-trajectory-trials.repeated-trials-emit-one-eval-scored-per-task-per-trial",
       "file": "v2-eval-trajectory-trials.test.ts",
-      "line": 314,
+      "line": 329,
       "title": "repeated trials emit one eval.scored per task per trial",
       "explicitId": "openwop.requirement.0244.trials-per-task",
       "citations": [
@@ -31286,7 +31286,7 @@
     {
       "id": "openwop.it.v2-eval-trajectory-trials.the-per-task-trials-block-and-trialstats-follow-the-b-formulas",
       "file": "v2-eval-trajectory-trials.test.ts",
-      "line": 325,
+      "line": 340,
       "title": "the per-task trials block and trialStats follow the §B formulas",
       "explicitId": "openwop.requirement.0244.trial-stats-formula",
       "citations": [
@@ -31325,7 +31325,7 @@
     {
       "id": "openwop.it.v2-eval-trajectory-trials.efficiency-toolcallcount-counts-the-scripted-agent-tool-called-events-c",
       "file": "v2-eval-trajectory-trials.test.ts",
-      "line": 348,
+      "line": 363,
       "title": "efficiency.toolCallCount counts the scripted agent.tool-called events (§C)",
       "explicitId": "openwop.requirement.0244.efficiency-tool-calls",
       "citations": [
@@ -31339,7 +31339,7 @@
     {
       "id": "openwop.it.v2-eval-trajectory-trials.efficiency-model-and-token-counts-match-the-agent-s-provider-usage-c",
       "file": "v2-eval-trajectory-trials.test.ts",
-      "line": 360,
+      "line": 375,
       "title": "efficiency model and token counts match the agent's provider.usage (§C)",
       "explicitId": "openwop.requirement.0244.efficiency-model-usage",
       "citations": [
@@ -31353,7 +31353,7 @@
     {
       "id": "openwop.it.v2-eval-trajectory-trials.totalcostusd-counts-every-trial-b",
       "file": "v2-eval-trajectory-trials.test.ts",
-      "line": 371,
+      "line": 386,
       "title": "totalCostUsd counts every trial (§B)",
       "explicitId": "openwop.requirement.0244.cost-every-trial",
       "citations": [
@@ -31364,10 +31364,38 @@
         }
       ]
     },
+    {
+      "id": "openwop.it.v2-eval-trajectory-trials.a-host-does-not-fail-a-task-or-a-suite-on-efficiency-c",
+      "file": "v2-eval-trajectory-trials.test.ts",
+      "line": 394,
+      "title": "a host does not fail a task or a suite on efficiency (§C)",
+      "explicitId": "openwop.requirement.0244.efficiency-never-gated",
+      "citations": [
+        {
+          "section": "eval-summary.schema.json efficiency (RFC 0244 §C)",
+          "requirement": null,
+          "interpolated": true
+        }
+      ]
+    },
+    {
+      "id": "openwop.it.v2-eval-trajectory-trials.memory-one-trial-writes-is-not-visible-to-another-trial-b",
+      "file": "v2-eval-trajectory-trials.test.ts",
+      "line": 409,
+      "title": "memory one trial writes is not visible to another trial (§B)",
+      "explicitId": "openwop.requirement.0244.trial-memory-isolated",
+      "citations": [
+        {
+          "section": "agent-eval-suite.schema.json trials, fixtures.md memory contract (RFC 0244 §B)",
+          "requirement": null,
+          "interpolated": true
+        }
+      ]
+    },
     {
       "id": "openwop.it.v2-eval-trajectory-trials.a-miss-under-onmiss-live-reaches-the-live-counter-and-is-counted-d",
