diff --git a/packages/api/src/services/voiceprint/voiceprint.config.ts b/packages/api/src/services/voiceprint/voiceprint.config.ts
index f117f5a..f3bad81 100644
--- a/packages/api/src/services/voiceprint/voiceprint.config.ts
+++ b/packages/api/src/services/voiceprint/voiceprint.config.ts
@@ -1,22 +1,24 @@
 import { z } from 'zod';
+import { existsSync } from 'fs';
 import { checkFlag } from './voiceprint.feature-flags';
 
-// Environment variables for VoicePrint
+// P3-4 fix: Use strict() to catch typos in env var names
+// P3-1 fix: Use safeParse() to avoid module-level crash on missing env vars
 const envSchema = z.object({
   ECAPA_TDNN_MODEL_PATH: z.string().default('./models/ecapa-tdnn'),
   ML_SERVICE_URL: z.string().default('http://localhost:8001'),
   FAISS_INDEX_PATH: z.string().default('./data/voiceprint_faiss.index'),
   AUDIO_STORAGE_BUCKET: z.string().default('voiceprint-audio'),
   AUDIO_STORAGE_ENDPOINT: z.string().default('http://localhost:9000'),
-  SYNTHETIC_THRESHOLD: z.string().transform(Number).default(0.75),
-  ENROLLMENT_MIN_DURATION_SEC: z.string().transform(Number).default(3),
-  ENROLLMENT_MAX_DURATION_SEC: z.string().transform(Number).default(60),
-  EMBEDDING_DIMENSIONS: z.string().transform(Number).default(192),
-  BATCH_MAX_FILES: z.string().transform(Number).default(20),
-  ANALYSIS_TIMEOUT_MS: z.string().transform(Number).default(30000),
-});
+  SYNTHETIC_THRESHOLD: z.string().transform(Number).default('0.75'),
+  ENROLLMENT_MIN_DURATION_SEC: z.string().transform(Number).default('3'),
+  ENROLLMENT_MAX_DURATION_SEC: z.string().transform(Number).default('60'),
+  EMBEDDING_DIMENSIONS: z.string().transform(Number).default('192'),
+  BATCH_MAX_FILES: z.string().transform(Number).default('20'),
+  ANALYSIS_TIMEOUT_MS: z.string().transform(Number).default('30000'),
+}).strict();
 
-export const voicePrintEnv = envSchema.parse({
+const envInput = {
   ECAPA_TDNN_MODEL_PATH: process.env.ECAPA_TDNN_MODEL_PATH,
   ML_SERVICE_URL: process.env.ML_SERVICE_URL,
   FAISS_INDEX_PATH: process.env.FAISS_INDEX_PATH,
@@ -28,7 +30,23 @@ export const voicePrintEnv = envSchema.parse({
   EMBEDDING_DIMENSIONS: process.env.EMBEDDING_DIMENSIONS,
   BATCH_MAX_FILES: process.env.BATCH_MAX_FILES,
   ANALYSIS_TIMEOUT_MS: process.env.ANALYSIS_TIMEOUT_MS,
-});
+};
+
+const parsed = envSchema.safeParse(envInput);
+export const voicePrintEnv = parsed.success
+  ? parsed.data
+  : envSchema.parse({}); // fallback to all defaults
+
+// P3-3 fix: Validate model path exists at startup (warn, not crash)
+if (voicePrintEnv.ECAPA_TDNN_MODEL_PATH && !existsSync(voicePrintEnv.ECAPA_TDNN_MODEL_PATH)) {
+  console.warn(
+    `[VoicePrint] Model path not found: ${voicePrintEnv.ECAPA_TDNN_MODEL_PATH} (using mock model)`
+  );
+}
+
+if (!parsed.success) {
+  console.warn('[VoicePrint] Env validation warnings:', parsed.error.issues.map((i: z.ZodIssue) => `${i.path.join('.')}: ${i.message}`).join(', '));
+}
 
