Capability 01

Long-context, evidence-backed code review.

The lab pins Kimi K3 to the first review pass, then asks MiniMax-M3 to cross-check and synthesize the highest-confidence findings.

Sample target
export function applyDiscount(cart, coupon) {
  if (!coupon) return cart.total;
  if (coupon.percent) return cart.total - coupon.percent;
  return cart.total;
}

export async function fetchUserProfile(fetcher, userId) {
  const res = await fetcher(`/api/users/${userId}`);
  return res.json();
}
K3 first-pass findings
  1. High: percent is subtracted as a flat amount instead of applied proportionally.
  2. Medium: raw userId interpolation can rewrite the URL path.
  3. Medium: res.json() runs without a response status check.
  4. Low: cart shape and numeric total are not validated.

This page summarizes the smoke-run outcome; it does not embed full transcripts or private model state.

Why this is a K3 capability demo

1

K3 reads the file and produces line-specific findings.

2

MiniMax-M3 verifies, refutes, and hunts missed edge cases.

3

The synthesis gate halts on high findings so issues are visible before promotion.