       "file": "v2-eval-trajectory-trials.test.ts",
-      "line": 401,
+      "line": 447,
       "title": "a miss under onMiss \"live\" reaches the live counter and is counted (§D)",
       "explicitId": "openwop.requirement.0244.live-miss-counted",
       "citations": [
@@ -31381,7 +31409,7 @@
     {
       "id": "openwop.it.v2-eval-trajectory-trials.repeated-calls-are-served-the-first-unconsumed-fixture-entry-once-each-d",
       "file": "v2-eval-trajectory-trials.test.ts",
-      "line": 409,
+      "line": 455,
       "title": "repeated calls are served the first unconsumed fixture entry, once each (§D)",
       "explicitId": "openwop.requirement.0244.first-unconsumed",
       "citations": [
@@ -31395,7 +31423,7 @@
     {
       "id": "openwop.it.v2-eval-trajectory-trials.a-miss-under-onmiss-fail-never-reaches-the-live-counter-d",
       "file": "v2-eval-trajectory-trials.test.ts",
-      "line": 417,
+      "line": 463,
       "title": "a miss under onMiss \"fail\" never reaches the live counter (§D)",
       "explicitId": "openwop.requirement.0244.miss-no-live-call",
       "citations": [
@@ -31409,7 +31437,7 @@
     {
       "id": "openwop.it.v2-eval-trajectory-trials.a-fixture-miss-under-onmiss-fail-fails-the-trial-and-is-counted",
       "file": "v2-eval-trajectory-trials.test.ts",
-      "line": 425,
+      "line": 471,
       "title": "a fixture miss under onMiss \"fail\" fails the trial and is counted",
       "explicitId": "openwop.requirement.0244.fixture-miss-fails",
       "citations": [
@@ -31446,7 +31474,7 @@
     {
       "id": "openwop.it.v2-eval-trajectory-trials.a-suite-using-an-unadvertised-rfc-0244-feature-is-refused-at-createrun",
       "file": "v2-eval-trajectory-trials.test.ts",
-      "line": 445,
+      "line": 491,
       "title": "a suite using an unadvertised RFC 0244 feature is refused at createRun",
       "explicitId": "openwop.requirement.0244.unadvertised-feature-refused",
       "citations": [
@@ -31464,7 +31492,7 @@
     {
       "id": "openwop.it.v2-eval-trajectory-trials.under-passrule-mean-a-task-whose-trials-failed-their-trajectory-does-not-pass-rf",
       "file": "v2-eval-trajectory-trials.test.ts",
-      "line": 461,
+      "line": 507,
       "title": "under passRule \"mean\" a task whose trials failed their trajectory does not pass (RFC 0246)",
       "explicitId": "openwop.requirement.0246.mean-gated",
       "citations": [
@@ -31478,7 +31506,7 @@
     {
       "id": "openwop.it.v2-eval-trajectory-trials.a-suite-whose-tasks-failed-their-trajectory-does-not-pass-on-the-output-scores-r",
       "file": "v2-eval-trajectory-trials.test.ts",
-      "line": 474,
+      "line": 520,
       "title": "a suite whose tasks failed their trajectory does not pass on the output scores (RFC 0246)",
       "explicitId": "openwop.requirement.0246.suite-gated",
       "citations": [
@@ -31492,7 +31520,7 @@
     {
       "id": "openwop.it.v2-eval-trajectory-trials.a-fixture-entry-marked-iserror-is-replayed-as-a-tool-error-not-a-success-rfc-024",
       "file": "v2-eval-trajectory-trials.test.ts",
-      "line": 485,
+      "line": 531,
       "title": "a fixture entry marked isError is replayed as a tool error, not a success (RFC 0245)",
