Harvesting Without Breaking

The professional partial-score submission is one program with tiers:

read n, the values
if (n <= 20)      → run brute force, output, exit
if (n <= 2000)    → run O(n²), output, exit
run full solution

Each tier is independently testable; a bug in the full solution cannot destroy the n ≤ 20 points if the tier exits before the heavy path runs.

Tier discipline

• Tier order = band order. Ship and verify the easiest tier first.
• Shared input parsing. Parse everything once, correctly; most tier bugs are actually parsing bugs (the tier never sees the real input).
• The upgrade rule. When adding tier k+1, re-run tier k's tests. The most expensive contest mistake is a refactor that silently breaks the guaranteed points.
• Fallback ladder. If the full solution TLEs on the real test, the tiers above it still answered — partial scoring rewards programs that degrade gracefully.

Honesty

Never special-case the sample tests themselves. Hard-coding sample outputs is detectable, worth zero learning, and trains exactly the wrong instinct. Specialize on constraint shapes, not on the sample files.