diff --git a/config/evolution.yaml b/config/evolution.yaml
index dcb325b..89cc4fd 100644
--- a/config/evolution.yaml
+++ b/config/evolution.yaml
@@ -28,6 +28,17 @@ reflection:
   effort: "high"
   max_budget_usd: 0.50
 
+# LLM Judge Configuration
+judges:
+  # "auto" enables when ANTHROPIC_API_KEY is available
+  # "always" enables unconditionally
+  # "never" disables unconditionally
+  enabled: "auto"
+  # Safety net against runaway costs (daily reset)
+  cost_cap_usd_per_day: 50.0
+  # Maximum golden suite entries (prune oldest when exceeded)
+  max_golden_suite_size: 50
+
 # Directory paths (relative to project root)
 paths:
   config_dir: "phantom-config"
diff --git a/src/evolution/__tests__/application.test.ts b/src/evolution/__tests__/application.test.ts
index b1b33b0..9ac42b9 100644
--- a/src/evolution/__tests__/application.test.ts
+++ b/src/evolution/__tests__/application.test.ts
@@ -12,6 +12,7 @@ function testConfig(): EvolutionConfig {
 		cadence: { reflection_interval: 1, consolidation_interval: 10, full_review_interval: 50, drift_check_interval: 20 },
 		gates: { drift_threshold: 0.7, max_file_lines: 200, auto_rollback_threshold: 0.1, auto_rollback_window: 5 },
 		reflection: { model: "claude-sonnet-4-20250514", effort: "high", max_budget_usd: 0.5 },
+		judges: { enabled: "auto", cost_cap_usd_per_day: 50.0, max_golden_suite_size: 50 },
 		paths: {
 			config_dir: TEST_DIR,
 			constitution: `${TEST_DIR}/constitution.md`,
diff --git a/src/evolution/__tests__/constitution.test.ts b/src/evolution/__tests__/constitution.test.ts
index 3e5c0a5..1cdfc71 100644
--- a/src/evolution/__tests__/constitution.test.ts
+++ b/src/evolution/__tests__/constitution.test.ts
@@ -11,6 +11,7 @@ function testConfig(): EvolutionConfig {
 		cadence: { reflection_interval: 1, consolidation_interval: 10, full_review_interval: 50, drift_check_interval: 20 },
 		gates: { drift_threshold: 0.7, max_file_lines: 200, auto_rollback_threshold: 0.1, auto_rollback_window: 5 },
 		reflection: { model: "claude-sonnet-4-20250514", effort: "high", max_budget_usd: 0.5 },
+		judges: { enabled: "auto", cost_cap_usd_per_day: 50.0, max_golden_suite_size: 50 },
 		paths: {
 			config_dir: TEST_DIR,
 			constitution: `${TEST_DIR}/constitution.md`,
diff --git a/src/evolution/__tests__/cost-cap.test.ts b/src/evolution/__tests__/cost-cap.test.ts
new file mode 100644
index 0000000..24135db
--- /dev/null
+++ b/src/evolution/__tests__/cost-cap.test.ts
@@ -0,0 +1,199 @@
+import { afterEach, beforeEach, describe, expect, test } from "bun:test";
+import { mkdirSync, readFileSync, rmSync, writeFileSync } from "node:fs";
+import { EvolutionEngine } from "../engine.ts";
+import type { SessionSummary } from "../types.ts";
+
+const TEST_DIR = "/tmp/phantom-test-cost-cap";
+const CONFIG_PATH = `${TEST_DIR}/config/evolution.yaml`;
+
+let savedApiKey: string | undefined;
+
+function setupTestEnv(costCap: number): void {
+	mkdirSync(`${TEST_DIR}/config`, { recursive: true });
+	mkdirSync(`${TEST_DIR}/phantom-config/meta`, { recursive: true });
+	mkdirSync(`${TEST_DIR}/phantom-config/strategies`, { recursive: true });
+	mkdirSync(`${TEST_DIR}/phantom-config/memory`, { recursive: true });
+
+	writeFileSync(
+		CONFIG_PATH,
+		[
+			"cadence:",
+			"  reflection_interval: 1",
+			"  consolidation_interval: 10",
+			"gates:",
+			"  drift_threshold: 0.7",
+			"  max_file_lines: 200",
+			"  auto_rollback_threshold: 0.1",
+			"  auto_rollback_window: 5",
+			"judges:",
+			'  enabled: "never"',
+			`  cost_cap_usd_per_day: ${costCap}`,
+			"  max_golden_suite_size: 50",
+			"paths:",
+			`  config_dir: "${TEST_DIR}/phantom-config"`,
+			`  constitution: "${TEST_DIR}/phantom-config/constitution.md"`,
+			`  version_file: "${TEST_DIR}/phantom-config/meta/version.json"`,
+			`  metrics_file: "${TEST_DIR}/phantom-config/meta/metrics.json"`,
+			`  evolution_log: "${TEST_DIR}/phantom-config/meta/evolution-log.jsonl"`,
+			`  golden_suite: "${TEST_DIR}/phantom-config/meta/golden-suite.jsonl"`,
+			`  session_log: "${TEST_DIR}/phantom-config/memory/session-log.jsonl"`,
+		].join("\n"),
+		"utf-8",
+	);
+
+	writeFileSync(`${TEST_DIR}/phantom-config/constitution.md`, "# Constitution\n1. Be honest.\n", "utf-8");
+	writeFileSync(`${TEST_DIR}/phantom-config/persona.md`, "", "utf-8");
+	writeFileSync(`${TEST_DIR}/phantom-config/user-profile.md`, "# User Profile\n", "utf-8");
+	writeFileSync(`${TEST_DIR}/phantom-config/domain-knowledge.md`, "", "utf-8");
+	writeFileSync(`${TEST_DIR}/phantom-config/strategies/task-patterns.md`, "", "utf-8");
+	writeFileSync(`${TEST_DIR}/phantom-config/strategies/tool-preferences.md`, "", "utf-8");
+	writeFileSync(`${TEST_DIR}/phantom-config/strategies/error-recovery.md`, "", "utf-8");
+	writeFileSync(`${TEST_DIR}/phantom-config/memory/session-log.jsonl`, "", "utf-8");
+	writeFileSync(
+		`${TEST_DIR}/phantom-config/meta/version.json`,
+		JSON.stringify({
+			version: 0,
+			parent: null,
+			timestamp: new Date().toISOString(),
+			changes: [],
+			metrics_at_change: { session_count: 0, success_rate_7d: 0, correction_rate_7d: 0 },
+		}),
+		"utf-8",
+	);
+	writeFileSync(
+		`${TEST_DIR}/phantom-config/meta/metrics.json`,
+		JSON.stringify({
+			session_count: 0,
+			success_count: 0,
+			failure_count: 0,
+			correction_count: 0,
+			evolution_count: 0,
+			rollback_count: 0,