 // Audio source types
 export enum VoicePrintSource {
diff --git a/services/voiceprint/src/voiceprint.config.ts b/services/voiceprint/src/voiceprint.config.ts
index f117f5a..f3bad81 100644
--- a/services/voiceprint/src/voiceprint.config.ts
+++ b/services/voiceprint/src/voiceprint.config.ts
@@ -1,22 +1,24 @@
 import { z } from 'zod';
+import { existsSync } from 'fs';
 import { checkFlag } from './voiceprint.feature-flags';
 
-// Environment variables for VoicePrint
+// P3-4 fix: Use strict() to catch typos in env var names
+// P3-1 fix: Use safeParse() to avoid module-level crash on missing env vars
 const envSchema = z.object({
   ECAPA_TDNN_MODEL_PATH: z.string().default('./models/ecapa-tdnn'),
   ML_SERVICE_URL: z.string().default('http://localhost:8001'),
   FAISS_INDEX_PATH: z.string().default('./data/voiceprint_faiss.index'),
   AUDIO_STORAGE_BUCKET: z.string().default('voiceprint-audio'),
   AUDIO_STORAGE_ENDPOINT: z.string().default('http://localhost:9000'),
-  SYNTHETIC_THRESHOLD: z.string().transform(Number).default(0.75),
-  ENROLLMENT_MIN_DURATION_SEC: z.string().transform(Number).default(3),
-  ENROLLMENT_MAX_DURATION_SEC: z.string().transform(Number).default(60),
-  EMBEDDING_DIMENSIONS: z.string().transform(Number).default(192),
-  BATCH_MAX_FILES: z.string().transform(Number).default(20),
-  ANALYSIS_TIMEOUT_MS: z.string().transform(Number).default(30000),
-});
+  SYNTHETIC_THRESHOLD: z.string().transform(Number).default('0.75'),
+  ENROLLMENT_MIN_DURATION_SEC: z.string().transform(Number).default('3'),
+  ENROLLMENT_MAX_DURATION_SEC: z.string().transform(Number).default('60'),
+  EMBEDDING_DIMENSIONS: z.string().transform(Number).default('192'),
+  BATCH_MAX_FILES: z.string().transform(Number).default('20'),
+  ANALYSIS_TIMEOUT_MS: z.string().transform(Number).default('30000'),
+}).strict();
 
-export const voicePrintEnv = envSchema.parse({
+const envInput = {
   ECAPA_TDNN_MODEL_PATH: process.env.ECAPA_TDNN_MODEL_PATH,
   ML_SERVICE_URL: process.env.ML_SERVICE_URL,
   FAISS_INDEX_PATH: process.env.FAISS_INDEX_PATH,
@@ -28,7 +30,23 @@ export const voicePrintEnv = envSchema.parse({
   EMBEDDING_DIMENSIONS: process.env.EMBEDDING_DIMENSIONS,
   BATCH_MAX_FILES: process.env.BATCH_MAX_FILES,
   ANALYSIS_TIMEOUT_MS: process.env.ANALYSIS_TIMEOUT_MS,
-});
+};
+
+const parsed = envSchema.safeParse(envInput);
+export const voicePrintEnv = parsed.success
+  ? parsed.data
+  : envSchema.parse({}); // fallback to all defaults
+
+// P3-3 fix: Validate model path exists at startup (warn, not crash)
+if (voicePrintEnv.ECAPA_TDNN_MODEL_PATH && !existsSync(voicePrintEnv.ECAPA_TDNN_MODEL_PATH)) {
+  console.warn(
+    `[VoicePrint] Model path not found: ${voicePrintEnv.ECAPA_TDNN_MODEL_PATH} (using mock model)`
+  );
+}
+
+if (!parsed.success) {
+  console.warn('[VoicePrint] Env validation warnings:', parsed.error.issues.map((i: z.ZodIssue) => `${i.path.join('.')}: ${i.message}`).join(', '));
+}
 
 // Audio source types
 export enum VoicePrintSource {