diff --git a/recipe/jemm/export.py b/recipe/jemm/export.py
index 21f05b20..f0f599ce 100644
--- a/recipe/jemm/export.py
+++ b/recipe/jemm/export.py
@@ -31,8 +31,6 @@
 LABELS = "ABCDEFGHIJKLMNOPQRSTUVWXYZ012345"
 SYSTEM = ("Choose the best available candidate for the question using only the supplied state. "
           "Return exactly one candidate label.")
-CALIBRATION = {"temperature": 1.3480874159655591, "mm_temperature": 1.3954832341582943,
-               "threshold": 0.9872681877423998}
 INVENTORY_PATH = Path(__file__).with_name("pinned_inventory.json")
 
 
@@ -189,9 +187,14 @@ def export_checkpoint(base, adapter, out, provenance, *, tokenizer=None, invento
     pairs, scale = adapter_pairs(inventory["base"]["tensors"], inventory["adapter"]["tensors"], adapter_config)
     if inventory.get("provenance") == PINS and (adapter_config["r"] != 16 or len(pairs) != 496 or scale != 2.):
         raise ValueError("pinned JEMM requires rank16, 496 FP32 LoRA pairs and scale2")
-    calibration = json.loads((adapter / "decision_config.json").read_text())
-    if any(calibration.get(key) != value for key, value in CALIBRATION.items()):
-        raise ValueError("decision calibration differs from pinned JEMM")
+    decision = json.loads((adapter / "decision_config.json").read_text())
+    calibration = {}
+    for key in ("temperature", "mm_temperature", "threshold"):
+        value = decision.get(key) if isinstance(decision, dict) else None
+        if type(value) not in (int, float) or not math.isfinite(value) or (
+                value <= 0 if key != "threshold" else not 0 <= value <= 1):
+            raise ValueError(f"invalid decision calibration: {key}")
+        calibration[key] = value
     if tokenizer is None:
         from transformers import AutoTokenizer
         tokenizer = AutoTokenizer.from_pretrained(str(base), local_files_only=True)
@@ -310,7 +313,7 @@ def export_checkpoint(base, adapter, out, provenance, *, tokenizer=None, invento
         raise ValueError("original FP32 adapter checksum mismatch")
     if "jemm_lm_head.safetensors" not in exports or "vision.safetensors" not in exports:
         raise ValueError("export lacks selected head or vision tensors")
-    manifest = {"format": "jemm-native/2", "model_id": "JEMM", **PINS, **CALIBRATION,
+    manifest = {"format": "jemm-native/2", "model_id": "JEMM", **PINS, **calibration,
                 "max_tokens": 8192, "max_mm_tokens": 3072, "chat_prefix": chat_prefix, "chat_suffix": chat_suffix,
                 "enable_thinking": False, "preserve_thinking": False, "labels": list(LABELS), "label_token_ids": label_ids,
                 "label_head_file": "jemm_lm_head.safetensors", "label_head_tensor": "weight", "label_head_dtype": "BF16",
diff --git a/src/models/jemm/README.md b/src/models/jemm/README.md
index 5238ebc1..ccd77bfe 100644
--- a/src/models/jemm/README.md
+++ b/src/models/jemm/README.md
@@ -21,10 +21,11 @@ rows from the untied LM head are extracted without merging adapter weights.
 `jemm_export.json` records every exported file's SHA-256 and the adapter contract.
 Startup verifies original shard and adapter pins before CUDA initialization;
 legacy premerged `jemm-native/1` exports are explicitly rejected. Keep files immutable
-while serving. This head requires the pinned temperatures 1.3480874159655591
-(text), 1.3954832341582943 (images) and threshold 0.9872681877423998. Follow-up
-#122 loads validated calibration from the checksum-covered decision config;
-its calibration policy is a separate change. The threshold is provenance and
+while serving. Calibration loads from checksum-covered `decision_config.json`;
+temperatures must be finite and positive, threshold finite and within [0,1],
+and all values must agree with the export manifest. The pinned artifact values
+are 1.3480874159655591 (text), 1.3954832341582943 (images) and
+0.9872681877423998 (threshold). The threshold is provenance and
 does not gate SystemOne responses.