+			last_session_at: null,
+			last_evolution_at: null,
+			success_rate_7d: 0,
+			correction_rate_7d: 0,
+			sessions_since_consolidation: 0,
+		}),
+		"utf-8",
+	);
+	writeFileSync(`${TEST_DIR}/phantom-config/meta/evolution-log.jsonl`, "", "utf-8");
+	writeFileSync(`${TEST_DIR}/phantom-config/meta/golden-suite.jsonl`, "", "utf-8");
+}
+
+function makeSession(overrides: Partial<SessionSummary> = {}): SessionSummary {
+	return {
+		session_id: `session-${Date.now()}`,
+		session_key: "cli:main",
+		user_id: "user-1",
+		user_messages: ["No, use TypeScript not JavaScript"],
+		assistant_messages: ["Got it."],
+		tools_used: [],
+		files_tracked: [],
+		outcome: "success",
+		cost_usd: 0.05,
+		started_at: "2026-03-25T10:00:00Z",
+		ended_at: "2026-03-25T10:05:00Z",
+		...overrides,
+	};
+}
+
+describe("Cost Cap", () => {
+	beforeEach(() => {
+		savedApiKey = process.env.ANTHROPIC_API_KEY;
+	});
+
+	afterEach(() => {
+		if (savedApiKey !== undefined) {
+			process.env.ANTHROPIC_API_KEY = savedApiKey;
+		} else {
+			process.env.ANTHROPIC_API_KEY = undefined;
+		}
+		rmSync(TEST_DIR, { recursive: true, force: true });
+	});
+
+	test("cost cap config is parsed from YAML", () => {
+		setupTestEnv(10.0);
+		const engine = new EvolutionEngine(CONFIG_PATH);
+		const config = engine.getEvolutionConfig();
+		expect(config.judges.cost_cap_usd_per_day).toBe(10.0);
+	});
+
+	test("cost cap defaults to 50 when not configured", () => {
+		mkdirSync(`${TEST_DIR}/config`, { recursive: true });
+		mkdirSync(`${TEST_DIR}/phantom-config/meta`, { recursive: true });
+		mkdirSync(`${TEST_DIR}/phantom-config/strategies`, { recursive: true });
+		mkdirSync(`${TEST_DIR}/phantom-config/memory`, { recursive: true });
+
+		writeFileSync(
+			CONFIG_PATH,
+			[
+				"paths:",
+				`  config_dir: "${TEST_DIR}/phantom-config"`,
+				`  constitution: "${TEST_DIR}/phantom-config/constitution.md"`,
+				`  version_file: "${TEST_DIR}/phantom-config/meta/version.json"`,
+				`  metrics_file: "${TEST_DIR}/phantom-config/meta/metrics.json"`,
+				`  evolution_log: "${TEST_DIR}/phantom-config/meta/evolution-log.jsonl"`,
+				`  golden_suite: "${TEST_DIR}/phantom-config/meta/golden-suite.jsonl"`,
+				`  session_log: "${TEST_DIR}/phantom-config/memory/session-log.jsonl"`,
+			].join("\n"),
+			"utf-8",
+		);
+		writeFileSync(`${TEST_DIR}/phantom-config/constitution.md`, "# Constitution\n", "utf-8");
+		writeFileSync(`${TEST_DIR}/phantom-config/persona.md`, "", "utf-8");
+		writeFileSync(`${TEST_DIR}/phantom-config/user-profile.md`, "", "utf-8");
+		writeFileSync(`${TEST_DIR}/phantom-config/domain-knowledge.md`, "", "utf-8");
+		writeFileSync(`${TEST_DIR}/phantom-config/strategies/task-patterns.md`, "", "utf-8");
+		writeFileSync(`${TEST_DIR}/phantom-config/strategies/tool-preferences.md`, "", "utf-8");
+		writeFileSync(`${TEST_DIR}/phantom-config/strategies/error-recovery.md`, "", "utf-8");
+		writeFileSync(`${TEST_DIR}/phantom-config/memory/session-log.jsonl`, "", "utf-8");
+		writeFileSync(
+			`${TEST_DIR}/phantom-config/meta/version.json`,
+			JSON.stringify({
+				version: 0,
+				parent: null,
+				timestamp: new Date().toISOString(),
+				changes: [],
+				metrics_at_change: { session_count: 0, success_rate_7d: 0, correction_rate_7d: 0 },
+			}),
+			"utf-8",
+		);
+		writeFileSync(
+			`${TEST_DIR}/phantom-config/meta/metrics.json`,
+			JSON.stringify({
+				session_count: 0,
+				success_count: 0,
+				failure_count: 0,
+				correction_count: 0,
+				evolution_count: 0,
+				rollback_count: 0,
+				last_session_at: null,
+				last_evolution_at: null,
+				success_rate_7d: 0,
+				correction_rate_7d: 0,
+				sessions_since_consolidation: 0,
+			}),
+			"utf-8",
+		);
+		writeFileSync(`${TEST_DIR}/phantom-config/meta/evolution-log.jsonl`, "", "utf-8");
+		writeFileSync(`${TEST_DIR}/phantom-config/meta/golden-suite.jsonl`, "", "utf-8");
+
+		const engine = new EvolutionEngine(CONFIG_PATH);
+		expect(engine.getEvolutionConfig().judges.cost_cap_usd_per_day).toBe(50.0);
+	});
+
+	test("engine uses heuristic path when judges are disabled", async () => {
+		setupTestEnv(50.0);
+		const engine = new EvolutionEngine(CONFIG_PATH);
+
+		// judges.enabled: "never" means heuristics
+		expect(engine.usesLLMJudges()).toBe(false);
+
+		const result = await engine.afterSession(makeSession());
+		// Should still work with heuristics
+		expect(result.changes_applied.length).toBeGreaterThan(0);
+
+		const userProfile = readFileSync(`${TEST_DIR}/phantom-config/user-profile.md`, "utf-8");
+		expect(userProfile).toContain("TypeScript");
+	});
+});
diff --git a/src/evolution/__tests__/engine.test.ts b/src/evolution/__tests__/engine.test.ts
index 47fef28..7af4012 100644
--- a/src/evolution/__tests__/engine.test.ts
+++ b/src/evolution/__tests__/engine.test.ts
@@ -27,6 +27,8 @@ function setupTestEnvironment(): void {
 			"  auto_rollback_window: 5",
 			"reflection:",
 			'  model: "claude-sonnet-4-20250514"',
+			"judges:",
+			'  enabled: "never"',
 			"paths:",
 			`  config_dir: "${TEST_DIR}/phantom-config"`,
 			`  constitution: "${TEST_DIR}/phantom-config/constitution.md"`,
diff --git a/src/evolution/__tests__/golden-suite-cap.test.ts b/src/evolution/__tests__/golden-suite-cap.test.ts
new file mode 100644
index 0000000..1741f83
--- /dev/null
+++ b/src/evolution/__tests__/golden-suite-cap.test.ts
