Pin generation identities and preserve portable model limits

.github/workflows/benchmark-calibration.yml

@@ -347,6 +347,7 @@ jobs:
           OPENROUTER_API_KEY: ${{ secrets.OPENROUTER_API_KEY }}
         run: |
           bun run bench:verify-generations -- \
+            --screen-profile ../provisional-models.json \
             --result "${RUNNER_TEMP}/verified-slots/01/report.json" \
             --receipt "${RUNNER_TEMP}/verified-slots/01/receipt.json" \
             --result "${RUNNER_TEMP}/verified-slots/02/report.json" \

.github/workflows/release.yml

@@ -511,6 +511,7 @@ jobs:
           OPENROUTER_API_KEY: ${{ secrets.OPENROUTER_API_KEY }}
         run: |
           bun run bench:verify-generations -- \
+            --screen-profile ../provisional-models.json \
             --result "${RUNNER_TEMP}/bench-live-reports/slots/01/report.json" \
             --receipt "${RUNNER_TEMP}/bench-live-reports/slots/01/receipt.json" \
             --result "${RUNNER_TEMP}/bench-live-reports/slots/02/report.json" \

ARCHITECTURE.md

@@ -210,10 +210,11 @@ acquire diff --> parse supported lockfiles --> parse + index --> bounded evidenc
   An operator can activate the embedded provisional roster with
   `POSTIL_PROVISIONAL_HOSTED_ROSTER=1` while the formal admission manifest is
   empty. `provisional-models.json` fixes the managed endpoint, upstream provider,
-  model chains, consensus, and price ceilings. The release verifier requires that
-  profile to match `config.toml`, and the runtime applies the same provider pin,
-  privacy policy, response-identity checks, and operation cost cap used by an
-  admitted roster. Removing the flag restores the attested-profile requirement.
+  model chains, canonical provider generation identities, consensus, and price
+  ceilings. The release verifier requires that profile to match `config.toml`,
+  and the runtime applies the same provider pin, privacy policy,
+  response-identity checks, and operation cost cap used by an admitted roster.
+  Removing the flag restores the attested-profile requirement.
 
 ## Prompt-injected policy sources
 

bench/README.md

@@ -468,9 +468,10 @@ canonical slot directory and attests the running receipt before inference,
 then attests its completed report and receipt together. Every accepted provider
 response contributes its OpenRouter generation ID to the report. The fan-in job
 verifies globally distinct generation IDs against OpenRouter's authenticated
-generation API, including the exact model, provider, token totals, and cost,
-then verifies every subject against the exact repository, release workflow,
-source commit, tag ref, OIDC issuer, and GitHub-hosted runner before parsing it.
+generation API, including the exact canonical provider model pinned for each
+logical profile model, provider, token totals, and cost. It also verifies every
+subject against the exact repository, release workflow, source commit, tag ref,
+OIDC issuer, and GitHub-hosted runner before parsing it.
 Only the unique first workflow run for
 the version tag is authoritative. Tag-scoped concurrency, an existing-release
 check, and duplicate-run rejection prevent a second publisher path. A failed

bench/baseline.json

@@ -2,7 +2,7 @@
   "schemaVersion": 2,
   "corpus": {
     "fixtureCorpusSha256": "8e4c2cb9ad5a7efdfe6a875566d20133e905155b6f693a873595adf6c069e065",
-    "evaluatorSha256": "00b424fcc4fb13a1c4bd4010ef79a992d0692d59ceb70eae0669ea9823e6645e"
+    "evaluatorSha256": "b3c6d4085f173ef3abf280b38d4e4221bd39b7752cfbe210e5c19706b700b7e2"
   },
   "profiles": {
     "z-ai/glm-5.2": {

bench/src/generation-evidence.test.ts

@@ -12,6 +12,9 @@ function sample(
       runId: "fixture-run",
       ranAt: "2026-08-26T12:00:30.000Z",
       model: "openai/gpt-5.6-luna",
+      scorerMode: "disabled",
+      scorerModel: null,
+      screeningProfileSha256: "c".repeat(64),
       upstreamProviderIdentity: "Azure",
       totalTokens: { prompt: 30, completion: 12, total: 42 },
       observedProviderCostUsdDecimal: "0.0042",
@@ -55,7 +58,7 @@ const records = {
   "gen-one": {
     id: "gen-one",
     created_at: "2026-08-26T12:00:10.000Z",
-    model: "openai/gpt-5.6-luna",
+    model: "openai/gpt-5.6-luna-20260709",
     provider_name: "Azure",
     tokens_prompt: 10,
     tokens_completion: 5,
@@ -64,18 +67,26 @@ const records = {
   "gen-two": {
     id: "gen-two",
     created_at: "2026-08-26T12:00:20.000Z",
-    model: "openai/gpt-5.6-luna",
+    model: "openai/gpt-5.6-luna-20260709",
     provider_name: "Azure",
     tokens_prompt: 20,
     tokens_completion: 7,
     total_cost: 0.0027,
   },
 };
 
