diff --git a/src/cots.ts b/src/cots.ts
index dee466a..9c70c3e 100644
--- a/src/cots.ts
+++ b/src/cots.ts
@@ -22,7 +22,7 @@ export interface Cots {
   inputs: { source_artifacts: SourceArtifact[] };
   context: Context;
   decisions: DecisionRecord[];
-  evidence: { evidence_chain: EvidenceChainEntry[] };
+  evidence: Evidence;
   validation: { feedback_surfaces: FeedbackSurface[] };
 }
 
@@ -104,6 +104,21 @@ export interface AuditEntry {
   notes: string;
 }
 
+export type SourceType = "human-authored" | "human-curated" | "model-generated" | "mixed";
+
+export type ChainTerminationPolicy = "terminal" | "bounded_recursion" | "unbounded_recursion";
+
+export interface ChainTermination {
+  policy: ChainTerminationPolicy;
+  max_levels?: number;
+}
+
+export interface Evidence {
+  /** RFC-0002 (COTS schema v1.2.0). Optional; the discipline's default is terminal. */
+  chain_termination?: ChainTermination;
+  evidence_chain: EvidenceChainEntry[];
+}
+
 export interface EvidenceChainEntry {
   decision_id: URI;
   supporting_evidence: EvidenceItem[];
@@ -127,6 +142,8 @@ export interface EvidenceItem {
   transformation: { operation: string; inputs: string[]; output: string; recorded_at: ISO8601 };
   uncertainty: { calibrated_score: CalibratedScore; basis: string };
   polarity: Polarity;
+  /** RFC-0002 (COTS schema v1.2.0). Optional; absent on pre-RFC-0002 COTSs. */
+  source_type?: SourceType;
 }
 
 export interface FeedbackSurface {
diff --git a/src/layer2/layer2.ts b/src/layer2/layer2.ts
index 79d2781..cf22f74 100644
--- a/src/layer2/layer2.ts
+++ b/src/layer2/layer2.ts
@@ -11,7 +11,7 @@
  * Layer 3, which is the layer that can reject.
  */
 
-import type { Cots, DecisionRecord, Context, ValidationRecord } from "../cots.js";
+import type { Cots, DecisionRecord, Context, ValidationRecord, DecisionOnValidation } from "../cots.js";
 
 export type FlagType =
   | "drift-detected"
@@ -101,6 +101,29 @@ const validateDecision = (
   };
 };
 
+const validateRecursiveChains = (cots: Cots, now: string): ValidationRecord["flags"] => {
+  const flags: ValidationRecord["flags"] = [];
+  const termination = cots.evidence.chain_termination;
+
+  for (const chain of cots.evidence.evidence_chain) {
+    const items = [...chain.supporting_evidence, ...chain.contradicting_evidence];
+    const recursive = items.filter((item) => item.source_type === "model-generated");
+    if (recursive.length > 0) {
+      const policy = termination?.policy ?? "terminal";
+      const admitted = policy === "bounded_recursion" || policy === "unbounded_recursion";
+      flags.push({
+        flag_type: "segment-signal-divergence",
+        message: admitted
+          ? `Recursive evidence chain: ${recursive.length} model-generated evidence item(s) under policy '${policy}'. The chain does not terminate at a human-readable source. (RFC-0002)`
+          : `Recursive evidence chain under terminal policy: ${recursive.length} model-generated evidence item(s). The chain must resolve to human-authored sources. (RFC-0002)`,
+        detected_at: now,
+      });
+    }
+  }
+
+  return flags;
+};
+
 export class Layer2 {
   constructor(private readonly calibration: CalibrationData) {}
 
@@ -118,9 +141,17 @@ export class Layer2 {
     const now = options.now ?? new Date().toISOString();
     const validatedBy = options.validatedBy ?? "layer2-runtime-validation";
 
+    const recursiveFlags = validateRecursiveChains(cots, now);
+
     const decisions = cots.decisions.map((decision) => {
       const record = validateDecision(decision, cots.context, this.calibration, now);
-      return { ...decision, validation: { ...record, validated_by: validatedBy } };
+      const flags = [...record.flags, ...recursiveFlags];
+      const decision_on_validation: DecisionOnValidation =
+        flags.length > 0 ? (record.decision_on_validation === "escalated" ? "escalated" : "flagged") : "passed";
+      return {
+        ...decision,
+        validation: { ...record, flags, decision_on_validation, validated_by: validatedBy },
+      };
     });
 
     return { ...cots, decisions };
diff --git a/tests/layer2.test.ts b/tests/layer2.test.ts
index 2596e64..2b4d530 100644
--- a/tests/layer2.test.ts
+++ b/tests/layer2.test.ts
@@ -100,3 +100,38 @@ describe("Layer 2 (Runtime Validation)", () => {
     expect(cots.decisions[0].validation).toBeUndefined();
   });
 });
+
+describe("Layer 2 recursive chain flagging (RFC-0002)", () => {
+  const buildRecursiveCots = async () => {
+    const cots = await buildCots();
+    cots.evidence.chain_termination = { policy: "bounded_recursion", max_levels: 1 };
+    cots.evidence.evidence_chain[0].supporting_evidence[0].source_type = "model-generated";
+    return cots;
+  };
+
+  it("flags a COTS with model-generated evidence (bounded policy)", async () => {
+    const cots = await buildRecursiveCots();
+    const layer2 = new Layer2({});
+    const validated = layer2.validate(cots, { now: NOW });
+    const v = validated.decisions[0].validation!;
+    expect(v.decision_on_validation).toBe("flagged");
+    expect(v.flags.some((f) => f.message.includes("Recursive evidence chain"))).toBe(true);
+  });
+
+  it("flags a model-generated item even under the terminal default", async () => {
+    const cots = await buildRecursiveCots();
+    cots.evidence.chain_termination = undefined;
+    const layer2 = new Layer2({});
+    const validated = layer2.validate(cots, { now: NOW });
+    expect(validated.decisions[0].validation!.flags.some((f) => f.message.includes("terminal policy"))).toBe(true);
+  });
+
+  it("does not flag a COTS whose evidence is human-authored", async () => {
+    const cots = await buildCots();
+    cots.evidence.chain_termination = { policy: "terminal" };
+    cots.evidence.evidence_chain[0].supporting_evidence[0].source_type = "human-authored";
+    const layer2 = new Layer2({});
+    const validated = layer2.validate(cots, { now: NOW });
+    expect(validated.decisions[0].validation!.decision_on_validation).toBe("passed");
+  });
+});