@@ -0,0 +1,116 @@
+import { afterEach, beforeEach, describe, expect, test } from "bun:test";
+import { mkdirSync, rmSync, writeFileSync } from "node:fs";
+import type { EvolutionConfig } from "../config.ts";
+import { addCase, loadSuite, pruneSuite } from "../golden-suite.ts";
+import type { GoldenCase } from "../types.ts";
+
+const TEST_DIR = "/tmp/phantom-test-golden-cap";
+
+function testConfig(): EvolutionConfig {
+	return {
+		cadence: { reflection_interval: 1, consolidation_interval: 10, full_review_interval: 50, drift_check_interval: 20 },
+		gates: { drift_threshold: 0.7, max_file_lines: 200, auto_rollback_threshold: 0.1, auto_rollback_window: 5 },
+		reflection: { model: "claude-sonnet-4-20250514", effort: "high", max_budget_usd: 0.5 },
+		judges: { enabled: "auto", cost_cap_usd_per_day: 50.0, max_golden_suite_size: 50 },
+		paths: {
+			config_dir: TEST_DIR,
+			constitution: `${TEST_DIR}/constitution.md`,
+			version_file: `${TEST_DIR}/meta/version.json`,
+			metrics_file: `${TEST_DIR}/meta/metrics.json`,
+			evolution_log: `${TEST_DIR}/meta/evolution-log.jsonl`,
+			golden_suite: `${TEST_DIR}/meta/golden-suite.jsonl`,
+			session_log: `${TEST_DIR}/memory/session-log.jsonl`,
+		},
+	};
+}
+
+function makeGoldenCase(index: number, daysAgo = 0): GoldenCase {
+	const date = new Date();
+	date.setDate(date.getDate() - daysAgo);
+	return {
+		id: `golden-${index}`,
+		description: `Correction ${index}`,
+		lesson: `Lesson for correction ${index}`,
+		session_id: `session-${index}`,
+		created_at: date.toISOString(),
+	};
+}
+
+describe("Golden Suite Cap", () => {
+	beforeEach(() => {
+		mkdirSync(`${TEST_DIR}/meta`, { recursive: true });
+		writeFileSync(`${TEST_DIR}/meta/golden-suite.jsonl`, "", "utf-8");
+	});
+
+	afterEach(() => {
+		rmSync(TEST_DIR, { recursive: true, force: true });
+	});
+
+	test("pruneSuite is a no-op when suite is under the cap", () => {
+		const config = testConfig();
+		for (let i = 0; i < 5; i++) {
+			addCase(config, makeGoldenCase(i));
+		}
+		const removed = pruneSuite(config, 50);
+		expect(removed).toBe(0);
+		expect(loadSuite(config)).toHaveLength(5);
+	});
+
+	test("pruneSuite removes oldest entries when suite exceeds cap", () => {
+		const config = testConfig();
+		// Add 10 cases with decreasing age (0 = newest, 9 = oldest)
+		for (let i = 0; i < 10; i++) {
+			addCase(config, makeGoldenCase(i, i));
+		}
+		expect(loadSuite(config)).toHaveLength(10);
+
+		const removed = pruneSuite(config, 5);
+		expect(removed).toBe(5);
+
+		const remaining = loadSuite(config);
+		expect(remaining).toHaveLength(5);
+
+		// Remaining should be the 5 newest (days ago 0-4)
+		for (const entry of remaining) {
+			const id = Number.parseInt(entry.id.replace("golden-", ""), 10);
+			expect(id).toBeLessThan(5);
+		}
+	});
+
+	test("pruneSuite with max_golden_suite_size defaults to 50", () => {
+		const config = testConfig();
+		// Default cap is 50 from the config
+		expect(config.judges.max_golden_suite_size).toBe(50);
+	});
+
+	test("pruneSuite handles empty suite", () => {
+		const config = testConfig();
+		const removed = pruneSuite(config, 50);
+		expect(removed).toBe(0);
+	});
+
+	test("pruneSuite handles suite at exactly the cap", () => {
+		const config = testConfig();
+		for (let i = 0; i < 5; i++) {
+			addCase(config, makeGoldenCase(i));
+		}
+		const removed = pruneSuite(config, 5);
+		expect(removed).toBe(0);
+		expect(loadSuite(config)).toHaveLength(5);
+	});
+
+	test("pruneSuite keeps newest entries when exceeding cap by 1", () => {
+		const config = testConfig();
+		// oldest first, then newest
+		addCase(config, makeGoldenCase(0, 10));
+		addCase(config, makeGoldenCase(1, 0));
+
+		const removed = pruneSuite(config, 1);
+		expect(removed).toBe(1);
+
+		const remaining = loadSuite(config);
+		expect(remaining).toHaveLength(1);
+		// The newest entry (days ago 0) should remain
+		expect(remaining[0].id).toBe("golden-1");
+	});
+});
diff --git a/src/evolution/__tests__/judge-activation.test.ts b/src/evolution/__tests__/judge-activation.test.ts
new file mode 100644
index 0000000..59935ce
--- /dev/null
+++ b/src/evolution/__tests__/judge-activation.test.ts
@@ -0,0 +1,205 @@
+import { afterEach, beforeEach, describe, expect, test } from "bun:test";
+import { mkdirSync, rmSync, writeFileSync } from "node:fs";
+import { EvolutionEngine } from "../engine.ts";
+
+const TEST_DIR = "/tmp/phantom-test-judge-activation";
+const CONFIG_PATH = `${TEST_DIR}/config/evolution.yaml`;
+
+function setupWithJudgeMode(enabled: "auto" | "always" | "never"): void {
+	mkdirSync(`${TEST_DIR}/config`, { recursive: true });
+	mkdirSync(`${TEST_DIR}/phantom-config/meta`, { recursive: true });
+	mkdirSync(`${TEST_DIR}/phantom-config/strategies`, { recursive: true });
+	mkdirSync(`${TEST_DIR}/phantom-config/memory`, { recursive: true });
+
+	writeFileSync(
+		CONFIG_PATH,
+		[
+			"cadence:",
+			"  reflection_interval: 1",
+			"  consolidation_interval: 10",
+			"gates:",
+			"  drift_threshold: 0.7",
+			"  max_file_lines: 200",
+			"  auto_rollback_threshold: 0.1",
+			"  auto_rollback_window: 5",
+			"reflection:",
+			'  model: "claude-sonnet-4-20250514"',
+			"judges:",
+			`  enabled: "${enabled}"`,
+			"paths:",
+			`  config_dir: "${TEST_DIR}/phantom-config"`,
