Summarize the merged PR diff below in at most 3 sentences and at most 90 words. State the concrete behavior changed, including a changed file or symbol from the diff. Use only facts from this input; do not introduce any number or identifier absent from it. Return only the summary.

Public repository: OmniNode-ai/omnimarket
Merged PR: omnimarket#4493
Title: fix(OMN-21128): require sampled evidence for exemplar revocation
Merged at: 2026-10-27T09:37:58Z
Merge SHA: 8ea4f7a0bc9794e48d6fe7d2a99d075f8c93a22a

BEGIN PUBLIC SOURCE
diff --git a/contracts/OMN-21090.yaml b/contracts/OMN-21090.yaml
index 045413459d..6a5b3f3d6a 100644
--- a/contracts/OMN-21090.yaml
+++ b/contracts/OMN-21090.yaml
@@ -257,3 +257,19 @@ dod_evidence:
       - check_type: "command"
         check_value: "uv run --frozen pytest tests/nodes/node_session_compaction_measurement_compute/test_measurement.py -q --no-cov -k 'dropped_decision_negative_control or wall_time_and_usage_stream_dedup or node_contract_and_overlay'"
         cwd: "${OMNI_HOME}/omnimarket"
+
+  - id: "dod-omn21090-ac4-sampled-exemplar-revocation"
+    description: "Track C TC.5 contribution: minimum-sample revocation and count-derived quality metrics. This does not assert every Track C child is Done; the epic criterion remains a Linear API check. All existing evidence entries for AC1 through AC5 are preserved."
+    source: "manual"
+    binds_ac: ["AC4"]
+    checks:
+      - check_type: "test_passes"
+        check_value: "uv run --frozen pytest tests/nodes/node_exemplar_registry_reconcile_compute/test_sampled_revocation.py -q -n 0 -p no:randomly --no-cov"
+        cwd: "${OMNI_HOME}/omnimarket"
+  - id: "dod-omn21090-deploy-sampled-exemplar-revocation"
+    description: "Hosted-runner probe executes the packaged exemplar reconcile handler and in-memory runtime against count-based revocation controls; a head without the change fails. No runtime host or credentials are used."
+    source: "manual"
+    checks:
+      - check_type: "command"
+        check_value: "uv run --frozen pytest tests/nodes/node_exemplar_registry_reconcile_compute/test_sampled_revocation.py -q -n 0 -p no:randomly --no-cov"
+        cwd: "${OMNI_HOME}/omnimarket"
diff --git a/contracts/OMN-21128.yaml b/contracts/OMN-21128.yaml
index 0d965e11ae..028d472fa9 100644
--- a/contracts/OMN-21128.yaml
+++ b/contracts/OMN-21128.yaml
@@ -15,6 +15,12 @@ is_seam_ticket: false
 interface_change: false
 interfaces_touched: []
 requirements:
+  - id: "req-exemplar-sampled-revocation"
+    statement: "Revocation requires a declared minimum sample; precision and recall are derived from labelled confusion counts."
+    acceptance:
+      - id: "AC12"
+        statement: >-
+          AC12: an exemplar is not revoked on a single failure with no prior runs (a minimum sample is required), and precision and recall come from counts, not constants -- falsifier: a one-run revocation, or a recall figure that does not move with the counts.
   - id: "req-exemplar-admission-refusals"
     statement: "A candidate is refused, with every reason listed, when it is in its repository's baseline, comes from a repository where the gate does not run, fails the gate, or is one of the shape-valid negatives."
     acceptance:
@@ -92,6 +98,21 @@ emergency_bypass:
   justification: ""
   follow_up_ticket_id: ""
 dod_evidence:
+  - id: "dod-omn21128-ac12-sampled-revocation"
+    description: "A first failure without history is deprecated rather than irrevocably revoked. Precision and recall derive from strict nonnegative confusion counts; each quality threshold requires enough denominator samples. Typed dispatch, RuntimeLocal in-memory dispatch and the packaged onex node CLI run the same cases. This slice does not complete the admission orchestrator or registry writer."