       "explicitId": "openwop.requirement.0245.fixture-error-replayed",
       "citations": [
@@ -31506,7 +31534,7 @@
     {
       "id": "openwop.it.v2-eval-trajectory-trials.a-suite-using-iserror-is-refused-at-createrun-without-fixtureerrors-rfc-0245-c",
       "file": "v2-eval-trajectory-trials.test.ts",
-      "line": 499,
+      "line": 545,
       "title": "a suite using isError is refused at createRun without fixtureErrors (RFC 0245 §C)",
       "explicitId": "openwop.requirement.0245.unadvertised-refused",
       "citations": [

package/schemas/CORPUS-STAMP.json

@@ -1,17 +1,17 @@
 {
   "_comment": "Provenance of @openwop/spec-artifacts (RFC 0168 §D.2). files: SHA-256 per file; the conformance suite compares the installed peer against dist/spec-artifacts.lock.json at start.",
   "package": "@openwop/spec-artifacts",
-  "version": "2.46.26",
-  "corpusTag": "v2.46.26",
+  "version": "2.46.27",
+  "corpusTag": "v2.46.27",
   "files": {
     "api/.redocly.lint-ignore.yaml": "bf5a8350b88a72fa43f59605ed8d903ed24b6cfccda5e45509c9f6ed9ee4e712",
     "api/asyncapi.yaml": "d5ecb9ee6114582be3b1f662c84bfac9ae96dae7bacb853e461168f70a8e1c7d",
     "api/grpc/openwop.proto": "c3e72bb17cba514ee98feb6434e6c9b6ea6795bfd086489ec69fd882dd1ad977",
     "api/openapi.yaml": "ac6de5303872664b78ff419c035915593e01599e02835e777a0123de1baeb398",
     "api/redocly.yaml": "b0604c89b2ca6d5076ec25725c539dad44a741a811fe524439ee6daef8baa09f",
-    "api/seams-v2.yaml": "fc5aaff321913531e04b13522de92a9a8fa3f61be836696e9ab83fbc563caa87",
-    "api/v2/asyncapi.yaml": "0ea85f3cb0af276cd7ed82e14810c1093917f0d1daa80bc0226b78e127c4e17d",
-    "api/v2/openapi.yaml": "7c7cd37cc0cabcd37285502594977f6ddc5a59f30938d7aede36f69e05322244",
+    "api/seams-v2.yaml": "fe9b3e83272ce29a2d0d13a09e12907cbf0e38e7de36d48bbb9343b6d88aa649",
+    "api/v2/asyncapi.yaml": "b647ccc7a8b7b6d62385d754bc597ef95725cf123655a9622510518ee2e64516",
+    "api/v2/openapi.yaml": "bfc7c189810490162495eb7e0f744b53274f6b72236004996cec951bb08af494",
     "api/v2/redocly.yaml": "1e66b60e6118ad11a823bb620678be464d99dfe50a40e3e6f93ec9429b88b34c",
     "schemas/README.md": "276fc1d51eb08e14e9127b9fd27761d7046ee33c12000f3a64a801322db8bf58",
     "schemas/a2a-task-state.schema.json": "75d5049dea8bd873ff0e7546f1c60c8d36c219bec7084264360be54a8be30ae1",
@@ -212,14 +212,14 @@
     "schemas/workspace-file.schema.json": "464de85c2a068243084ee9c1d969bc7cd5d8f7948574e58450d6493c38a0e1e4",
     "spec/v1/alias-detectors.json": "22444c22bb88c568339ef512553f9fc38146e22b566d11ab8608f655d688d19c",
     "spec/v1/capability-declaration-classes.json": "e7729aed5c4b4e1dd02abab0530f14cc95f5d4070fe51fb139e7f5cccefa00c6",
-    "spec/v1/core-standard-manifest.json": "580b853bb37383d5acea488fba6747193d1db4e6e63025255433dababf008d60",
+    "spec/v1/core-standard-manifest.json": "0511812bdd86bbbe79f7a1dddd60878b01c66b1d4a43122822f74f5cd6a12bf9",
     "spec/v1/deprecations.json": "e723bc11d7272c71789ad1303bb7f21809a8a024abf57594f9d5c35dc199ece9",
     "spec/v1/deprecations.schema.json": "3e393c405d2a41b467d8c5e3c468549078df1ce6a6d2588fc488097b95d9b55b",
     "spec/v1/end-of-support.json": "35f0bf3b70fe78b16e2a2fd81a153fc27e703f6c16376c6fe7cdf1f4f2c4e6d4",
     "spec/v1/event-codemap.json": "3da60d884157793a360da532a9fcbbfb5285636db325a74cec94b34622186d97",
