diff --git a/src/maple/core/calibration/calibration_target_models.py b/src/maple/core/calibration/calibration_target_models.py
index 48954fe..f7a88db 100644
--- a/src/maple/core/calibration/calibration_target_models.py
+++ b/src/maple/core/calibration/calibration_target_models.py
@@ -53,6 +53,7 @@
 )
 from maple.core.calibration.scenario import Scenario
 from maple.core.calibration.shared_models import (
+    DistributionShape,
     EstimateInput,
     InputType,
     ModelingAssumption,
@@ -339,6 +340,34 @@ def validate_observed_distribution_consistency(self) -> "CalibrationTargetEstima
             )
         return self
 
+    @model_validator(mode="after")
+    def validate_bounded_observable_uses_logit_normal(self) -> "CalibrationTargetEstimates":
+        """A bounded observable's population spread (``moments`` form) must use
+        ``shape: logit_normal``, not normal/lognormal.
+
+        Parity with the SubmodelTarget validator of the same name: for a
+        fraction / proportion / probability / percent observable, ``normal`` puts
+        mass outside the bound and ``lognormal`` is unbounded above; ``logit_normal``
+        keeps expanded quartiles in (0, 1). Only applies to the ``moments`` form —
+        the ``quantiles`` form carries the empirical shape directly and is exempt.
+        """
+        od = self.observed_distribution
+        if od is None or od.moments is None:
+            return self
+        if od.moments.shape == DistributionShape.LOGIT_NORMAL:
+            return self
+        BOUNDED_UNITS = {"percent", "%", "fraction", "proportion", "probability"}
+        if (self.units or "").strip().lower() not in BOUNDED_UNITS:
+            return self
+        raise ValueError(
+            f"Observable units='{self.units}' are a bounded fraction/percentage, but "
+            f"observed_distribution.moments uses shape='{od.moments.shape.value}'. "
+            "Bounded observables must use shape='logit_normal', which expands quartiles "
+            "in logit space so they never escape (0, 1); normal puts mass outside the "
+            "bound and lognormal is unbounded above. logit_normal requires center in "
+            "(0, 1) with center_type='median' — express a percent as a fraction (12% -> 0.12)."
+        )
+
     @field_validator("sample_size")
     @classmethod
     def validate_sample_size_positive(cls, v: int) -> int:
diff --git a/src/maple/prompts/calibration_target_prompt.md b/src/maple/prompts/calibration_target_prompt.md
index 20d0f20..1a9a11f 100644
--- a/src/maple/prompts/calibration_target_prompt.md
+++ b/src/maple/prompts/calibration_target_prompt.md
@@ -215,7 +215,7 @@ Use the `quantiles` form when the paper gives quartiles/percentiles/samples dire
     experimental_unit_type: biological
 ```
 
-Provide EXACTLY ONE of `moments` / `quantiles`. A population spread (`spread_source: across_patient`) REQUIRES `n_biological` + `experimental_unit_type: biological`. When both `observed_distribution` and `population_spread` are present they must agree that the width is (or is not) genuine spread — a validator enforces this. For a `[0,1]`-bounded observable (a fraction / proportion / response rate), use `shape: logit_normal` (with `center_type: median`) so expanded quartiles stay in-bounds. If a target ever carries several observables measured on the SAME patients/units and you want them treated as one shared spread, tag them with a common `unit_group` string (rare for full-model calibration targets, which are usually one observable each — omit it otherwise).
+Provide EXACTLY ONE of `moments` / `quantiles`. A population spread (`spread_source: across_patient`) REQUIRES `n_biological` + `experimental_unit_type: biological`. When both `observed_distribution` and `population_spread` are present they must agree that the width is (or is not) genuine spread — a validator enforces this. For a `[0,1]`-bounded observable (a fraction / proportion / response rate), use `shape: logit_normal` (with `center_type: median`) so expanded quartiles stay in-bounds — **a validator rejects a `percent`/`fraction`/`proportion`/`probability` observable whose `moments` shape is `normal` or `lognormal`.** If the source reports the unit count as a LOWER BOUND ("n≥40", "at least 40 patients"), set `n_biological_is_floor: true` (with the floor value for `n_biological`) so precision-weighting consumers don't over-weight the panel. If a target ever carries several observables measured on the SAME patients/units and you want them treated as one shared spread, tag them with a common `unit_group` string (rare for full-model calibration targets, which are usually one observable each — omit it otherwise).
 
 ### Source Relevance Assessment
 
diff --git a/tests/unit/core/test_calibration_target_validators.py b/tests/unit/core/test_calibration_target_validators.py
index 32c0334..69ebf88 100644
--- a/tests/unit/core/test_calibration_target_validators.py
+++ b/tests/unit/core/test_calibration_target_validators.py
@@ -1671,3 +1671,59 @@ def test_degenerate_samples_rejected(
         )
         with pytest.raises(ValidationError, match="zero variance"):
             CalibrationTarget.model_validate(data, context={"model_structure": model_structure})
+
+
+# ============================================================================
+# Cal-side parity for the submodel bounded->logit_normal validator: a bounded
+# observable's moments-form observed_distribution must use shape=logit_normal.
+# ============================================================================
+
+from maple.core.calibration.calibration_target_models import CalibrationTargetEstimates
+
+
+def _bounded_cal_estimates(shape: str) -> dict:
+    return {
+        "median": [0.5],
+        "ci95": [[0.3, 0.7]],
+        "units": "percent",
+        "sample_size": 40,
+        "sample_size_rationale": "n=40 patients, Table 1",
+        "inputs": [
+            {
+                "name": "resp_fraction",
+                "value": 0.5,
+                "units": "percent",
+                "description": "objective response fraction",
+                "source_ref": "smith_2020",
+                "value_location": "Table 1",
+                "value_snippet": "response rate 0.5",
+            }
+        ],
+        "distribution_code": (
+            "def derive_distribution(inputs, ureg):\n"
+            "    v = inputs['resp_fraction']\n"
+            "    return {'median_obs': v, 'ci95_lower': v * 0.6, 'ci95_upper': v * 1.4}"
+        ),
+        "population_spread": "center_only",
+        "observed_distribution": {
+            "moments": {
+                "center": 0.5,
+                "center_type": "median",
+                "scale": 0.1,
+                "scale_type": "sd",
+                "shape": shape,
+            },
+            "spread_source": "center_only",
+        },
+    }
+
+
+class TestCalBoundedObservableLogitNormal:
+    """CalibrationTargetEstimates: bounded moments-form observable must use logit_normal."""
+
+    def test_percent_with_normal_shape_raises(self):
+        with pytest.raises(ValidationError, match="logit_normal"):
+            CalibrationTargetEstimates.model_validate(_bounded_cal_estimates("normal"))
+
+    def test_percent_with_logit_normal_passes(self):
+        CalibrationTargetEstimates.model_validate(_bounded_cal_estimates("logit_normal"))