2026-09-19THE STANDARD5 min read

Done is a claim. GOLD is a receipt.

The whole standard exists because of one sentence: "it works." It's the most expensive sentence in software. It sounds like evidence. It contains none. "It works" tells you someone ran something, once, and liked the result. It doesn't tell you whether the result was skill, luck, or a baseline that was never tuned honestly.

Six gates, each with a named enemy

The standard didn't invent these checks — it enforces patterns that already existed, each aimed at a specific way claims die:

Gate 01 — Commit. The enemy is data dredging: tuning until the data looks significant. Procedures are fingerprinted before the sealed seed exists. There is no tuning window.

Gate 02 — Parity. The enemy is resource cheating. Candidate and baseline see identical instances with identical budgets. Same random.Random(1) for both. Winning by burning more compute is a cost failure, not a win.

Gate 03 — Placebo. The enemy is specification gaming. Gain must beat a mismatched hidden truth. If it survives a placebo, it's the prior, not skill.

Gate 04 — Paired win. The enemy is luck and the multiple comparisons problem. A bootstrap lower bound on the paired difference must clear zero, with Bonferroni correction across slices. "It won somewhere" is not "it won."

Gate 05 — Goodhart. The enemy is Goodhart's law. Gain on the optimizable proxy must not outrun gain on the sealed truth. Optimizing the scoreboard instead of the outcome fails here.

Gate 06 — Cost. The enemy is brute force. Runtime is held against a slack multiple of the baseline. Victory by compute is a cost failure.

Then five seeds and a holdout

All six gates, on five independent fresh seeds, plus one frozen holdout the procedure never touched. Only then does the loop stop. It sounds brutal. It's the minimum honest protocol given what we know about contamination.

The part that makes it worth trusting

A standard that only reports its wins is just marketing with a p-value. The receipts on this site include the numbers that hurt: a correlation of −0.144 between the rubric and the judge. Eleven Goodhart flags, undeleted. Those stay in the report permanently, because a receipt you can't check is just another claim with better typography.

The rule: GOLD reports everything — wins, losses, flags, bounds, cost. It only passes what survives all of it.

The takeaway: the next time someone says "it works," ask for the receipt: the commit hash, the seed, the lower bound, the placebo, the Goodhart delta, the runtime, and the holdout. If they can't produce those, they've made a claim — and claims are cheap.