     "spec/v1/event-codemap.schema.json": "d05933b2e88103aff51a2f774df97b9bda065e0fc88dbd9b114cc47a32189174",
     "spec/v1/extensions.json": "79a60754aa16cbdbdb604f8e5c00038af535a1690a4d07bd5ff7c38f0256a7bc",
-    "spec/v1/gaps.json": "21d53eee36f47b3cd44ba54802acc5cdc810d242d1fd929b1e76167da8566957",
+    "spec/v1/gaps.json": "c590927affc3d909b5096faf341995100e17717d9e302bc9f00dd1729353de45",
     "spec/v1/gaps.schema.json": "8fd83259f556553c9df0f53e7a82ca8c2a4771d471197f69b4e2ae0ceeacfd66",
     "spec/v1/migrations.json": "4c50760be70cb644d02427bcaf5749606700edf4513f423903cf3b1d12f5c300",
     "spec/v1/migrations.schema.json": "886779aa6c22e646db097f5df210adb018a4dd14a7b815465a18c8a7056c8f72",
@@ -318,9 +318,9 @@
     "spec/v2/path-manifest.json": "20f598f56df460ad3ed50aba3d6bc2933ea16a3c96ce125da1c5dd6989d4155c",
     "spec/v2/peer-dependency-aliases.json": "d10299280abee08258502925bc327293ee413e0108cd6e6ec75ff6110653308d",
     "spec/v2/profiles.json": "b382b261c7844798575f7c1b37c934f0da19846016a2337d883d4f5ef6e05479",
-    "spec/v2/release.json": "e2a9a7d13015eb3ca85bdbcd6ec422b6e91780fb9019b0e75e2855b826728b1d",
+    "spec/v2/release.json": "6452f97c5ee21eddfce6f69085a8c4828d39d8337f0a6b4830025ecb3c6abc98",
     "spec/v2/retention-floors.json": "eaf3722d95c79947af1d4269ef85117e126518c588cfcf1a2b21b97269f51624",
-    "spec/v2/surface-baseline.json": "4a1e5ab2349349eeeb9fa173e392627d49c2bee77ac8a42897aeacf2c3a4d784"
+    "spec/v2/surface-baseline.json": "aaef2430679e649dc80e1eb961a0e6900cea84ce2d8999a45ef58ca9911fcb3b"
   },
-  "corpusCommit": "e5e8306a4f9c8c41394cfbe63fb58a08a74fe5b9"
+  "corpusCommit": "7d8354c0ba855f4d721cdaa6ae34ace611650353"
 }

package/src/lib/eval-efficiency-witness.ts

@@ -12,7 +12,8 @@
  * agent calls a model, which the fixture contract leaves to the host.
  *
  * `efficiency` is optional on a summary entry, so a run whose entries carry
- * none is not evidence either way (`applicable: false`).
+ * none is not evidence either way (`applicable: false`). The never-gated rule
+ * (`judgeNeverGated`) holds whether or not a host reports `efficiency`.
  *
  * @see RFCS/0244-eval-trajectories-trials-and-recorded-fixtures.md §C
  */
@@ -137,3 +138,30 @@ export function judgeCostEveryTrial(summary: Rec, log: { bracketed: boolean; usa
   const sum = (costs as number[]).reduce((a, b) => a + b, 0);
   return { applicable: true, findings: [{ ok: total >= sum || close(total, sum), message: `totalCostUsd MUST count every trial: it MUST be at least the ${sum} costEstimateUsd of the agent's ${log.usage.length} provider.usage events across all trials (got ${total})` }] };
 }
+
+/**
+ * §C "MUST NOT fail a task or a suite on" efficiency, on
+ * `conformance-eval-efficiency-ungated`: `lean` and `busy` share an output and
+ * an expectation and differ only in their scripted call counts, so efficiency
+ * is the only thing that can tell them apart. `busy` MUST pass and score what
+ * `lean` scores, and the suite (`passScore` 1) MUST pass. The verdict is
+ * `blocked` unless the log carries every scripted call (otherwise the
+ * difference was never produced) and `lean` passes (otherwise a failing `busy`
+ * could have another cause).