+const profile = {
+  sha256: "c".repeat(64),
+  providerGenerationModels: {
+    "openai/gpt-5.6-luna": "openai/gpt-5.6-luna-20260709",
+  },
+};
+
 describe("provider generation evidence", () => {
   test("verifies distinct generation identity, route, tokens, and cost", async () => {
     await expect(verifyGenerationEvidence([sample(["gen-one", "gen-two"])], {
       apiKey: "fixture",
+      profile,
       fetchImpl: generationFetch(records),
     })).resolves.toBe(2);
   });
@@ -86,6 +97,7 @@ describe("provider generation evidence", () => {
       sample(["gen-one", "gen-two"]),
     ], {
       apiKey: "fixture",
+      profile,
       fetchImpl: generationFetch(records),
     })).rejects.toThrow("duplicate provider generation IDs");
   });
@@ -95,20 +107,116 @@ describe("provider generation evidence", () => {
       sample(["gen-one", "gen-two"], { totalTokens: { prompt: 31, completion: 12, total: 43 } }),
     ], {
       apiKey: "fixture",
+      profile,
       fetchImpl: generationFetch(records),
     })).rejects.toThrow("token totals do not match provider generations");
     await expect(verifyGenerationEvidence([sample(["gen-one", "gen-two"])], {
       apiKey: "fixture",
+      profile,
       fetchImpl: generationFetch({
         ...records,
         "gen-two": { ...records["gen-two"], provider_name: "Other" },
       }),
     })).rejects.toThrow("generation from another provider");
   });
 
+  test("binds the logical alias to the exact provider generation model", async () => {
+    await expect(verifyGenerationEvidence([sample(["gen-one", "gen-two"])], {
+      apiKey: "fixture",
+      profile,
+      fetchImpl: generationFetch({
+        ...records,
+        "gen-two": { ...records["gen-two"], model: "openai/gpt-5.6-luna-20260801" },
+      }),
+    })).rejects.toThrow("generation for another model");
+
+    await expect(verifyGenerationEvidence([
+      sample(["gen-one", "gen-two"], { screeningProfileSha256: "d".repeat(64) }),
+    ], {
+      apiKey: "fixture",
+      profile,
+      fetchImpl: generationFetch(records),
+    })).rejects.toThrow("does not match its screening profile");
+  });
+
+  test("binds each report to only its generator and scorer identities", async () => {
+    const multipleModels = {
+      sha256: "c".repeat(64),
+      providerGenerationModels: {
+        "logical/generator": "provider/generator-20260801",
+        "logical/other": "provider/other-20260801",
+        "logical/scorer": "provider/scorer-20260801",
+      },
+    };
+    const generatorRecord = {
+      ...records["gen-one"],
+      model: "provider/generator-20260801",
+    };
+    const otherRecord = {
+      ...records["gen-two"],
+      model: "provider/other-20260801",
+    };
+    const generatorReport = sample(["gen-one", "gen-two"], {
+      model: "logical/generator",
+    });
+    await expect(verifyGenerationEvidence([generatorReport], {
+      apiKey: "fixture",
+      profile: multipleModels,
+      fetchImpl: generationFetch({
+        "gen-one": generatorRecord,
+        "gen-two": otherRecord,
+      }),
+    })).rejects.toThrow("generation for another model");
+
+    const scorerRecord = {
+      ...records["gen-two"],
+      model: "provider/scorer-20260801",
+    };
+    await expect(verifyGenerationEvidence([
+      sample(["gen-one", "gen-two"], {
+        model: "logical/generator",
+        scorerMode: "enabled",
+        scorerModel: "logical/scorer",
+      }),
+    ], {
+      apiKey: "fixture",
+      profile: multipleModels,
+      fetchImpl: generationFetch({
+        "gen-one": generatorRecord,
+        "gen-two": scorerRecord,
+      }),
+    })).resolves.toBe(2);
+  });
+
+  test("rejects ambiguous provider identity maps", async () => {
+    await expect(verifyGenerationEvidence([sample(["gen-one", "gen-two"])], {
+      apiKey: "fixture",
+      profile: {
+        sha256: "c".repeat(64),
+        providerGenerationModels: {
+          "logical/one": "provider/shared-20260801",
+          "logical/two": "provider/shared-20260801",
+        },
+      },
+      fetchImpl: generationFetch(records),
+    })).rejects.toThrow("must not repeat canonical models");
+
+    await expect(verifyGenerationEvidence([sample(["gen-one", "gen-two"])], {
+      apiKey: "fixture",
+      profile: {
+        sha256: "c".repeat(64),
+        providerGenerationModels: {
+          "openai/gpt-5.6-luna": "openai/gpt-5.6-luna",
+        },
+      },
+      fetchImpl: generationFetch(records),
+    })).rejects.toThrow("must be distinct from logical model IDs");
+  });
+
   test("rejects a lookup whose returned generation identity differs", async () => {
     await expect(verifyGenerationEvidence([sample(["gen-one", "gen-two"])], {
       apiKey: "fixture",
+      profile,
       fetchImpl: generationFetch({
         ...records,
         "gen-two": { ...records["gen-two"], id: "gen-one" },
@@ -119,6 +227,7 @@ describe("provider generation evidence", () => {
   test("binds every generation to the attested receipt interval", async () => {
     await expect(verifyGenerationEvidence([sample(["gen-one", "gen-two"])], {
       apiKey: "fixture",
+      profile,
       fetchImpl: generationFetch({
         ...records,
         "gen-two": { ...records["gen-two"], created_at: "2024-01-01T00:00:00.000Z" },
@@ -131,13 +240,15 @@ describe("provider generation evidence", () => {
       sample(["gen-one", "gen-two"], {}, { reportRawSha256: "b".repeat(64) }),
     ], {
       apiKey: "fixture",
+      profile,
       fetchImpl: generationFetch(records),
     })).rejects.toThrow("does not match its receipt digest");
 
     await expect(verifyGenerationEvidence([
       sample(["gen-one", "gen-two"], {}, { runId: "another-run" }),
     ], {
       apiKey: "fixture",
+      profile,
       fetchImpl: generationFetch(records),
     })).rejects.toThrow("does not match its receipt run identity");
   });