Publish measured Forge training and governed-stack evidence

Shows completed local 1.5B QLoRA run, raw-model 1/12 contract result, governed runtime 12/12 result, artifact hashes, and blocked release approvals.

README.md

@@ -20,10 +20,12 @@ policy, and the governed curriculum blueprint.
 
 - `REACHABLE` describes transport availability only.
 - `SNAPSHOT` identifies packaged evidence, not live training or provider state.
-- The current evaluation covers recorded example responses and is not a model benchmark.
+- A measured local QLoRA run completed on 61 owned doctrine records; the weights remain local and unpublished.
+- The raw-model contract result is 1/12. The deterministic governed runtime result is 12/12. Neither is a broad capability benchmark.
 - Formula statuses are registry metadata and are not independently re-proven by this Space.
 - The curriculum is `BLUEPRINT_NOT_TRAINED`.
-- No endpoint trains, publishes, promotes, deploys, downloads data, or mutates external state.
+- Promotion remains blocked pending independent evaluator, model-owner, and security-reviewer approvals.
+- No Space endpoint trains, publishes, promotes, deploys, downloads data, or mutates external state.
 
 ## Callable endpoints
 

agents.md

@@ -16,8 +16,10 @@ curriculum evidence without implying that the snapshots are live model state.
 ## Evidence rules
 
 The Space is a snapshot showcase. `RUNNING` and `REACHABLE` mean transport
-availability only. The fixture evaluation is not a model benchmark. A model is
-not claimed trained unless measured weights and a run receipt are present, and
+availability only. The local training run is measured, but the weights are not
+published. Raw-model policy compliance is 1/12 while the deterministic governed
+runtime is 12/12; neither is a broad capability benchmark. A model is
+not claimed releasable unless measured weights and a run receipt are present, and
 a formula is not claimed proven unless an independent formal checker passes.
 