+ */
+export function judgeNeverGated(summary: Rec, lean: ScriptedTask, busy: ScriptedTask, toolCallsInLog: number): { blocked: string } | EfficiencyVerdict {
+  const want = lean.scriptLength + busy.scriptLength;
+  if (toolCallsInLog !== want) return { blocked: `the log carries ${toolCallsInLog} agent.tool-called events, not the ${want} scripted calls, so the two tasks' efficiency never differed` };
+  const entries = (Array.isArray(summary['tasks']) ? (summary['tasks'] as unknown[]) : []).filter(isRec);
+  const l = entries.find((e) => e['taskId'] === lean.taskId);
+  const b = entries.find((e) => e['taskId'] === busy.taskId);
+  if (l?.['passed'] !== true) return { blocked: `the control task ${lean.taskId} did not pass, so a failing ${busy.taskId} could have a cause other than efficiency` };
+  return {
+    applicable: true,
+    findings: [
+      { ok: b?.['passed'] === true, message: `task ${busy.taskId} (${busy.scriptLength} tool calls) MUST pass like ${lean.taskId} (${lean.scriptLength}): a host MUST NOT fail a task on efficiency (got passed ${String(b?.['passed'])})` },
+      { ok: typeof b?.['score'] === 'number' && b['score'] === l['score'], message: `task ${busy.taskId} MUST score what ${lean.taskId} scores (${String(l['score'])}): efficiency is informational (got ${String(b?.['score'])})` },
+      { ok: summary['passed'] === true, message: `the suite MUST pass at passScore 1 when both tasks pass: a host MUST NOT fail a suite on efficiency (got passed ${String(summary['passed'])})` },
+    ],
+  };
+}

package/src/lib/eval-memory-witness.ts

@@ -0,0 +1,82 @@
+/**
+ * RFC 0244 §B, G3 — trial isolation: each trial starts from the same
+ * `memorySeed` snapshot, and memory one trial writes is not visible to another
+ * trial of the same eval run. The suite-owned memory tool
+ * `conformance.eval:memory` makes both observable from outside. A pure judge
+ * over one eval run of `conformance-eval-memory-isolation`, for
+ * `v2-eval-trajectory-trials.test.ts`.
+ *
+ * The tool reads and writes the trial's own agent memory, the store
+ * `memorySeed` seeds (contract in `fixtures.md` §Eval-suite fixtures):
+ * `write {id, content}` stores an entry, `read {id}` returns `{id, content}`,
+ * with `content: null` for an entry the trial's memory does not hold. Every
+ * trial runs the same script:
+ *   1. read `seeded`      — the seed, in every trial;
+ *   2. read `trial-note`  — nothing yet, unless another trial's write leaked;
+ *   3. write `trial-note` = `written`;
+ *   4. read `trial-note`  — `written`, the positive control.
+ *
+ * Trials may run concurrently (scheduling is host-internal), so the judge
+ * never groups calls by trial. Over n trials a conforming host returns n
+ * `null` and n `written` reads of `trial-note`. A leak turns a step-2 `null`
+ * into `written`, so the run reads fewer than n `null`. A host whose writes
+ * never read back (fewer than n `written`) shows nothing about isolation, and
+ * the verdict is `blocked`, never a pass.