+    source: "manual"
+    binds_ac: ["AC12"]
+    checks:
+      - check_type: "test_passes"
+        check_value: "uv run --frozen pytest tests/nodes/node_exemplar_registry_reconcile_compute/test_sampled_revocation.py -q -n 0 -p no:randomly --no-cov"
+        cwd: "${OMNI_HOME}/omnimarket"
+  - id: "dod-omn21128-deploy-sampled-revocation"
+    description: "Hosted-runner packaged-handler and in-memory-runtime probe checks the sampled revocation behavior, strict evidence validation and checked-in contract/overlay parity. A head without this implementation fails these tests. No live services or credentials are required."
+    source: "manual"
+    checks:
+      - check_type: "command"
+        check_value: "uv run --frozen pytest tests/nodes/node_exemplar_registry_reconcile_compute/test_sampled_revocation.py -q -n 0 -p no:randomly --no-cov"
+        cwd: "${OMNI_HOME}/omnimarket"
   - id: "dod-omn21128-refusals"
     description: "Baseline, gate-absent, gate-failed and shape-valid negative candidates are refused with every reason."
     source: "manual"
diff --git a/src/omnimarket/nodes/node_exemplar_registry_reconcile_compute/contract.yaml b/src/omnimarket/nodes/node_exemplar_registry_reconcile_compute/contract.yaml
index 812fc8c432..6e7e2bc1a0 100644
--- a/src/omnimarket/nodes/node_exemplar_registry_reconcile_compute/contract.yaml
+++ b/src/omnimarket/nodes/node_exemplar_registry_reconcile_compute/contract.yaml
@@ -3,8 +3,8 @@
 name: node_exemplar_registry_reconcile_compute
 node_type: compute
 runtime_profiles: [main]
-contract_version: {major: 1, minor: 0, patch: 0}
-node_version: "1.0.0"
+contract_version: {major: 1, minor: 1, patch: 0}
+node_version: "1.1.0"
 description: >-
   Pure TC.5 exemplar registry reconciliation (OMN-21128). Plans deterministic
   create, revise, retain, deprecate and revoke operations from caller-supplied
@@ -39,12 +39,22 @@ inputs:
   content_sha256: {type: object, required: true, description: Content digests keyed by admitted node ID.}
   current_rows: {type: list, required: false, description: Current registry rows for the repository.}
   policy: {type: object, required: false, description: Versioned removal policy.}
+  evaluation_counts: {type: object, required: false, description: Labelled confusion counts by node ID; missing counts mean unknown evidence.}
+configuration:
+  removal_policy:
+    model: omnimarket.nodes.node_exemplar_registry_reconcile_compute.models.model_exemplar_registry_reconcile.ModelExemplarRemovalPolicy
+    policy_version: exemplar-registry-removal.v2
+    minimum_sample_size: 3
+    minimum_precision: 0.8
+    minimum_recall: 0.8
+    revoke_on_codes: [gate_failed, deny_secret_literal, deny_eval_exec, deny_pickle]
 outputs:
   operations: {type: list}
   counts: {type: object}
   current_count: {type: int}
   admitted_count: {type: int}
   reconciliation_hash: {type: str}
+  policy: {type: object}
 terminal_event: onex.evt.omnimarket.exemplar-registry-reconciled.v1
 runtime_dispatch:
   command_topic: onex.cmd.omnimarket.exemplar-registry-reconcile-requested.v1
diff --git a/src/omnimarket/nodes/node_exemplar_registry_reconcile_compute/handlers/handler_exemplar_registry_reconcile.py b/src/omnimarket/nodes/node_exemplar_registry_reconcile_compute/handlers/handler_exemplar_registry_reconcile.py
index 1ad6775c72..a643179a77 100644
--- a/src/omnimarket/nodes/node_exemplar_registry_reconcile_compute/handlers/handler_exemplar_registry_reconcile.py
+++ b/src/omnimarket/nodes/node_exemplar_registry_reconcile_compute/handlers/handler_exemplar_registry_reconcile.py
@@ -11,6 +11,7 @@
     ModelAdmissionDecision,
 )
 from omnimarket.nodes.node_exemplar_registry_reconcile_compute.models.model_exemplar_registry_reconcile import (
+    ModelExemplarEvaluationCounts,
     ModelExemplarRegistryOperation,
     ModelExemplarRegistryReconcileRequest,
     ModelExemplarRegistryReconcileResult,
@@ -80,6 +81,25 @@ def _operation(
         prior_digest = current.version_digest
         version_digest = current.version_digest
         operation, lifecycle_after = "retain", current.lifecycle
+        evaluation = request.evaluation_counts.get(
