The Three-Program Harness

Stress testing is three programs in a loop:

1. The oracle — a brute-force solution, too slow for real constraints but obviously correct at small n.
2. The candidate — your fast solution, the thing under test.
3. The generator — a small seeded-random input producer.

The loop: generate → run both on the input → compare byte-exactly. On the first mismatch, you hold in your hands a failing input, guaranteed. Then shrink it to the minimal counterexample (Module 12) and diagnose.

Making the harness bite

• Seed everything. A reproducible failure is fixable; a one-time mismatch you cannot regenerate is a ghost.
• Vary the generator's shape, not just size. Alternate sorted, reversed, tie-heavy, tiny-alphabet, extreme-values profiles — each shape is a different attack surface (Module 12's checklist, automated).
• Compare outputs byte-exactly. Trailing whitespace differences are real verdicts on judges; let the harness teach you that early.

What stress finds and what it misses

Stress testing finds behavioral divergence reachable by your generator's shapes. It cannot find overflow at n = 200000 (the oracle cannot run there) or wrong-complexity death. For those, static analysis of the code and the budget arithmetic remain the tools — stress complements, never replaces, the checklist.