 
 ## Request and response contract
diff --git a/src/models/jemm/native/src/artifacts.rs b/src/models/jemm/native/src/artifacts.rs
index f687dcb0..f3fd7847 100644
--- a/src/models/jemm/native/src/artifacts.rs
+++ b/src/models/jemm/native/src/artifacts.rs
@@ -104,6 +104,7 @@ pub fn verify_export_files(dir: &Path, manifest: &Value) -> Result<()> {
         "config.json",
         "tokenizer.json",
         "preprocessor_config.json",
+        "decision_config.json",
         "model.safetensors.index.json",
         "vision.safetensors",
         "jemm_lm_head.safetensors",
diff --git a/src/models/jemm/native/src/contract.rs b/src/models/jemm/native/src/contract.rs
index a63147d0..ec26403d 100644
--- a/src/models/jemm/native/src/contract.rs
+++ b/src/models/jemm/native/src/contract.rs
@@ -11,9 +11,6 @@ pub const LABELS: &str = "ABCDEFGHIJKLMNOPQRSTUVWXYZ012345";
 pub const SYSTEM: &str = "Choose the best available candidate for the question using only the supplied state. Return exactly one candidate label.";
 pub const BASE_REVISION: &str = "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0";
 pub const CHECKPOINT_REVISION: &str = "76e3c209e8441fa658221c7ba2725bad2f811176";
-pub const TEMPERATURE: f64 = 1.3480874159655591;
-pub const MM_TEMPERATURE: f64 = 1.3954832341582943;
-pub const THRESHOLD: f64 = 0.9872681877423998;
 pub const SOURCE_REVISION: &str = "6822fe0fd53c5e6670af6ba99fb2c857a661e532";
 #[derive(Debug)]
 pub struct Question {
diff --git a/src/models/jemm/native/src/processing.rs b/src/models/jemm/native/src/processing.rs
index 3538da9f..759b811d 100644
--- a/src/models/jemm/native/src/processing.rs
+++ b/src/models/jemm/native/src/processing.rs
@@ -200,16 +200,33 @@ impl Processor {
             prefix.contains(contract::SYSTEM) && suffix.contains("<think>\n\n</think>"),
             "expected JEMM system and disabled-thinking template"
         );
-        let temperature = manifest["temperature"].as_f64().context("temperature")?;
-        let mm_temperature = manifest["mm_temperature"]
-            .as_f64()
-            .context("mm_temperature")?;
-        ensure!(
-            temperature == contract::TEMPERATURE
-                && mm_temperature == contract::MM_TEMPERATURE
-                && manifest["threshold"].as_f64() == Some(contract::THRESHOLD),
-            "expected pinned JEMM calibration"
-        );
+        let calibration: Value = serde_json::from_slice(
+            &std::fs::read(dir.join("decision_config.json")).context("decision_config.json")?,
+        )?;
+        let mut values = [0.0; 3];
+        for (index, key) in ["temperature", "mm_temperature", "threshold"]
+            .iter()
+            .enumerate()
+        {
+            let value = calibration[key]
+                .as_f64()
+                .with_context(|| format!("calibration {key}"))?;
+            ensure!(
+                value.is_finite()
+                    && if index < 2 {
+                        value > 0.0
+                    } else {
+                        (0.0..=1.0).contains(&value)
+                    },
+                "invalid calibration {key}"
+            );
+            ensure!(
+                manifest[key].as_f64() == Some(value),
+                "manifest disagrees with decision_config.json: {key}"
+            );
+            values[index] = value;
+        }
+        let [temperature, mm_temperature, _threshold] = values;
         let config: Value = serde_json::from_slice(&std::fs::read(dir.join("config.json"))?)?;
         let image_token = config["image_token_id"]
             .as_u64()
diff --git a/tests/jemm/test_export.py b/tests/jemm/test_export.py
index 75a652dd..c4997920 100644
--- a/tests/jemm/test_export.py
+++ b/tests/jemm/test_export.py
@@ -62,7 +62,9 @@ def setUp(self):
         self.write_adapter()
         self.config = {"peft_type": "LORA", "base_model_name_or_path": EXPORT.BASE_MODEL_ID, "r": 16, "lora_alpha": 32}
         (self.adapter / "adapter_config.json").write_text(json.dumps(self.config))
-        (self.adapter / "decision_config.json").write_text(json.dumps(EXPORT.CALIBRATION))
+        (self.adapter / "decision_config.json").write_text(json.dumps({"temperature": 1.3480874159655591,
+                                                                    "mm_temperature": 1.3954832341582943,
+                                                                    "threshold": 0.9872681877423998}))
         self.refresh_inventory()
 
     def tearDown(self):
@@ -103,6 +105,31 @@ def test_preserves_original_base_and_fp32_adapter_without_merging(self):
         with safe_open(str(self.out / "vision.safetensors"), framework="pt", device="cpu") as reader:
             self.assertTrue(torch.equal(reader.get_tensor("model.visual.patch_embed.weight"), torch.ones(2, 2, dtype=torch.bfloat16)))
 
+    def test_v2_export_uses_checksum_covered_file_calibration(self):