+            node_id, ModelExemplarEvaluationCounts()
+        )
+        enough_samples = evaluation.sample_size >= request.policy.minimum_sample_size
+        quality_reasons = []
+        if (
+            evaluation.true_positive + evaluation.false_positive
+            >= request.policy.minimum_sample_size
+            and evaluation.precision is not None
+            and evaluation.precision < request.policy.minimum_precision
+        ):
+            quality_reasons.append("precision_below_threshold")
+        if (
+            evaluation.true_positive + evaluation.false_negative
+            >= request.policy.minimum_sample_size
+            and evaluation.recall is not None
+            and evaluation.recall < request.policy.minimum_recall
+        ):
+            quality_reasons.append("recall_below_threshold")
         if admitted:
             if (
                 current.lifecycle == "deprecated"
@@ -88,10 +108,20 @@ def _operation(
                 operation, lifecycle_after = "revise", "active"
                 version_digest = new_digest
         elif current.lifecycle != "revoked":
-            if any(code in request.policy.revoke_on_codes for code in reason_codes):
+            if enough_samples and any(
+                code in request.policy.revoke_on_codes for code in reason_codes
+            ):
                 operation, lifecycle_after = "revoke", "revoked"
             elif current.lifecycle == "active":
                 operation, lifecycle_after = "deprecate", "deprecated"
+        if current.lifecycle != "revoked":
+            if enough_samples and quality_reasons:
+                operation, lifecycle_after = "revoke", "revoked"
+                reason_codes.extend(quality_reasons)
+            elif not enough_samples and any(
+                code in request.policy.revoke_on_codes for code in reason_codes
+            ):
+                reason_codes.append("insufficient_evaluation_sample")
 
     return ModelExemplarRegistryOperation(
         operation_id=_hash(
@@ -101,6 +131,12 @@ def _operation(
                 "logical_source_id": logical_source_id,
                 "operation": operation,
                 "version_digest": version_digest,
+                "lifecycle_after": lifecycle_after,
+                "reason_codes": reason_codes,
+                "policy": request.policy.model_dump(mode="json"),
+                "evaluation": request.evaluation_counts.get(
+                    node_id, ModelExemplarEvaluationCounts()
+                ).model_dump(mode="json"),
             }
         ),
         logical_source_id=logical_source_id,
@@ -110,6 +146,9 @@ def _operation(
         prior_version_digest=prior_digest,
         version_digest=version_digest,
         reason_codes=reason_codes,
+        evaluation=request.evaluation_counts.get(
+            node_id, ModelExemplarEvaluationCounts()
+        ),
         receipt=decision.receipt
         if decision is not None and admitted and lifecycle_after == "active"
         else None,
@@ -159,6 +198,7 @@ def handle(
             repository=request.repository,
             commit=request.admission.commit,
             policy_version=request.policy.policy_version,
+            policy=request.policy,
             operations=operations,
             counts=counts,
             current_count=len(request.current_rows),
@@ -168,6 +208,7 @@ def handle(
                     "repository": request.repository,
                     "commit": request.admission.commit,
                     "policy_version": request.policy.policy_version,
+                    "policy": request.policy.model_dump(mode="json"),
                     "operations": [op.model_dump(mode="json") for op in operations],
                 }
             ),
diff --git a/src/omnimarket/nodes/node_exemplar_registry_reconcile_compute/metadata.yaml b/src/omnimarket/nodes/node_exemplar_registry_reconcile_compute/metadata.yaml
index f73336ecba..b0e8844242 100644
--- a/src/omnimarket/nodes/node_exemplar_registry_reconcile_compute/metadata.yaml
+++ b/src/omnimarket/nodes/node_exemplar_registry_reconcile_compute/metadata.yaml
