breakthrough-loop

Breakthrough-loop discipline (drop into .cursor/rules/)

Full methodology: https://github.com/GuoCheng24/breakthrough-harness

When running research experiments in this project:

• Triage the target first: cheap scoring function + existing groundwork + externally recognisable win + an oracle-ceiling/trivial-baseline gap worth mining (measure before round 1). Any missing -> flag it and stop.
• Build the harness before the method: single evaluation entry point; null models scored first and pinned to the floor; metric conventions pinned (double-report if the published one differs); calibration and evaluation data physically separated - tuning touches calibration only.
• Reproduce a published baseline number before trusting ours; an unmatched baseline means the recipe has unread layers (optimizer, loss, metric convention, operator).
• Loop: 10-30 variants/round each with a one-line reason it might win; sanity-check before entry; executors only call the entry point, never write method code inline; select on calibration; confirm top-k held-out. Unreproduced gains do not exist; report confirmations with a paired error bar, budget held-out accesses. Persist numbers to disk, then write.
• Claim polarity: the main sentence of any result is constructive ("we propose X, it solves Y, number Z"); audits and negatives are supporting material, never the headline.
• Verify every new guard by deliberately breaking what it watches, and confirm it fails for the right reason.
• Stop conditions: a well-scoring null model or any firing anti-cheat check freezes all conclusions until the harness is repaired; when candidates stop carrying new reasons, record the ceiling number and escalate to a more ambitious target class instead of continuing to sweep.