+			`  constitution: "${TEST_DIR}/phantom-config/constitution.md"`,
+			`  version_file: "${TEST_DIR}/phantom-config/meta/version.json"`,
+			`  metrics_file: "${TEST_DIR}/phantom-config/meta/metrics.json"`,
+			`  evolution_log: "${TEST_DIR}/phantom-config/meta/evolution-log.jsonl"`,
+			`  golden_suite: "${TEST_DIR}/phantom-config/meta/golden-suite.jsonl"`,
+			`  session_log: "${TEST_DIR}/phantom-config/memory/session-log.jsonl"`,
+		].join("\n"),
+		"utf-8",
+	);
+
+	writeFileSync(`${TEST_DIR}/phantom-config/constitution.md`, "# Constitution\n1. Be honest.\n", "utf-8");
+	writeFileSync(`${TEST_DIR}/phantom-config/persona.md`, "", "utf-8");
+	writeFileSync(`${TEST_DIR}/phantom-config/user-profile.md`, "", "utf-8");
+	writeFileSync(`${TEST_DIR}/phantom-config/domain-knowledge.md`, "", "utf-8");
+	writeFileSync(`${TEST_DIR}/phantom-config/strategies/task-patterns.md`, "", "utf-8");
+	writeFileSync(`${TEST_DIR}/phantom-config/strategies/tool-preferences.md`, "", "utf-8");
+	writeFileSync(`${TEST_DIR}/phantom-config/strategies/error-recovery.md`, "", "utf-8");
+	writeFileSync(`${TEST_DIR}/phantom-config/memory/session-log.jsonl`, "", "utf-8");
+	writeFileSync(
+		`${TEST_DIR}/phantom-config/meta/version.json`,
+		JSON.stringify({
+			version: 0,
+			parent: null,
+			timestamp: new Date().toISOString(),
+			changes: [],
+			metrics_at_change: { session_count: 0, success_rate_7d: 0, correction_rate_7d: 0 },
+		}),
+		"utf-8",
+	);
+	writeFileSync(
+		`${TEST_DIR}/phantom-config/meta/metrics.json`,
+		JSON.stringify({
+			session_count: 0,
+			success_count: 0,
+			failure_count: 0,
+			correction_count: 0,
+			evolution_count: 0,
+			rollback_count: 0,
+			last_session_at: null,
+			last_evolution_at: null,
+			success_rate_7d: 0,
+			correction_rate_7d: 0,
+			sessions_since_consolidation: 0,
+		}),
+		"utf-8",
+	);
+	writeFileSync(`${TEST_DIR}/phantom-config/meta/evolution-log.jsonl`, "", "utf-8");
+	writeFileSync(`${TEST_DIR}/phantom-config/meta/golden-suite.jsonl`, "", "utf-8");
+}
+
+let savedApiKey: string | undefined;
+
+describe("Judge Activation", () => {
+	beforeEach(() => {
+		savedApiKey = process.env.ANTHROPIC_API_KEY;
+	});
+
+	afterEach(() => {
+		if (savedApiKey !== undefined) {
+			process.env.ANTHROPIC_API_KEY = savedApiKey;
+		} else {
+			process.env.ANTHROPIC_API_KEY = undefined;
+		}
+		rmSync(TEST_DIR, { recursive: true, force: true });
+	});
+
+	test("auto mode enables judges when ANTHROPIC_API_KEY is set", () => {
+		process.env.ANTHROPIC_API_KEY = "sk-test-key";
+		setupWithJudgeMode("auto");
+		const engine = new EvolutionEngine(CONFIG_PATH);
+		expect(engine.usesLLMJudges()).toBe(true);
+	});
+
+	test("auto mode disables judges when ANTHROPIC_API_KEY is missing", () => {
+		process.env.ANTHROPIC_API_KEY = undefined;
+		setupWithJudgeMode("auto");
+		const engine = new EvolutionEngine(CONFIG_PATH);
+		expect(engine.usesLLMJudges()).toBe(false);
+	});
+
+	test("never mode disables judges even when API key is set", () => {
+		process.env.ANTHROPIC_API_KEY = "sk-test-key";
+		setupWithJudgeMode("never");
+		const engine = new EvolutionEngine(CONFIG_PATH);
+		expect(engine.usesLLMJudges()).toBe(false);
+	});
+
+	test("always mode enables judges regardless of API key", () => {
+		process.env.ANTHROPIC_API_KEY = undefined;
+		setupWithJudgeMode("always");
+		const engine = new EvolutionEngine(CONFIG_PATH);
+		expect(engine.usesLLMJudges()).toBe(true);
+	});
+
+	test("usesLLMJudges accessor matches resolved state", () => {
+		process.env.ANTHROPIC_API_KEY = "sk-test-key";
+		setupWithJudgeMode("auto");
+		const engine = new EvolutionEngine(CONFIG_PATH);
+		expect(engine.usesLLMJudges()).toBe(true);
+
+		process.env.ANTHROPIC_API_KEY = undefined;
+		setupWithJudgeMode("auto");
+		const engine2 = new EvolutionEngine(CONFIG_PATH);
+		expect(engine2.usesLLMJudges()).toBe(false);
+	});
+
+	test("missing judges section defaults to auto mode", () => {
+		process.env.ANTHROPIC_API_KEY = undefined;
+		mkdirSync(`${TEST_DIR}/config`, { recursive: true });
+		mkdirSync(`${TEST_DIR}/phantom-config/meta`, { recursive: true });
+		mkdirSync(`${TEST_DIR}/phantom-config/strategies`, { recursive: true });
+		mkdirSync(`${TEST_DIR}/phantom-config/memory`, { recursive: true });
+
+		// Config without judges section
+		writeFileSync(
+			CONFIG_PATH,
+			[
+				"cadence:",
+				"  reflection_interval: 1",
+				"paths:",
+				`  config_dir: "${TEST_DIR}/phantom-config"`,
+				`  constitution: "${TEST_DIR}/phantom-config/constitution.md"`,
+				`  version_file: "${TEST_DIR}/phantom-config/meta/version.json"`,
+				`  metrics_file: "${TEST_DIR}/phantom-config/meta/metrics.json"`,
+				`  evolution_log: "${TEST_DIR}/phantom-config/meta/evolution-log.jsonl"`,
+				`  golden_suite: "${TEST_DIR}/phantom-config/meta/golden-suite.jsonl"`,
+				`  session_log: "${TEST_DIR}/phantom-config/memory/session-log.jsonl"`,
+			].join("\n"),
+			"utf-8",
+		);
+
+		writeFileSync(`${TEST_DIR}/phantom-config/constitution.md`, "# Constitution\n", "utf-8");
+		writeFileSync(`${TEST_DIR}/phantom-config/persona.md`, "", "utf-8");
+		writeFileSync(`${TEST_DIR}/phantom-config/user-profile.md`, "", "utf-8");
+		writeFileSync(`${TEST_DIR}/phantom-config/domain-knowledge.md`, "", "utf-8");
+		writeFileSync(`${TEST_DIR}/phantom-config/strategies/task-patterns.md`, "", "utf-8");