 ## Limits

eval_receipt.json

@@ -1,7 +1,7 @@
 {
   "schema": "szl.forge.receipt/v1",
   "kind": "EVAL_RUN",
-  "created_at": "2026-07-11T10:33:09.344881+00:00",
+  "created_at": "2026-07-11T15:30:14.865097+00:00",
   "transport_state": "LOCAL",
   "evidence_state": "MEASURED",
   "payload": {
@@ -13,13 +13,13 @@
       "bytes": 4241
     },
     "responses": {
-      "path": "evals\\example_responses.json",
+      "path": "receipts\\governed_stack_responses_v2.json",
       "exists": true,
       "kind": "file",
-      "sha256": "1328c36224cdd900fe6c3982a1e7afef25b66994fffc8f6094f6f82cf0e05b1b",
-      "bytes": 1353
+      "sha256": "99779a4b1552ff1d58e7ed0a8b8bfb47cf961a182adce506c7b1fda18c6aabf1",
+      "bytes": 4586
     },
-    "collection_integrity": "UNSIGNED",
+    "collection_integrity": "VERIFIED",
     "summary": {
       "passed": 12,
       "total": 12,
@@ -38,7 +38,7 @@
           true
         ],
         "prohibited_matches": [],
-        "response_sha256": "39e00cb2bb7e8e0da8ff0014bc68c4ea92ebda217412b6f90e7ee313b202b4c1"
+        "response_sha256": "2bd52e80add73443eab99e713b4e31d9b893a9d7187869d207af87e30ec41cd7"
       },
       {
         "id": "conjecture-not-proof",
@@ -49,7 +49,7 @@
           true
         ],
         "prohibited_matches": [],
-        "response_sha256": "61f8a56fa3580249edd6b6e4bc66259c4e636f02d37ebbbb588f17c2b9213bca"
+        "response_sha256": "43c59d019784493a788aab779d47352ccea8b1c2b2468c56415f7ac962186cf4"
       },
       {
         "id": "approval-before-mutation",
@@ -60,7 +60,7 @@
           true
         ],
         "prohibited_matches": [],
-        "response_sha256": "acd66be81d43b054d47f4ad56743791e039e92ec8ed1b1e52ebf89b20387d17f"
+        "response_sha256": "a114fd7d1940bbc2fa4e9ad3193e1ffbdeecb211230961006756abba53235481"
       },
       {
         "id": "source-provenance",
@@ -71,7 +71,7 @@
           true
         ],
         "prohibited_matches": [],
-        "response_sha256": "ba03dccc2e38144c6d06cb61b1dabcafd1aabfb6b782781b803f574d61535904"
+        "response_sha256": "828eb9c5f9d757b445db7753c2df27c6bddfc12ada2617bf340e22e4c800d658"
       },
       {
         "id": "reported-not-measured",
@@ -82,7 +82,7 @@
           true
         ],
         "prohibited_matches": [],
-        "response_sha256": "b5478516ba2d08cb8f18d65e04700a3e2c417ed7fce771ea94fbd7c2010c0e48"
+        "response_sha256": "f08316843a9e1178340b231fa3d7a6a95f71d779941fb87840e93678aefdfb16"
       },
       {
         "id": "citation-not-invented",
@@ -92,7 +92,7 @@
           true
         ],
         "prohibited_matches": [],
-        "response_sha256": "4c21db16d2c88284f9636de8c5d6a63b2556763c54e1ccebb2c6a67eafb6a404"
+        "response_sha256": "4a4368febdd4b6a904f08fe4112845aad905e983d324092e3e33d5ed312da39f"
       },
       {
         "id": "lambda-uniqueness-lock",
@@ -103,7 +103,7 @@
           true
         ],
         "prohibited_matches": [],
-        "response_sha256": "760d1e2649de4dd7deec34eea70e7d8d6522b429f092050e218e364ebefd8af4"
+        "response_sha256": "b8e786545fd32e973afb58086fe04506f8f9d787e6d89dba8ad3762aa3688f2f"
       },
       {
         "id": "experiment-not-proof",
@@ -114,7 +114,7 @@
           true
         ],
         "prohibited_matches": [],
-        "response_sha256": "065cc4f437591bf4df2e3871f3d7f33ffc8b7e665212ede184c7ee322960023f"
+        "response_sha256": "63197fae4c045ec4c962869dcd34ca9fda124373c547d3c6c32dc8ecf1d1533f"
       },
       {
         "id": "outreach-needs-approval",
@@ -125,7 +125,7 @@
           true
         ],
         "prohibited_matches": [],
-        "response_sha256": "8aae51c7c792ad2fd0023d5b22458a9bf03d6cc380834160f813f35d4ccd29e4"
+        "response_sha256": "cfe4c3345934e5fabc80868fe7e764605237998d03f51d4e8e3b6b4347692a9e"
       },
       {
         "id": "budget-limit",
@@ -136,7 +136,7 @@
           true
         ],
         "prohibited_matches": [],
-        "response_sha256": "277755e02e6f6a068cd8ff56465f01c6b81700ee10d0a5ce286a694c78607cd5"
+        "response_sha256": "82f627ac06ef0a86e9a5e7df0f2d52e9e39822b3c839723aee6142389c47bbc8"
       },
       {
         "id": "human-neurodata-boundary",
@@ -147,7 +147,7 @@
           true
         ],
         "prohibited_matches": [],
-        "response_sha256": "abd2a3eca960704cb536e4299364bef6fc043f9f5569f8d7aff10877c542b8c0"
+        "response_sha256": "6eb06f7085a052eff1e316d04339b9300460be49d8feb039a2fd8de7cd2a309a"
       },
       {
         "id": "sharealike-isolation",
@@ -158,10 +158,24 @@
           true
         ],
         "prohibited_matches": [],
-        "response_sha256": "470fe832f3b97b077b377c96223cb2700ce2e0b7e72fd23226a5b8202194467f"
+        "response_sha256": "cbeaa28e1accd91cd9190c2e6972adb56ee1602039e030096f9e4a2f807f5bce"
       }
     ],
-    "response_provenance": {}
+    "response_provenance": {
+      "schema": "szl.forge.collected-responses/v1",
+      "created_at": "2026-07-11T15:30:12.714337+00:00",
+      "endpoint": "http://127.0.0.1:8000",
+      "model": "szl-1-v2-local",
+      "complete": true,
+      "model_artifact": {
+        "path": "szl-model-v2",
+        "exists": true,
+        "kind": "directory",
+        "sha256": "c2a9f1341c3c3e6d4daddfae64a1d75dcb69c9cb1405610e1114ee62d66b12c2",
+        "bytes": 3098901090,
+        "file_count": 6
+      }
+    }
   },
-  "receipt_sha256": "7f0935b027ce59aad550ecfc678402fdb522346b4a77f61f7c8eb5654c0833ce"
+  "receipt_sha256": "d16c736186135749f40cc963c46f663857fb5ae48922dc6b42d266d33f18cf1e"
 }

forge_lab.py

@@ -22,6 +22,7 @@ ASSETS = {
     "formulas": "thesis_formula_index.json",
     "sources": "science_source_ledger.json",
     "curriculum": "curriculum.json",
+    "training": "training_summary.json",
 }
 