@@ -1,8 +1,8 @@
 # SPDX-FileCopyrightText: 2026 OmniNode.ai Inc.
 # SPDX-License-Identifier: MIT
 name: node_exemplar_registry_reconcile_compute
-version: "1.0.0"
-description: Pure TC.5 exemplar registry reconciliation and lifecycle planning
+version: "1.1.0"
+description: Pure TC.5 exemplar registry reconciliation with sampled quality evidence
 omnibase_core_compat: ">=0.47.28,<0.48.0"
 entry_points:
   onex.nodes:
diff --git a/src/omnimarket/nodes/node_exemplar_registry_reconcile_compute/models/model_exemplar_registry_reconcile.py b/src/omnimarket/nodes/node_exemplar_registry_reconcile_compute/models/model_exemplar_registry_reconcile.py
index ce8e968a0e..02b6cbbcce 100644
--- a/src/omnimarket/nodes/node_exemplar_registry_reconcile_compute/models/model_exemplar_registry_reconcile.py
+++ b/src/omnimarket/nodes/node_exemplar_registry_reconcile_compute/models/model_exemplar_registry_reconcile.py
@@ -36,7 +36,10 @@ class ModelExemplarRemovalPolicy(BaseModel):
 
     model_config = ConfigDict(frozen=True, extra="forbid")
 
-    policy_version: str = "exemplar-registry-removal.v1"
+    policy_version: str = "exemplar-registry-removal.v2"
+    minimum_sample_size: int = Field(default=3, ge=2, strict=True)
+    minimum_precision: float = Field(default=0.8, ge=0.0, le=1.0)
+    minimum_recall: float = Field(default=0.8, ge=0.0, le=1.0)
     revoke_on_codes: tuple[ViolationCode, ...] = (
         "gate_failed",
         "deny_secret_literal",
@@ -45,6 +48,39 @@ class ModelExemplarRemovalPolicy(BaseModel):
     )
 
 
+class ModelExemplarEvaluationCounts(BaseModel):
+    """Observed confusion counts; undefined ratios remain unknown."""
+
+    model_config = ConfigDict(frozen=True, extra="forbid")
+
+    true_positive: int = Field(default=0, ge=0, strict=True)
+    false_positive: int = Field(default=0, ge=0, strict=True)
+    true_negative: int = Field(default=0, ge=0, strict=True)
+    false_negative: int = Field(default=0, ge=0, strict=True)
+
+    @property
+    def sample_size(self) -> int:
+        """Every observed labelled outcome contributes one sample."""
+        return (
+            self.true_positive
+            + self.false_positive
+            + self.true_negative
+            + self.false_negative
+        )
+
+    @property
+    def precision(self) -> float | None:
+        """Share of predicted positives that were correct."""
+        denominator = self.true_positive + self.false_positive
+        return self.true_positive / denominator if denominator else None
+
+    @property
+    def recall(self) -> float | None:
+        """Share of actual positives recovered by the exemplar."""
+        denominator = self.true_positive + self.false_negative
+        return self.true_positive / denominator if denominator else None
+
+
 class ModelExemplarRegistryReconcileRequest(BaseModel):
     """Admission evidence and current registry rows supplied by the caller."""