+		writeFileSync(`${TEST_DIR}/phantom-config/strategies/tool-preferences.md`, "", "utf-8");
+		writeFileSync(`${TEST_DIR}/phantom-config/strategies/error-recovery.md`, "", "utf-8");
+		writeFileSync(`${TEST_DIR}/phantom-config/memory/session-log.jsonl`, "", "utf-8");
+		writeFileSync(
+			`${TEST_DIR}/phantom-config/meta/version.json`,
+			JSON.stringify({
+				version: 0,
+				parent: null,
+				timestamp: new Date().toISOString(),
+				changes: [],
+				metrics_at_change: { session_count: 0, success_rate_7d: 0, correction_rate_7d: 0 },
+			}),
+			"utf-8",
+		);
+		writeFileSync(
+			`${TEST_DIR}/phantom-config/meta/metrics.json`,
+			JSON.stringify({
+				session_count: 0,
+				success_count: 0,
+				failure_count: 0,
+				correction_count: 0,
+				evolution_count: 0,
+				rollback_count: 0,
+				last_session_at: null,
+				last_evolution_at: null,
+				success_rate_7d: 0,
+				correction_rate_7d: 0,
+				sessions_since_consolidation: 0,
+			}),
+			"utf-8",
+		);
+		writeFileSync(`${TEST_DIR}/phantom-config/meta/evolution-log.jsonl`, "", "utf-8");
+		writeFileSync(`${TEST_DIR}/phantom-config/meta/golden-suite.jsonl`, "", "utf-8");
+
+		// No API key + auto = disabled
+		const engine = new EvolutionEngine(CONFIG_PATH);
+		expect(engine.usesLLMJudges()).toBe(false);
+	});
+});
diff --git a/src/evolution/__tests__/metrics.test.ts b/src/evolution/__tests__/metrics.test.ts
index 2a640b4..fcd24e4 100644
--- a/src/evolution/__tests__/metrics.test.ts
+++ b/src/evolution/__tests__/metrics.test.ts
@@ -18,6 +18,7 @@ function testConfig(): EvolutionConfig {
 		cadence: { reflection_interval: 1, consolidation_interval: 10, full_review_interval: 50, drift_check_interval: 20 },
 		gates: { drift_threshold: 0.7, max_file_lines: 200, auto_rollback_threshold: 0.1, auto_rollback_window: 5 },
 		reflection: { model: "claude-sonnet-4-20250514", effort: "high", max_budget_usd: 0.5 },
+		judges: { enabled: "auto", cost_cap_usd_per_day: 50.0, max_golden_suite_size: 50 },
 		paths: {
 			config_dir: TEST_DIR,
 			constitution: `${TEST_DIR}/constitution.md`,
diff --git a/src/evolution/__tests__/validation.test.ts b/src/evolution/__tests__/validation.test.ts
index 9c42ba4..9eae6ce 100644
--- a/src/evolution/__tests__/validation.test.ts
+++ b/src/evolution/__tests__/validation.test.ts
@@ -20,6 +20,7 @@ function testConfig(): EvolutionConfig {
 		cadence: { reflection_interval: 1, consolidation_interval: 10, full_review_interval: 50, drift_check_interval: 20 },
 		gates: { drift_threshold: 0.7, max_file_lines: 200, auto_rollback_threshold: 0.1, auto_rollback_window: 5 },
 		reflection: { model: "claude-sonnet-4-20250514", effort: "high", max_budget_usd: 0.5 },
+		judges: { enabled: "auto", cost_cap_usd_per_day: 50.0, max_golden_suite_size: 50 },
 		paths: {
 			config_dir: TEST_DIR,
 			constitution: `${TEST_DIR}/constitution.md`,
diff --git a/src/evolution/__tests__/versioning.test.ts b/src/evolution/__tests__/versioning.test.ts
index 2703f35..72868ca 100644
--- a/src/evolution/__tests__/versioning.test.ts
+++ b/src/evolution/__tests__/versioning.test.ts
@@ -11,6 +11,7 @@ function testConfig(): EvolutionConfig {
 		cadence: { reflection_interval: 1, consolidation_interval: 10, full_review_interval: 50, drift_check_interval: 20 },
 		gates: { drift_threshold: 0.7, max_file_lines: 200, auto_rollback_threshold: 0.1, auto_rollback_window: 5 },
 		reflection: { model: "claude-sonnet-4-20250514", effort: "high", max_budget_usd: 0.5 },
+		judges: { enabled: "auto", cost_cap_usd_per_day: 50.0, max_golden_suite_size: 50 },
 		paths: {
 			config_dir: TEST_DIR,
 			constitution: `${TEST_DIR}/constitution.md`,
diff --git a/src/evolution/config.ts b/src/evolution/config.ts
index abbcd89..ea2f6d2 100644
--- a/src/evolution/config.ts
+++ b/src/evolution/config.ts
@@ -26,6 +26,13 @@ export const EvolutionConfigSchema = z.object({
 			max_budget_usd: z.number().positive().default(0.5),
 		})
 		.default({}),
+	judges: z
+		.object({
+			enabled: z.enum(["auto", "always", "never"]).default("auto"),
+			cost_cap_usd_per_day: z.number().positive().default(50.0),
+			max_golden_suite_size: z.number().int().positive().default(50),
+		})
+		.default({}),
 	paths: z
 		.object({
 			config_dir: z.string().default("phantom-config"),
diff --git a/src/evolution/engine.ts b/src/evolution/engine.ts
index 87bd38d..c32b4df 100644
--- a/src/evolution/engine.ts
+++ b/src/evolution/engine.ts
@@ -4,8 +4,7 @@ import { applyApproved } from "./application.ts";
 import { type EvolutionConfig, loadEvolutionConfig } from "./config.ts";
 import { recordObservations, runConsolidation } from "./consolidation.ts";
 import { ConstitutionChecker } from "./constitution.ts";
-import { addCase } from "./golden-suite.ts";
-import { loadSuite } from "./golden-suite.ts";
+import { addCase, loadSuite, pruneSuite } from "./golden-suite.ts";
 import { runQualityJudge } from "./judges/quality-judge.ts";
 import { type JudgeCosts, emptyJudgeCosts } from "./judges/types.ts";
 import {
@@ -30,24 +29,45 @@ import { getHistory, readVersion, rollback as versionRollback } from "./versioni
 export class EvolutionEngine {
 	private config: EvolutionConfig;
 	private checker: ConstitutionChecker;
-	private useLLMJudges: boolean;
+	private llmJudgesEnabled: boolean;
+	private dailyCostUsd = 0;
+	private dailyCostResetDate = "";
 