+ *
+ * @see RFCS/0244-eval-trajectories-trials-and-recorded-fixtures.md §B
+ */
+
+export const MEMORY_TOOL = 'conformance.eval:memory';
+export const SEEDED = { id: 'seeded', content: 'seed' } as const;
+export const NOTE = { id: 'trial-note', content: 'written' } as const;
+
+export interface MemoryFinding { readonly ok: boolean; readonly message: string }
+export type MemoryVerdict =
+  | { readonly kind: 'blocked'; readonly reason: string }
+  | { readonly kind: 'judged'; readonly findings: readonly MemoryFinding[] };
+
+type Rec = Record<string, unknown>;
+export interface MemoryEvent { readonly type: string; readonly payload: Rec }
+export interface MemoryCall { readonly op: unknown; readonly id: unknown; readonly outcome: unknown; readonly returned: boolean }
+
+const isRec = (v: unknown): v is Rec => typeof v === 'object' && v !== null && !Array.isArray(v);
+
+/** The memory-tool calls of one run, in log order, each paired with its return by `callId`. */
+export function memoryCalls(events: readonly MemoryEvent[], agentId: string): MemoryCall[] {
+  const returns = new Map<string, Rec>();
+  for (const e of events) {
+    if (e.type === 'agent.tool-returned' && e.payload['toolName'] === MEMORY_TOOL && typeof e.payload['callId'] === 'string') returns.set(e.payload['callId'], e.payload);
+  }
+  const out: MemoryCall[] = [];
+  for (const e of events) {
+    if (e.type !== 'agent.tool-called' || e.payload['toolName'] !== MEMORY_TOOL || e.payload['agentId'] !== agentId) continue;
+    const inputs = isRec(e.payload['inputs']) ? e.payload['inputs'] : {};
+    const r = typeof e.payload['callId'] === 'string' ? returns.get(e.payload['callId']) : undefined;
+    out.push({ op: inputs['op'], id: inputs['id'], outcome: r?.['outcome'], returned: r !== undefined });
+  }
+  return out;
+}
+
+const content = (c: MemoryCall): unknown => (isRec(c.outcome) ? c.outcome['content'] : undefined);
+
+/** §B isolation over `trials` trials of the four-step script. */
+export function judgeTrialMemoryIsolated(calls: readonly MemoryCall[], trials: number): MemoryVerdict {
+  if (calls.length !== 4 * trials || !calls.every((c) => c.returned)) {
+    return { kind: 'blocked', reason: `the run MUST log the ${4 * trials} scripted memory calls (four per trial), each with an agent.tool-returned (got ${calls.length}, ${calls.filter((c) => c.returned).length} returned)` };
+  }
+  const reads = calls.filter((c) => c.op === 'read');
+  const seeded = reads.filter((c) => c.id === SEEDED.id);
+  const note = reads.filter((c) => c.id === NOTE.id);
+  const written = note.filter((c) => content(c) === NOTE.content).length;
+  if (written < trials) {
+    return { kind: 'blocked', reason: `only ${written} of ${trials} trials read back their own write of ${NOTE.id}, so the memory tool does not reach the trial's memory (fixtures.md memory contract) and isolation is not observable` };
+  }
+  const empty = note.filter((c) => content(c) === null).length;
+  return {
+    kind: 'judged',
+    findings: [
+      { ok: seeded.length === trials && seeded.every((c) => content(c) === SEEDED.content), message: `every trial MUST start from the memorySeed snapshot: each of the ${trials} reads of ${SEEDED.id} MUST return content "${SEEDED.content}" (got ${JSON.stringify(seeded.map(content))})` },
+      { ok: empty === trials && note.length === 2 * trials, message: `memory one trial writes MUST NOT be visible to another trial: each trial's first read of ${NOTE.id} MUST find nothing, so ${trials} of the ${note.length} reads return content null (got ${empty})` },
+    ],
+  };
+}

package/src/scenarios/v2-eval-trajectory-trials.test.ts

@@ -36,8 +36,20 @@
  *      length (the mean over identical trials), and on the one-trial run the
  *      counts sum to the log's `agent.tool-called` events. `efficiency` is
  *      optional, so a host that reports none records `inapplicable`. Judge in
- *      `lib/eval-efficiency-witness.ts`. The never-gated rule is not judged:
- *      no fixture makes efficiency the only difference between two verdicts.
