Every post is a named pattern with a real source, written from the other side of the gate — where claims go to be proven or die.
One demo, one seed, one lucky pass — and everyone calls it done. Statistical power, noise floors, and why n=300 paired instances per slice is sized, not improvised.
Goodhart's law is not a warning label — it's a gate. How the proxy-vs-truth check catches procedures that optimize the scoreboard instead of the outcome, and why the flags stay in the report.
DeepMind's catalogue of agents that satisfy the letter of the objective while violating its spirit — and the placebo test that catches them scoring against a truth they were never meant to see.
Black-box models memorize their evaluation sets — it's been proven. Five fresh sealed seeds plus one frozen holdout is the minimum honest protocol, and a single GOLD True never counts.
The origin story of the six gates: commit, parity, placebo, paired win, Goodhart, cost — and why a standard that reports its failures is the only kind worth trusting.