-	constructor(configPath?: string, useLLMJudges = false) {
+	constructor(configPath?: string) {
 		this.config = loadEvolutionConfig(configPath);
 		this.checker = new ConstitutionChecker(this.config);
-		this.useLLMJudges = useLLMJudges;
+		this.llmJudgesEnabled = this.resolveJudgeMode();
+		if (this.llmJudgesEnabled) {
+			console.log("[evolution] LLM judges enabled (API key detected)");
+		} else {
+			console.log("[evolution] LLM judges disabled (no API key or config override)");
+		}
 	}
 
-	getEvolutionConfig(): EvolutionConfig {
-		return this.config;
+	private resolveJudgeMode(): boolean {
+		const setting = this.config.judges?.enabled ?? "auto";
+		if (setting === "never") return false;
+		if (setting === "always") return true;
+		return !!process.env.ANTHROPIC_API_KEY;
 	}
 
-	enableLLMJudges(): void {
-		this.useLLMJudges = true;
+	usesLLMJudges(): boolean {
+		return this.llmJudgesEnabled;
 	}
 
-	disableLLMJudges(): void {
-		this.useLLMJudges = false;
+	/** Memory consolidation runs outside afterSession() but still needs to respect the daily cap. */
+	isWithinCostCap(): boolean {
+		return !this.isDailyCostCapReached();
+	}
+
+	/** Consolidation judge costs happen outside the evolution pipeline but count toward the daily cap. */
+	trackExternalJudgeCost(cost: { totalUsd: number }): void {
+		this.resetDailyCostIfNewDay();
+		this.dailyCostUsd += cost.totalUsd;
+	}
+
+	getEvolutionConfig(): EvolutionConfig {
+		return this.config;
 	}
 
 	/**
@@ -62,12 +82,13 @@ export class EvolutionEngine {
 
 		// Step 1: Observation Extraction (LLM or heuristic)
 		let observations: import("./types.ts").SessionObservation[];
-		if (this.useLLMJudges) {
+		if (this.llmJudgesEnabled && !this.isDailyCostCapReached()) {
 			const currentConfig = this.getConfig();
 			const result = await extractObservationsWithLLM(session, currentConfig);
 			observations = result.observations;
 			if (result.judgeCost) {
 				addCost(judgeCosts.observation_extraction, result.judgeCost);
+				this.incrementDailyCost(result.judgeCost.totalUsd);
 			}
 		} else {
 			observations = extractObservations(session);
@@ -98,10 +119,11 @@ export class EvolutionEngine {
 		const goldenSuite = loadSuite(this.config);
 		let validationResults: import("./types.ts").ValidationResult[];
 
-		if (this.useLLMJudges) {
+		if (this.llmJudgesEnabled && !this.isDailyCostCapReached()) {
 			const judgeResult = await validateAllWithJudges(deltas, this.checker, goldenSuite, this.config, currentConfig);
 			validationResults = judgeResult.results;
 			mergeCosts(judgeCosts, judgeResult.judgeCosts);
+			this.incrementDailyCost(totalCostFromJudgeCosts(judgeResult.judgeCosts));
 		} else {
 			validationResults = validateAll(deltas, this.checker, goldenSuite, this.config);
 		}
@@ -139,13 +161,14 @@ export class EvolutionEngine {
 		}
 
 		// Quality Assessment (LLM only, non-blocking)
-		if (this.useLLMJudges) {
+		if (this.llmJudgesEnabled && !this.isDailyCostCapReached()) {
 			try {
 				const qualityResult = await runQualityJudge(session, currentConfig);
 				judgeCosts.quality_assessment.calls++;
 				judgeCosts.quality_assessment.totalUsd += qualityResult.costUsd;
 				judgeCosts.quality_assessment.totalInputTokens += qualityResult.inputTokens;
 				judgeCosts.quality_assessment.totalOutputTokens += qualityResult.outputTokens;
+				this.incrementDailyCost(qualityResult.costUsd);
 
 				if (qualityResult.data.regression_signal) {
 					console.warn(
@@ -184,11 +207,14 @@ export class EvolutionEngine {
 			this.rollback(this.getCurrentVersion() - 1);
 		}
 
