diff --git a/control_plane/tests/accuracy_evidence.rs b/control_plane/tests/accuracy_evidence.rs
new file mode 100644
index 00000000..2e6dca8e
--- /dev/null
+++ b/control_plane/tests/accuracy_evidence.rs
@@ -0,0 +1,107 @@
+//! Evidence applicability is independent of prices and observed resource use.
+#[path = "support/erp_contract.rs"]
+mod fixture;
+use control_plane::physical::erp::{ErpAccuracyMode, ErpParameterDecision, ErpPlanningInput};
+use planner_types::post_asap::{SketchAlgorithm, SketchParams};
+use serde_json::json;
+
+fn decision(policy: &ErpPlanningInput) -> ErpParameterDecision {
+    policy.select(
+        SketchAlgorithm::Cms,
+        0.01,
+        SketchParams::Cms {
+            width: 4096,
+            depth: 5,
+        },
+    )
+}
+
+/// Wrong distribution/implementation/metric, insufficient trials and excessive
+/// error reject empirical admission even when that record is nearly free.
+#[test]
+fn inapplicable_accuracy_cannot_be_rescued_by_a_low_price() {
+    assert!(matches!(
+        decision(&fixture::policy(ErpAccuracyMode::Empirical)),
+        ErpParameterDecision::Empirical { .. }
+    ));
+    for mismatch in [
+        "distribution",
+        "implementation",
+        "metric",
+        "trials",
+        "error",
+        "parameters",
+    ] {
+        let mut policy = fixture::policy(ErpAccuracyMode::Empirical);
+        let row = &mut policy.artifact.records[0];
+        row.resources.update_cpu_seconds = 1e-30;
+        row.resources.query_cpu_seconds = 1e-30;
+        match mismatch {
+            "distribution" => row.distribution = json!({"different_population":true}),
+            "implementation" => row.implementation = "different-implementation".into(),
+            "metric" => row.error_metrics.clear(),
+            "trials" => row.trials = 1,
+            "error" => {
+                row.error_metrics.insert("relative_error".into(), 0.2);
+            }
+            "parameters" => row.parameters = json!({"rows":0,"cols":0}),
+            _ => unreachable!(),
+        }
+        assert!(
+            matches!(
+                decision(&policy),
+                ErpParameterDecision::ExactFallback { .. }
+            ),
+            "{mismatch}"
+        );
+    }
+}
+
+/// Theoretical fallback retains its identity; it is never relabeled empirical.
+#[test]
+fn missing_evidence_preserves_the_declared_fallback_mode() {
+    for mode in [ErpAccuracyMode::Hybrid, ErpAccuracyMode::Empirical] {
+        let mut policy = fixture::policy(mode);
+        policy.artifact.records.clear();
+        match (mode, decision(&policy)) {
+            (ErpAccuracyMode::Hybrid, ErpParameterDecision::TheoreticalFallback { params, .. }) => {
+                assert_eq!(
+                    params,
+                    SketchParams::Cms {
+                        width: 4096,
+                        depth: 5
+                    }
+                )
+            }
+            (ErpAccuracyMode::Empirical, ErpParameterDecision::ExactFallback { .. }) => {}
+            unexpected => panic!("wrong evidence/fallback provenance: {unexpected:?}"),
+        }
+    }
+}
+
+/// Both parameter choice and error must be traceable to the same admitted row.
+#[test]
+fn admitted_accuracy_retains_record_identity() {
+    let policy = fixture::policy(ErpAccuracyMode::Empirical);
+    let ErpParameterDecision::Empirical {
+        record_id,
+        observed_error,
+        params,
+        ..
+    } = decision(&policy)
+    else {
+        panic!("matching evidence should be admitted")
+    };
+    assert_eq!(record_id, policy.artifact.records[0].id);
+    assert_eq!(
+        observed_error,
+        policy.artifact.records[0].error_metrics["relative_error"]
+    );
+    assert_eq!(
+        params,
+        SketchParams::Cms {
+            width: 512,
+            depth: 3
+        }
+    );
+}
diff --git a/control_plane/tests/support/erp_contract.rs b/control_plane/tests/support/erp_contract.rs
new file mode 100644
index 00000000..704c4bb6
--- /dev/null
+++ b/control_plane/tests/support/erp_contract.rs
@@ -0,0 +1,44 @@
+//! Synthetic evidence for admission contracts, not a measured calibration run.
+use asap_aware_mapping::erp::{ErpArtifact, ErpRecord, ErpResourceProfile, ERP_SCHEMA_VERSION};
+use control_plane::physical::erp::{ErpAccuracyMode, ErpPlanningInput, ErpRuntimeCapabilities};
+use std::collections::BTreeMap;
+pub fn policy(mode: ErpAccuracyMode) -> ErpPlanningInput {
+    ErpPlanningInput {
+        artifact: ErpArtifact {
+            schema_version: ERP_SCHEMA_VERSION,
+            producer_version: "bench-rev".into(),
+            records: vec![ErpRecord {
+                id: "cms-512".into(),
+                sketch: "cms-fastpath-vector2d".into(),
+                implementation: "oxide".into(),
+                parameters: serde_json::json!({"rows": 3, "cols": 512}),
+                distribution: serde_json::json!({"synthetic":{"kind":"zipf","s":1.1}}),
+                trials: 20,
+                error_metrics: BTreeMap::from([("relative_error".into(), 0.009)]),
+                resources: ErpResourceProfile {
+                    memory_bytes: 12_288.0,
+                    update_cpu_seconds: 1e-7,
+                    merge_cpu_seconds: 1e-5,
+                    query_cpu_seconds: 1e-6,
+                },
+            }],
+        },
+        distribution: serde_json::json!({"synthetic":{"kind":"zipf","s":1.1}}),
+        implementation: Some("oxide".into()),
+        error_metric: "relative_error".into(),
+        min_trials: 10,
+        expected_updates: 1_000.0,
+        expected_queries: 100.0,
+        expected_merges: 0.0,
+        retention_seconds: 60.0,
+        cpu_weight: 1.0,
+        byte_second_weight: 1e-9,
+        mode,
+        observed_shape: None,
+        observed_populations: None,
+        resolved_data_descriptor: None,
+        observed_shape_source: None,
+        shape_match: None,
+        runtime: ErpRuntimeCapabilities::default(),
+    }
+}
diff --git a/docs/design_docs/accuracy-evidence-validation.md b/docs/design_docs/accuracy-evidence-validation.md
new file mode 100644
index 00000000..03d09ce2
--- /dev/null
+++ b/docs/design_docs/accuracy-evidence-validation.md
@@ -0,0 +1,14 @@
+# Accuracy evidence validation
+
+This PR isolates evidence applicability from ranking. The public ERP adapter is
+exercised with explicitly synthetic records in `accuracy_evidence.rs`: mismatched
+distribution, implementation, metric, parameters, too few trials and excessive
+error cannot admit a cheap candidate. Successful decisions retain record identity;
+Hybrid theoretical fallback is distinguishable from empirical admission.
+
+These tests validate the decision contract, not the quality of a real benchmark.
+Level 3 needs applicable measured sketch errors, query-specific accuracy checks
+against exact results, and provenance for the implementation/data/parameters.
+Observed mean error alone cannot establish a requested tail-probability guarantee.
+Existing readout-specific and online evidence freshness checks remain in `erp.rs`;
+this PR neither replaces those checks nor adds a second semantic matcher.