diff --git a/src/cli/args.ts b/src/cli/args.ts
index 7f5b5d6..73905d2 100644
--- a/src/cli/args.ts
+++ b/src/cli/args.ts
@@ -66,14 +66,27 @@ export function boolFlag(args: ParsedArgs, keys: string | string[]): boolean {
   return false;
 }
 
+/** Raised when a numeric flag is present but not a valid number. */
+export class InvalidNumericFlagError extends Error {
+  constructor(flag: string, raw: string) {
+    super(`[evalgate] invalid numeric flag --${flag}: "${raw}"`);
+    this.name = "InvalidNumericFlagError";
+  }
+}
+
 /** Read a numeric flag with an optional default. */
 export function numFlag(
   args: ParsedArgs,
   keys: string | string[],
   fallback?: number,
 ): number | undefined {
-  const raw = strFlag(args, keys);
-  if (raw === undefined) return fallback;
-  const n = Number(raw);
-  return Number.isNaN(n) ? fallback : n;
+  const keyList = Array.isArray(keys) ? keys : [keys];
+  for (const k of keyList) {
+    const v = args.flags[k];
+    if (typeof v !== "string") continue;
+    const n = Number(v);
+    if (Number.isNaN(n)) throw new InvalidNumericFlagError(k, v);
+    return n;
+  }
+  return fallback;
 }
diff --git a/tests/args.test.ts b/tests/args.test.ts
new file mode 100644
index 0000000..02c31f3
--- /dev/null
+++ b/tests/args.test.ts
@@ -0,0 +1,70 @@
+import { describe, it, expect } from "vitest";
+import { mkdtemp, writeFile, rm } from "node:fs/promises";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import { numFlag, parseArgs, InvalidNumericFlagError } from "../src/cli/args.js";
+import { cmdCompare } from "../src/cli/index.js";
+import type { RunResult } from "../src/types.js";
+
+function runResult(score: number): RunResult {
+  return {
+    version: "1",
+    suite: "s",
+    timestamp: "now",
+    score,
+    passed: true,
+    total: 1,
+    passedCount: 1,
+    latencyMs: 0,
+    costUsd: 0,
+    cases: [
+      {
+        id: "c1",
+        model: "mock",
+        provider: "mock",
+        output: "",
+        latencyMs: 1,
+        costUsd: 0,
+        score,
+        passed: true,
+        scores: [],
+      },
+    ],
+  };
+}
+
+describe("numFlag", () => {
+  it("returns fallback when the flag is omitted", () => {
+    const args = parseArgs([]);
+    expect(numFlag(args, "tolerance", 0)).toBe(0);
+  });
+
+  it("parses a numeric flag value", () => {
+    const args = parseArgs(["--tolerance", "0.05"]);
+    expect(numFlag(args, "tolerance", 0)).toBe(0.05);
+  });
+
+  it("rejects non-numeric flag values", () => {
+    const args = parseArgs(["--tolerance", "abc"]);
+    expect(() => numFlag(args, "tolerance", 0)).toThrow(InvalidNumericFlagError);
+  });
+});
+
+describe("cmdCompare invalid tolerance", () => {
+  it("rejects a non-numeric --tolerance instead of silently using the default", async () => {
+    const dir = await mkdtemp(join(tmpdir(), "evalgate-"));
+    const base = join(dir, "base.json");
+    const head = join(dir, "head.json");
+    await writeFile(base, JSON.stringify(runResult(1)));
+    await writeFile(head, JSON.stringify(runResult(1)));
+
+    await expect(
+      cmdCompare({
+        _: ["compare"],
+        flags: { base, head, tolerance: "abc" },
+      }),
+    ).rejects.toThrow(/tolerance/i);
+
+    await rm(dir, { recursive: true, force: true });
+  });
+});