-		// Record judge costs
-		if (this.useLLMJudges) {
+		// Record judge costs to persistent metrics (daily tracking already done incrementally above)
+		if (this.llmJudgesEnabled) {
 			this.recordJudgeCosts(judgeCosts);
 		}
 
+		// Enforce golden suite cap
+		this.pruneGoldenSuite();
+
 		return {
 			version: this.getCurrentVersion(),
 			changes_applied: applied,
@@ -236,6 +262,39 @@ export class EvolutionEngine {
 		console.log(`[evolution] Rolled back to version ${toVersion}`);
 	}
 
+	private resetDailyCostIfNewDay(): void {
+		const today = new Date().toISOString().slice(0, 10);
+		if (this.dailyCostResetDate !== today) {
+			this.dailyCostUsd = 0;
+			this.dailyCostResetDate = today;
+		}
+	}
+
+	private isDailyCostCapReached(): boolean {
+		this.resetDailyCostIfNewDay();
+		const cap = this.config.judges?.cost_cap_usd_per_day ?? 50.0;
+		if (this.dailyCostUsd >= cap) {
+			console.warn(
+				`[evolution] Daily cost cap reached ($${this.dailyCostUsd.toFixed(2)} >= $${cap}), using heuristics`,
+			);
+			return true;
+		}
+		return false;
+	}
+
+	private incrementDailyCost(usd: number): void {
+		this.resetDailyCostIfNewDay();
+		this.dailyCostUsd += usd;
+	}
+
+	private pruneGoldenSuite(): void {
+		const maxSize = this.config.judges?.max_golden_suite_size ?? 50;
+		const removed = pruneSuite(this.config, maxSize);
+		if (removed > 0) {
+			console.log(`[evolution] Pruned ${removed} oldest golden suite entries (cap: ${maxSize})`);
+		}
+	}
+
 	private recordJudgeCosts(costs: JudgeCosts): void {
 		const metricsPath = this.config.paths.metrics_file;
 		try {
@@ -277,3 +336,11 @@ function mergeCosts(target: JudgeCosts, source: JudgeCosts): void {
 		addCost(target[key], source[key]);
 	}
 }
+
+function totalCostFromJudgeCosts(costs: JudgeCosts): number {
+	let total = 0;
+	for (const key of Object.keys(costs) as Array<keyof JudgeCosts>) {
+		total += costs[key].totalUsd;
+	}
+	return total;
+}
diff --git a/src/evolution/golden-suite.ts b/src/evolution/golden-suite.ts
index 22f780f..5643a9a 100644
--- a/src/evolution/golden-suite.ts
+++ b/src/evolution/golden-suite.ts
@@ -45,6 +45,22 @@ export function addCase(config: EvolutionConfig, goldenCase: GoldenCase): void {
 	}
 }
 
+/**
+ * Prune the golden suite to the given max size, removing oldest entries.
+ * No-op if the suite is within the limit.
+ */
+export function pruneSuite(config: EvolutionConfig, maxSize: number): number {
+	const suite = loadSuite(config);
+	if (suite.length <= maxSize) return 0;
+
+	const sorted = suite.sort((a, b) => b.created_at.localeCompare(a.created_at));
+	const pruned = sorted.slice(0, maxSize);
+	const content = pruned.map((c) => JSON.stringify(c)).join("\n");
+	writeFileSync(config.paths.golden_suite, `${content}\n`, "utf-8");
+
+	return suite.length - maxSize;
+}
+
 /**
  * Run the golden suite against a proposed change description.
  * Returns cases that might be affected.
diff --git a/src/evolution/judges/client.ts b/src/evolution/judges/client.ts
index 088f54d..6254e99 100644
--- a/src/evolution/judges/client.ts
+++ b/src/evolution/judges/client.ts
@@ -1,6 +1,7 @@
 import Anthropic from "@anthropic-ai/sdk";
 import { zodOutputFormat } from "@anthropic-ai/sdk/helpers/zod";
-import type { z } from "zod";
+// zod/v4 required: matches schemas.ts for zodOutputFormat compatibility
+import type { z } from "zod/v4";
 import {
 	JUDGE_MAX_TOKENS,
 	JUDGE_TEMPERATURE,
@@ -23,6 +24,10 @@ export function setClient(client: Anthropic | null): void {
 	_client = client;
 }
 
+export function isJudgeAvailable(): boolean {
+	return !!process.env.ANTHROPIC_API_KEY;
+}
+
 /**
  * Call a single LLM judge with structured output.
  * Uses the raw Anthropic SDK (not the Agent SDK).
@@ -46,7 +51,9 @@ export async function callJudge<T>(options: {
 		system: options.systemPrompt,
 		messages: [{ role: "user", content: options.userMessage }],
 		output_config: {
-			format: zodOutputFormat(options.schema),
+			// Cast needed: SDK .d.ts references zod v3 types but runtime uses zod/v4
+			// biome-ignore lint/suspicious/noExplicitAny: bridging zod v3/v4 type mismatch
+			format: zodOutputFormat(options.schema as any),
 		},
 	});
 
diff --git a/src/evolution/judges/schemas.ts b/src/evolution/judges/schemas.ts
index b9fb0ef..a96f69d 100644
--- a/src/evolution/judges/schemas.ts
+++ b/src/evolution/judges/schemas.ts
@@ -1,4 +1,5 @@
-import { z } from "zod";
+// zod/v4 required: the Anthropic SDK's zodOutputFormat reads schema._zod.def (v4 only)
+import { z } from "zod/v4";
 