 
@@ -54,6 +90,9 @@ class ModelExemplarRegistryReconcileRequest(BaseModel):
     admission: ModelExemplarAdmissionResult
     content_sha256: dict[str, str]
     current_rows: list[ModelExemplarRegistryRow] = Field(default_factory=list)
+    evaluation_counts: dict[str, ModelExemplarEvaluationCounts] = Field(
+        default_factory=dict
+    )
     policy: ModelExemplarRemovalPolicy = Field(
         default_factory=ModelExemplarRemovalPolicy
     )
@@ -82,6 +121,9 @@ class ModelExemplarRegistryOperation(BaseModel):
     version_digest: str | None
     reason_codes: list[str]
     receipt: ModelGateReceipt | None
+    evaluation: ModelExemplarEvaluationCounts = Field(
+        default_factory=ModelExemplarEvaluationCounts
+    )
 
 
 class ModelExemplarRegistryReconcileResult(BaseModel):
@@ -92,6 +134,9 @@ class ModelExemplarRegistryReconcileResult(BaseModel):
     repository: str
     commit: str
     policy_version: str
+    policy: ModelExemplarRemovalPolicy = Field(
+        default_factory=ModelExemplarRemovalPolicy
+    )
     operations: list[ModelExemplarRegistryOperation]
     counts: dict[str, int]
     current_count: int
diff --git a/src/omnimarket/nodes/node_exemplar_registry_reconcile_compute/overlay.yaml b/src/omnimarket/nodes/node_exemplar_registry_reconcile_compute/overlay.yaml
new file mode 100644
index 0000000000..401d798fae
--- /dev/null
+++ b/src/omnimarket/nodes/node_exemplar_registry_reconcile_compute/overlay.yaml
@@ -0,0 +1,10 @@
+# SPDX-FileCopyrightText: 2026 OmniNode.ai Inc.
+# SPDX-License-Identifier: MIT
+# Input policy injected by the caller; this file does not install a live binding.
+node: node_exemplar_registry_reconcile_compute
+policy:
+  policy_version: exemplar-registry-removal.v2
+  minimum_sample_size: 3
+  minimum_precision: 0.8
+  minimum_recall: 0.8
+  revoke_on_codes: [gate_failed, deny_secret_literal, deny_eval_exec, deny_pickle]
diff --git a/tests/nodes/node_exemplar_registry_reconcile_compute/test_golden_chain_exemplar_registry_reconcile_compute.py b/tests/nodes/node_exemplar_registry_reconcile_compute/test_golden_chain_exemplar_registry_reconcile_compute.py
index b0602164ac..126c13986d 100644
--- a/tests/nodes/node_exemplar_registry_reconcile_compute/test_golden_chain_exemplar_registry_reconcile_compute.py
+++ b/tests/nodes/node_exemplar_registry_reconcile_compute/test_golden_chain_exemplar_registry_reconcile_compute.py
@@ -248,7 +248,11 @@ def test_refusal_with_revoke_code_revokes(code: str) -> None:
     for lifecycle in ("active", "deprecated"):
         row = current_row("node_a", lifecycle=lifecycle)
         refused = decision("node_a", admitted=False, codes=("in_baseline", code))
-        [op] = reconcile(request(refused, rows=[row])).operations
+        [op] = reconcile(
+            request(
+                refused, rows=[row], evaluation_counts={"node_a": {"true_positive": 3}}
+            )
+        ).operations
         assert (op.operation, op.lifecycle_after) == ("revoke", "revoked")
         assert op.reason_codes == ["in_baseline", code]
 
@@ -257,6 +261,7 @@ def test_revoke_codes_are_declared_policy_data() -> None:
     row = current_row("node_a")
     refused = decision("node_a", admitted=False, codes=("in_baseline",))
     payload = request(refused, rows=[row])
+    payload["evaluation_counts"] = {"node_a": {"true_positive": 3}}
     payload["policy"] = {"policy_version": "p.test", "revoke_on_codes": ["in_baseline"]}
     result = reconcile(payload)
     assert ops(result) == {"node_a": "revoke"}
@@ -305,6 +310,7 @@ def test_plan_is_ordered_by_logical_source_and_counts_bind_it() -> None:
         decision("node_e", admitted=False, codes=("deny_pickle",)),
         decision("node_a"),
         rows=rows,
+        evaluation_counts={"node_e": {"true_positive": 3}},
     )
     result = reconcile(payload)
     assert [op.node_id for op in result.operations] == [
diff --git a/tests/nodes/node_exemplar_registry_reconcile_compute/test_sampled_revocation.py b/tests/nodes/node_exemplar_registry_reconcile_compute/test_sampled_revocation.py
new file mode 100644
index 0000000000..22a2826043
--- /dev/null
+++ b/tests/nodes/node_exemplar_registry_reconcile_compute/test_sampled_revocation.py
@@ -0,0 +1,181 @@
+# SPDX-FileCopyrightText: 2026 OmniNode.ai Inc.