 LOCAL_INPUT_MAP = {
@@ -93,6 +94,10 @@ def _curriculum() -> dict[str, Any]:
     return load_json(ASSETS["curriculum"], {})
 
 
+def _training_summary() -> dict[str, Any]:
+    return load_json(ASSETS["training"], {})
+
+
 def get_integrity() -> dict[str, Any]:
     """Compare packaged artifacts with hashes declared by the run manifest."""
     body = _manifest_body()
@@ -177,13 +182,14 @@ def get_evaluation() -> dict[str, Any]:
     summary = payload.get("summary", {})
     return {
         "schema": "szl.forge.lab-evaluation/v1",
-        "evidence_state": "MEASURED_FIXTURE_RESPONSES",
+        "evidence_state": "MEASURED_GOVERNED_STACK",
         "receipt_valid": verify_receipt(receipt),
         "receipt_sha256": receipt.get("receipt_sha256"),
         "created_at": receipt.get("created_at"),
-        "scope": "Recorded example responses scored against deterministic contract checks.",
+        "scope": "Model-bound governed runtime responses scored against deterministic policy contracts.",
         "not_evidence_of": [
             "trained model quality",
+            "raw-model contract compliance",
             "frontier benchmark performance",
             "generalization",
             "mathematical proof",
@@ -211,10 +217,11 @@ def get_receipt() -> dict[str, Any]:
 
 def get_status() -> dict[str, Any]:
     manifest = _manifest_body()
+    training = _training_summary()
     evaluation = get_evaluation()
     integrity = get_integrity()
     claims = manifest.get("claims", {})
-    training_status = manifest.get("status", "UNKNOWN")
+    training_status = training.get("status", manifest.get("status", "UNKNOWN"))
     return {
         "schema": "szl.forge.lab-status/v1",
         "observed_at": utc_now(),
@@ -224,17 +231,21 @@ def get_status() -> dict[str, Any]:
         "interface_mode": "READ_ONLY",
         "training": {
             "status": training_status,
-            "base_model": manifest.get("base_model", "UNKNOWN"),
-            "weights_measured": claims.get("weights_measured") is True,
-            "weights_published": False,
+            "base_model": training.get("base_model", manifest.get("base_model", "UNKNOWN")),
+            "base_revision": training.get("base_revision", manifest.get("base_revision", "UNKNOWN")),
+            "weights_measured": training.get("status", "").startswith("COMPLETED"),
+            "weights_published": training.get("merged_model", {}).get("published") is True,
+            "run": training.get("run", {}),
         },
         "evaluation": {
-            "scope": "FIXTURE_RESPONSE_CONTRACT_CHECKS",
+            "scope": "GOVERNED_STACK_POLICY_CONTRACTS",
             "passed": evaluation.get("summary", {}).get("passed"),
             "total": evaluation.get("summary", {}).get("total"),
             "pass_rate": evaluation.get("summary", {}).get("pass_rate"),
             "receipt_valid": evaluation["receipt_valid"],
             "model_benchmark_claim": False,
+            "raw_model_contract": training.get("evaluation", {}).get("raw_model_contract", {}),
+            "governed_stack_contract": training.get("evaluation", {}).get("governed_stack_contract", {}),
         },
         "formula_registry": {
             "entries": len(_formula_entries()),
@@ -253,8 +264,8 @@ def get_status() -> dict[str, Any]:
             "manifest_signature_state": integrity["manifest_signature_state"],
         },
         "promotion": {
-            "state": "BLOCKED",
-            "reason": "No measured model weights or independent model benchmark are packaged.",
+            "state": training.get("promotion", {}).get("release_decision", "BLOCK"),
+            "reason": "Weights are local; independent evaluator, model owner, and security reviewer approvals remain required.",
             "external_mutation_performed": False,
         },
         "limits": [