+ *      `lib/eval-efficiency-witness.ts`.
+ *   6d. §C never-gated (2.46.27), `conformance-eval-efficiency-ungated`: two
+ *      tasks with one output and one expectation, `lean` (1 scripted call) and
+ *      `busy` (8). `busy` passes and scores what `lean` scores, and the suite
+ *      passes at `passScore` 1 (`0244.efficiency-never-gated`). `blocked` when
+ *      the calls never ran or `lean` itself fails. Gated on `agents.evalSuite`
+ *      and the fixture only: the rule binds whether or not a host reports
+ *      `efficiency`.
+ *   6e. §B trial isolation (2.46.27, G3), `conformance-eval-memory-isolation`,
+ *      gated on `trials`: two trials of read seed, read note, write note, read
+ *      note through the reserved memory tool `conformance.eval:memory`. Every
+ *      trial reads the seed, and each trial's first note read finds nothing
+ *      (`0244.trial-memory-isolated`); a tool whose writes never read back is
+ *      `blocked`. Judge in `lib/eval-memory-witness.ts`.
  *   6c. §C model usage and §B cost (2.46.26), on the same runs. The per-task
  *      `modelCallCount`, `inputTokens` and `outputTokens`, times the trial
  *      count, sum to the agent's `provider.usage` inside its
@@ -109,7 +121,8 @@ import { judgeEvalEnvelope, judgeSummaryServed } from '../lib/eval-envelope-witn
 import { gradeTrajectory, passAtK, passHatK, scoreStderr, withinTolerance, type TrajectorySpec } from '../lib/eval-trajectory.js';
 import { judgeMeanGated, judgeSuiteGated } from '../lib/eval-gate-witness.js';
 import { judgeFixtureErrorReplayed } from '../lib/eval-fixture-error-witness.js';
-import { agentUsage, judgeCostEveryTrial, judgeModelUsage, judgeToolCallCount } from '../lib/eval-efficiency-witness.js';
+import { agentUsage, judgeCostEveryTrial, judgeModelUsage, judgeNeverGated, judgeToolCallCount } from '../lib/eval-efficiency-witness.js';
+import { judgeTrialMemoryIsolated, memoryCalls } from '../lib/eval-memory-witness.js';
 import { counterCalls, judgeFirstUnconsumed, judgeLiveMiss, judgeMissNoLiveCall, type CounterCall } from '../lib/eval-counter-witness.js';
 
 const PROFILE = majorProfile(2);
@@ -128,6 +141,8 @@ const ID_REFUSED = 'openwop.requirement.0244.unadvertised-feature-refused';
 const ID_EFFICIENCY = 'openwop.requirement.0244.efficiency-tool-calls';
 const ID_MODEL_USAGE = 'openwop.requirement.0244.efficiency-model-usage';
 const ID_COST_TRIALS = 'openwop.requirement.0244.cost-every-trial';
+const ID_NEVER_GATED = 'openwop.requirement.0244.efficiency-never-gated';
+const ID_MEMORY_ISOLATED = 'openwop.requirement.0244.trial-memory-isolated';
 const ID_LIVE_MISS = 'openwop.requirement.0244.live-miss-counted';
 const ID_FIRST_UNCONSUMED = 'openwop.requirement.0244.first-unconsumed';
 const ID_NO_LIVE_CALL = 'openwop.requirement.0244.miss-no-live-call';
@@ -224,7 +239,7 @@ describe('v2 RFC 0244 — wire shape (server-free)', () => {
   });
 
   it('the eval-suite fixtures validate against the v2 suite schema', () => {
-    for (const id of ['conformance-eval-trajectory', 'conformance-eval-trials', 'conformance-eval-fixture-miss', 'conformance-eval-trajectory-gate', 'conformance-eval-trajectory-gate-mean', 'conformance-eval-fixture-error', 'conformance-eval-counter-live', 'conformance-eval-counter-miss']) {