+        calibration = {"temperature": 2., "mm_temperature": 3., "threshold": 0.5}
+        source = self.adapter / "decision_config.json"
+        source.write_text(json.dumps(calibration))
+        self.refresh_inventory()
+        manifest = self.run_export()
+        self.assertEqual(manifest["format"], "jemm-native/2")
+        self.assertEqual(manifest["lora_mode"], "unmerged_fp32")
+        self.assertEqual({key: manifest[key] for key in calibration}, calibration)
+        self.assertEqual((self.out / source.name).read_bytes(), source.read_bytes())
+        self.assertEqual(manifest["export_sha256"][source.name], EXPORT.sha256(source))
+
+    def test_invalid_file_calibration_is_refused_before_output_creation(self):
+        for key, value in [("temperature", 0), ("temperature", True),
+                           ("mm_temperature", -1), ("threshold", 1.1),
+                           ("threshold", float("nan")), ("temperature", None)]:
+            with self.subTest(key=key, value=value):
+                calibration = {"temperature": 2., "mm_temperature": 3., "threshold": 0.5}
+                calibration[key] = value
+                (self.adapter / "decision_config.json").write_text(json.dumps(calibration))
+                self.refresh_inventory()
+                with self.assertRaisesRegex(ValueError, "invalid decision calibration"):
+                    self.run_export()
+                self.assertFalse(self.out.exists())
+
     def test_cross_device_copy_fallback_preserves_identical_source_bytes(self):
         with patch.object(EXPORT.os, "link", side_effect=OSError(errno.EXDEV, "cross-device")):
             manifest = self.run_export()
diff --git a/tests/jemm/tokenization.rs b/tests/jemm/tokenization.rs
index f9bc30f6..8aaca82e 100644
--- a/tests/jemm/tokenization.rs
+++ b/tests/jemm/tokenization.rs
@@ -43,6 +43,13 @@ fn fixture() -> (Fixture, Value) {
     )
     .unwrap();
     let manifest = json!({"format":"jemm-native/2","model_id":"JEMM","base_model_id":"Qwen/Qwen3.8-27B","base_revision":contract::BASE_REVISION,"checkpoint_revision":contract::CHECKPOINT_REVISION,"source_revision":contract::SOURCE_REVISION,"max_tokens":8192,"max_mm_tokens":3072,"label_token_ids":(1..=32).collect::<Vec<_>>(),"chat_prefix":format!("<|im_start|>system\n{}<|im_end|>\n<|im_start|>user\n",contract::SYSTEM),"chat_suffix":"<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n","temperature":1.3480874159655591,"mm_temperature":1.3954832341582943,"threshold":0.9872681877423998});
+    std::fs::write(
+        dir.join("decision_config.json"),
+        serde_json::to_vec(&json!({"temperature":manifest["temperature"],
+            "mm_temperature":manifest["mm_temperature"], "threshold":manifest["threshold"]}))
+        .unwrap(),
+    )
+    .unwrap();
     (Fixture(dir), manifest)
 }
 #[test]
@@ -86,7 +93,7 @@ fn exported_tokenizer_truncation_and_padding_cannot_change_request_tokens() {
     assert!(prepared.inputs[0].ids.len() > 3);
 }
 #[test]
-fn rejects_custom_calibration_for_pinned_adapter() {
+fn rejects_manifest_calibration_disagreement() {
     let (f, mut m) = fixture();
     m["mm_temperature"] = json!(1.);
     assert!(
@@ -94,7 +101,52 @@ fn rejects_custom_calibration_for_pinned_adapter() {
             .err()
             .unwrap()
             .to_string()
-            .contains("pinned JEMM calibration")
+            .contains("manifest disagrees with decision_config.json")
+    );
+}
+
+#[test]
+fn file_driven_calibration_must_be_finite_in_range_and_agree_with_manifest() {
+    for (field, invalid) in [
+        ("temperature", json!(0.0)),
+        ("mm_temperature", json!(-1.0)),
+        ("threshold", json!(1.1)),
+    ] {
+        let (f, mut m) = fixture();
+        m[field] = invalid;
+        std::fs::write(
+            f.0.join("decision_config.json"),
+            serde_json::to_vec(&json!({"temperature":m["temperature"],
+                "mm_temperature":m["mm_temperature"], "threshold":m["threshold"]}))
+            .unwrap(),
+        )
+        .unwrap();
+        assert!(
+            Processor::load(&f.0, &m)
+                .err()
+                .unwrap()
+                .to_string()
+                .contains("invalid calibration")
+        );
+    }
+    let (f, mut m) = fixture();
+    m["temperature"] = json!(2.0);
+    m["mm_temperature"] = json!(3.0);
+    m["threshold"] = json!(0.5);
+    std::fs::write(
+        f.0.join("decision_config.json"),
+        serde_json::to_vec(&json!({"temperature":2.0,"mm_temperature":3.0,"threshold":0.5}))
+            .unwrap(),
+    )
+    .unwrap();
+    let p = Processor::load(&f.0, &m).unwrap();
+    let prepared = p
+        .prepare(br#"{"questions":{"q":{"type":"noul"}}}"#)
+        .unwrap();
+    let response = prepared.context.finish(vec![vec![2.0, 0.0]]).unwrap();
+    assert!(
+        (response["answers"]["q"]["noul"].as_f64().unwrap() - 1.0 / (1.0 + (-1.0f64).exp())).abs()
+            < 1e-12
     );
 }