 // -- Observation Extraction --
 
diff --git a/src/evolution/judges/types.ts b/src/evolution/judges/types.ts
index 6402013..755972d 100644
--- a/src/evolution/judges/types.ts
+++ b/src/evolution/judges/types.ts
@@ -1,6 +1,6 @@
-export const JUDGE_MODEL_SONNET = "claude-sonnet-4-6-20250514";
-export const JUDGE_MODEL_HAIKU = "claude-haiku-4-5-20250929";
-export const JUDGE_MODEL_OPUS = "claude-opus-4-6-20250918";
+export const JUDGE_MODEL_SONNET = "claude-sonnet-4-6";
+export const JUDGE_MODEL_HAIKU = "claude-haiku-4-5";
+export const JUDGE_MODEL_OPUS = "claude-opus-4-6";
 
 export const JUDGE_TIMEOUT_MS = 30_000;
 export const JUDGE_MAX_TOKENS = 4096;
diff --git a/src/index.ts b/src/index.ts
index 4a25539..47fd660 100644
--- a/src/index.ts
+++ b/src/index.ts
@@ -37,7 +37,7 @@ import { PeerHealthMonitor } from "./mcp/peer-health.ts";
 import { PeerManager } from "./mcp/peers.ts";
 import { PhantomMcpServer } from "./mcp/server.ts";
 import { loadMemoryConfig } from "./memory/config.ts";
-import { type SessionData, consolidateSession } from "./memory/consolidation.ts";
+import { type SessionData, consolidateSession, consolidateSessionWithLLM } from "./memory/consolidation.ts";
 import { MemoryContextBuilder } from "./memory/context-builder.ts";
 import { MemorySystem } from "./memory/system.ts";
 import { isFirstRun, isOnboardingInProgress } from "./onboarding/detection.ts";
@@ -99,7 +99,8 @@ async function main(): Promise<void> {
 	try {
 		evolution = new EvolutionEngine();
 		const currentVersion = evolution.getCurrentVersion();
-		console.log(`[evolution] Engine initialized (v${currentVersion})`);
+		const judgeMode = evolution.usesLLMJudges() ? "LLM judges" : "heuristic";
+		console.log(`[evolution] Engine initialized (v${currentVersion}, ${judgeMode})`);
 		setEvolutionVersionProvider(() => evolution?.getCurrentVersion() ?? 0);
 	} catch (err: unknown) {
 		const msg = err instanceof Error ? err.message : String(err);
@@ -491,19 +492,41 @@ async function main(): Promise<void> {
 				outcome: response.text.startsWith("Error:") ? "failure" : "success",
 			};
 
-			consolidateSession(memory, sessionData)
-				.then((result) => {
-					if (result.episodesCreated > 0 || result.factsExtracted > 0) {
-						console.log(
-							`[memory] Consolidated: ${result.episodesCreated} episodes, ` +
-								`${result.factsExtracted} facts (${result.durationMs}ms)`,
-						);
-					}
-				})
-				.catch((err: unknown) => {
-					const errMsg = err instanceof Error ? err.message : String(err);
-					console.warn(`[memory] Consolidation failed: ${errMsg}`);
-				});
+			const useLLMConsolidation = evolution?.usesLLMJudges() && evolution.isWithinCostCap();
+			if (useLLMConsolidation) {
+				const evolvedConfig = evolution?.getConfig();
+				const existingFacts = evolvedConfig ? `${evolvedConfig.userProfile}\n${evolvedConfig.domainKnowledge}` : "";
+				consolidateSessionWithLLM(memory, sessionData, existingFacts)
+					.then(({ result, judgeCost }) => {
+						if (judgeCost) {
+							evolution?.trackExternalJudgeCost(judgeCost);
+						}
+						if (result.episodesCreated > 0 || result.factsExtracted > 0) {
+							console.log(
+								`[memory] Consolidated (LLM): ${result.episodesCreated} episodes, ` +
+									`${result.factsExtracted} facts (${result.durationMs}ms)`,
+							);
+						}
+					})
+					.catch((err: unknown) => {
+						const errMsg = err instanceof Error ? err.message : String(err);
+						console.warn(`[memory] LLM consolidation failed: ${errMsg}`);
+					});
+			} else {
+				consolidateSession(memory, sessionData)
+					.then((result) => {
+						if (result.episodesCreated > 0 || result.factsExtracted > 0) {
+							console.log(
+								`[memory] Consolidated: ${result.episodesCreated} episodes, ` +
+									`${result.factsExtracted} facts (${result.durationMs}ms)`,
+							);
+						}
+					})
+					.catch((err: unknown) => {
+						const errMsg = err instanceof Error ? err.message : String(err);
+						console.warn(`[memory] Consolidation failed: ${errMsg}`);
+					});
+			}
 		}
 
 		// Evolution pipeline (non-blocking)
diff --git a/src/mcp/__tests__/dynamic-handlers.test.ts b/src/mcp/__tests__/dynamic-handlers.test.ts
index 1c5376b..86aa33b 100644
--- a/src/mcp/__tests__/dynamic-handlers.test.ts
+++ b/src/mcp/__tests__/dynamic-handlers.test.ts
@@ -1,6 +1,6 @@
 import { describe, expect, test } from "bun:test";
-import type { DynamicToolDef } from "../dynamic-tools.ts";
 import { buildSafeEnv, executeDynamicHandler } from "../dynamic-handlers.ts";
+import type { DynamicToolDef } from "../dynamic-tools.ts";
 
 describe("buildSafeEnv", () => {
 	test("includes only safe environment variables", () => {
@@ -24,7 +24,7 @@ describe("buildSafeEnv", () => {
 			if (origKey !== undefined) {
 				process.env.ANTHROPIC_API_KEY = origKey;
 			} else {
-				delete process.env.ANTHROPIC_API_KEY;
+				process.env.ANTHROPIC_API_KEY = undefined;
 			}
 		}
 	});
@@ -40,7 +40,7 @@ describe("buildSafeEnv", () => {
 			if (origToken !== undefined) {
 				process.env.SLACK_BOT_TOKEN = origToken;
 			} else {
-				delete process.env.SLACK_BOT_TOKEN;
+				process.env.SLACK_BOT_TOKEN = undefined;
 			}
 		}
 	});
@@ -82,7 +82,7 @@ describe("executeDynamicHandler", () => {
 			if (origKey !== undefined) {
 				process.env.ANTHROPIC_API_KEY = origKey;
 			} else {
-				delete process.env.ANTHROPIC_API_KEY;
+				process.env.ANTHROPIC_API_KEY = undefined;
 			}
 		}
 	});
@@ -154,7 +154,7 @@ describe("executeDynamicHandler", () => {
 			if (origKey !== undefined) {
 				process.env.ANTHROPIC_API_KEY = origKey;
 			} else {
-				delete process.env.ANTHROPIC_API_KEY;
+				process.env.ANTHROPIC_API_KEY = undefined;
 			}
 		}
 	});