+# SPDX-License-Identifier: MIT
+"""TC.5 AC12: count-derived metrics and minimum-sample revocation."""
+
+from __future__ import annotations
+
+import json
+import subprocess
+import sys
+from pathlib import Path
+from typing import Any
+
+import pytest
+import yaml
+from pydantic import ValidationError
+
+from omnimarket.nodes.node_exemplar_registry_reconcile_compute.models.model_exemplar_registry_reconcile import (
+    ModelExemplarRegistryReconcileResult,
+)
+from tests.nodes.node_exemplar_registry_reconcile_compute.test_golden_chain_exemplar_registry_reconcile_compute import (
+    NODE_DIR,
+    _run,
+    current_row,
+    decision,
+    reconcile,
+    request,
+)
+
+pytestmark = pytest.mark.unit
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("binding", ["typed", "inmemory", "cli"])
+@pytest.mark.parametrize(
+    ("counts", "admitted", "expected", "precision", "recall"),
+    [
+        ({}, False, "deprecate", None, None),
+        ({"false_positive": 1}, False, "deprecate", 0.0, None),
+        ({"false_positive": 3}, False, "revoke", 0.0, None),
+        ({"true_positive": 9, "false_positive": 1}, True, "retain", 0.9, 1.0),
+        ({"true_positive": 1, "false_positive": 3}, True, "revoke", 0.25, 1.0),
+        ({"true_positive": 1, "false_negative": 3}, True, "revoke", 1.0, 0.25),
+        ({"true_negative": 10}, True, "retain", None, None),
+        ({"true_negative": 10, "false_positive": 1}, True, "retain", 0.0, None),
+        ({"true_negative": 10, "false_negative": 1}, True, "retain", None, 0.0),
+    ],
+)
+async def test_revocation_uses_counts_on_each_dispatch_binding(
+    tmp_path: Path,
+    binding: str,
+    counts: dict[str, int],
+    admitted: bool,
+    expected: str,
+    precision: float | None,
+    recall: float | None,
+) -> None:
+    candidate = decision(
+        "node_a", admitted=admitted, codes=() if admitted else ("gate_failed",)
+    )
+    payload = request(candidate, rows=[current_row("node_a")])
+    payload["evaluation_counts"] = {"node_a": counts}
+    if binding == "typed":
+        result = reconcile(payload)
+    elif binding == "inmemory":
+        result = (await _run(tmp_path, payload)).handler_result
+    else:
+        input_path = tmp_path / "input.json"
+        input_path.write_text(json.dumps(payload))
+        state = tmp_path / "cli-state"
+        completed = subprocess.run(
+            [
+                str(Path(sys.executable).parent / "onex"),
+                "node",
+                "node_exemplar_registry_reconcile_compute",
+                "--input",
+                str(input_path),
+                "--state-root",
+                str(state),
+                "--backend",
+                "event_bus=inmemory",
+                "--timeout",
+                "10",
+            ],
+            capture_output=True,
+            text=True,
+            timeout=30,
+            check=False,
+        )
+        assert completed.returncode == 0, completed.stdout + completed.stderr
+        saved = json.loads((state / "workflow_result.json").read_text())
+        assert saved["handler_locus"] == "in_process"
+        result = ModelExemplarRegistryReconcileResult.model_validate(
+            saved["handler_result"]
+        )
+    assert result == reconcile(payload)
+    [operation] = result.operations
+    assert operation.operation == expected