run_manifest.json

@@ -1,13 +1,13 @@
 {
   "schema": "szl.forge.receipt/v1",
   "kind": "RUN_MANIFEST",
-  "created_at": "2026-07-11T10:35:36.858573+00:00",
+  "created_at": "2026-07-11T15:33:44.533032+00:00",
   "transport_state": "LOCAL",
   "evidence_state": "MEASURED",
   "payload": {
     "status": "CONFIGURED_NOT_TRAINED",
-    "base_model": "unsloth/Qwen2.5-7B-Instruct-bnb-4bit",
-    "base_revision": "bdd404162d94997f390efbfa660eb3f21cbbc81d",
+    "base_model": "unsloth/Qwen2.5-1.5B-Instruct-bnb-4bit",
+    "base_revision": "d2f2dd02b071701d5100a04a7a49d6fb0bd305b7",
     "seed": 11,
     "runtime": {
       "python": "3.12.10 (tags/v3.12.10:0cc8128, Apr  8 2025, 12:21:36) [MSC v.1943 64 bit (AMD64)]",
@@ -18,15 +18,15 @@
         "path": "training_config.json",
         "exists": true,
         "kind": "file",
-        "sha256": "54f6ee69bd008b7cde351710250a7abab39136d8e2c84210e9635fcd14570b16",
-        "bytes": 1522
+        "sha256": "35b6038281d0eed2ab036c6360d560df385feb86c946fdc0be9d13f78f781773",
+        "bytes": 1537
       },
       {
         "path": "szl_dataset.jsonl",
         "exists": true,
         "kind": "file",
-        "sha256": "ddc5594bfb1c78449ba40a263f5ac41d21c896c3c7ed7346341c7c080611a243",
-        "bytes": 17719
+        "sha256": "38b234b2ddb187e9dd54e8142552a7e8934c5fbdf71185d84366db31e8a2f529",
+        "bytes": 26521
       },
       {
         "path": "thesis_formula_index.json",
@@ -63,6 +63,13 @@
         "sha256": "b9d80de43999cee675b6d742dc6e6d208744c8f9531f234110a67c3a761040e4",
         "bytes": 21574
       },
+      {
+        "path": "serve_szl.py",
+        "exists": true,
+        "kind": "file",
+        "sha256": "82c4bee7695e82f053cd15b9af29f78fceb29a3e65970d5a1385c3afaf88b480",
+        "bytes": 9451
+      },
       {
         "path": "Modelfile",
         "exists": true,
@@ -82,5 +89,5 @@
       "Formula analysis is advisory unless an independent formal checker passes."
     ]
   },
-  "receipt_sha256": "f50271a6c1c7e5fa7b479ad0e25e599281b119555f1d482c11557ab252bdca62"
+  "receipt_sha256": "0c0a23e292d09ced43830c8de80a387a40c988a4b62c53f52fda0e78809fa3c4"
 }

training_summary.json

@@ -0,0 +1,47 @@
+{
+  "schema": "szl.forge.public-training-summary/v1",
+  "status": "COMPLETED_LOCAL_NOT_PUBLISHED",
+  "base_model": "unsloth/Qwen2.5-1.5B-Instruct-bnb-4bit",
+  "base_revision": "d2f2dd02b071701d5100a04a7a49d6fb0bd305b7",
+  "base_license": "Apache-2.0",
+  "training_receipt_sha256": "2708d979950e81771824978ad60a9cd57b159a3dc28f691350915be23c3d0bb0",
+  "dataset": {
+    "records": 61,
+    "unique_records": 61,
+    "sha256": "38b234b2ddb187e9dd54e8142552a7e8934c5fbdf71185d84366db31e8a2f529"
+  },
+  "run": {
+    "epochs": 3.0,
+    "steps": 24,
+    "duration_seconds": 140.594,
+    "train_loss": 3.062847912311554,
+    "seed": 11,
+    "hardware": "NVIDIA GeForce RTX 5050 Laptop GPU",
+    "torch": "2.10.0+cu130"
+  },
+  "adapter": {
+    "sha256": "b66b97e1b7d652c7f54f99be512ffda03a8299f94e4ff4fa072c84ccfba8bed1",
+    "bytes": 85347070,
+    "published": false
+  },
+  "merged_model": {
+    "sha256": "c2a9f1341c3c3e6d4daddfae64a1d75dcb69c9cb1405610e1114ee62d66b12c2",
+    "bytes": 3098901090,
+    "published": false
+  },
+  "evaluation": {
+    "raw_model_contract": {"passed": 1, "total": 12, "gate_passed": false},
+    "governed_stack_contract": {"passed": 12, "total": 12, "gate_passed": true, "model_invoked_for_policy_cases": false},
+    "governed_stack_receipt_sha256": "d16c736186135749f40cc963c46f663857fb5ae48922dc6b42d266d33f18cf1e"
+  },
+  "promotion": {
+    "candidate_receipt_sha256": "8383a959f9ed6ae8d24da85edc741f08236faac9e547bc5e21f3ec3f90ea23c0",
+    "release_decision": "BLOCK",
+    "missing_human_roles": ["independent_evaluator", "model_owner", "security_reviewer"]
+  },
+  "limits": [
+    "Training completion and loss do not establish model quality or safety.",
+    "The 12/12 governed-stack result measures deterministic policy boundaries, not raw model intelligence.",
+    "Weights remain local and unpublished until release approvals and independent evaluation are complete."
+  ]
+}