+    for (const id of ['conformance-eval-trajectory', 'conformance-eval-trials', 'conformance-eval-fixture-miss', 'conformance-eval-trajectory-gate', 'conformance-eval-trajectory-gate-mean', 'conformance-eval-fixture-error', 'conformance-eval-counter-live', 'conformance-eval-counter-miss', 'conformance-eval-efficiency-ungated', 'conformance-eval-memory-isolation']) {
       const r = suiteSchema(fixture(id));
       expect(r.ok, req(ID_FIXTURES, 'fixtures.md §eval-suites', `${id} MUST validate: ${r.errors}`)).toBe(true);
     }
@@ -376,6 +391,37 @@ describe('v2 RFC 0244 — host behaviour', () => {
     for (const x of v.findings) expect(x.ok, req(ID_COST_TRIALS, 'agent-eval-suite.schema.json thresholds.maxCostUsd, eval-summary.schema.json totalCostUsd (RFC 0244 §B)', x.message)).toBe(true);
   });
 
+  it('a host does not fail a task or a suite on efficiency (§C)', async () => {
+    const { doc, facet } = await evalSuiteFacet();
+    if (!doc) return softSkip('blocked', UNREADABLE);
+    if (facet === null) return softSkip('inapplicable', 'the host does not advertise agents.evalSuite');
+    const id = 'conformance-eval-efficiency-ungated';
+    if (!isFixtureAdvertised(id)) return softSkip('inapplicable', `the host does not advertise fixture ${id}`);
+    const o = await runSuite(id);
+    if (o.kind === 'skip') return softSkip(o.disposition, o.reason);
+    const task = (taskId: string) => ({ taskId, scriptLength: fixture(id).tasks.find((t) => t.taskId === taskId)!.input.script.length });
+    const logCalls = o.events.filter((e) => e.type === 'agent.tool-called' && e.payload['agentId'] === AGENT).length;
+    const v = judgeNeverGated(o.summary, task('lean'), task('busy'), logCalls);
+    if ('blocked' in v) return softSkip('blocked', v.blocked);
+    for (const x of v.findings) expect(x.ok, req(ID_NEVER_GATED, 'eval-summary.schema.json efficiency (RFC 0244 §C)', x.message)).toBe(true);
+  });
+
+  it('memory one trial writes is not visible to another trial (§B)', async () => {
+    const { doc, facet } = await evalSuiteFacet();
+    if (!doc) return softSkip('blocked', UNREADABLE);
+    const trials = facet?.['trials'];
+    if (!isRecord(trials)) return softSkip('inapplicable', 'the host does not advertise agents.evalSuite.trials');
+    const id = 'conformance-eval-memory-isolation';
+    const n = fixture(id).trials!.count;
+    if (typeof trials['maxTrials'] === 'number' && trials['maxTrials'] < n) return softSkip('inapplicable', `the host advertises trials.maxTrials ${trials['maxTrials']}, below the fixture's ${n}`);
+    if (!isFixtureAdvertised(id)) return softSkip('inapplicable', `the host does not advertise fixture ${id}, its opt-in to the memory-tool contract`);
+    const o = await runSuite(id);
+    if (o.kind === 'skip') return softSkip(o.disposition, o.reason);
+    const v = judgeTrialMemoryIsolated(memoryCalls(o.events, AGENT), n);
+    if (v.kind === 'blocked') return softSkip('blocked', v.reason);
+    for (const x of v.findings) expect(x.ok, req(ID_MEMORY_ISOLATED, 'agent-eval-suite.schema.json trials, fixtures.md memory contract (RFC 0244 §B)', x.message)).toBe(true);
+  });
+
   type CounterRuns = { skip: [Parameters<typeof softSkip>[0], string] } | { a: CounterCall[]; b: CounterCall[]; c: CounterCall[]; aMisses: unknown };
   let counterSeq: Promise<CounterRuns> | undefined;
   /** A → B → C, strictly in order, once per worker: the three counter legs read the same runs. */