+    assert operation.evaluation.sample_size == sum(counts.values())
+    assert operation.evaluation.precision == precision
+    assert operation.evaluation.recall == recall
+    if expected == "revoke":
+        assert operation.lifecycle_after == "revoked"
+        assert operation.receipt is None
+
+
+def test_missing_history_does_not_revoke_a_first_failure() -> None:
+    payload = request(
+        decision("node_a", admitted=False, codes=("gate_failed",)),
+        rows=[current_row("node_a")],
+    )
+    [operation] = reconcile(payload).operations
+    assert operation.operation == "deprecate"
+    assert operation.evaluation.sample_size == 0
+    assert "insufficient_evaluation_sample" in operation.reason_codes
+
+
+def test_overlay_policy_controls_minimum_sample_and_thresholds() -> None:
+    payload = request(decision("node_a"), rows=[current_row("node_a")])
+    payload["evaluation_counts"] = {"node_a": {"true_positive": 1, "false_positive": 3}}
+    payload["policy"] = {
+        "minimum_sample_size": 5,
+        "minimum_precision": 0.8,
+        "minimum_recall": 0.8,
+    }
+    [operation] = reconcile(payload).operations
+    assert operation.operation == "retain"
+    payload["policy"]["minimum_sample_size"] = 4
+    [operation] = reconcile(payload).operations
+    assert operation.operation == "revoke"
+    assert "precision_below_threshold" in operation.reason_codes
+    payload["policy"]["minimum_precision"] = 0.25
+    [operation] = reconcile(payload).operations
+    assert operation.operation == "retain"
+
+
+@pytest.mark.parametrize(
+    "counts",
+    [
+        {"false_positive": -1},
+        {"true_positive": True},
+        {"true_positive": 1.5},
+        {"recall": 0.5},
+    ],
+)
+def test_evidence_refuses_negative_noninteger_or_asserted_metrics(
+    counts: dict[str, Any],
+) -> None:
+    payload = request(decision("node_a"), rows=[current_row("node_a")])
+    payload["evaluation_counts"] = {"node_a": counts}
+    with pytest.raises(ValidationError):
+        reconcile(payload)
+
+
+@pytest.mark.parametrize(
+    "policy",
+    [{"minimum_sample_size": 1}, {"minimum_precision": 1.1}, {"minimum_recall": -0.1}],
+)
+def test_invalid_policy_is_refused(policy: dict[str, Any]) -> None:
+    with pytest.raises(ValidationError):
+        reconcile(request(decision("node_a"), policy=policy))
+
+
+def test_evidence_and_policy_bind_the_reconciliation_hash() -> None:
+    payload = request(decision("node_a"), rows=[current_row("node_a")])
+    payload["evaluation_counts"] = {"node_a": {"true_positive": 4}}
+    first = reconcile(payload)
+    payload["evaluation_counts"]["node_a"]["true_positive"] = 5
+    second = reconcile(payload)
+    assert first.reconciliation_hash != second.reconciliation_hash
+    payload["policy"] = {"minimum_sample_size": 6}
+    assert reconcile(payload).reconciliation_hash != second.reconciliation_hash
+
+
+def test_contract_overlay_and_model_agree_on_policy() -> None:
+    contract = yaml.safe_load((NODE_DIR / "contract.yaml").read_text())
+    overlay = yaml.safe_load((NODE_DIR / "overlay.yaml").read_text())
+    declared = dict(contract["configuration"]["removal_policy"])
+    declared.pop("model")
+    assert declared == overlay["policy"]
+    payload = request(decision("node_a"), policy=overlay["policy"])
+    assert reconcile(payload).policy.model_dump(mode="json") == declared

END PUBLIC SOURCE
