研究 · 覚書 · 更新Blossom journal
Notes from the lab.
Papers, evals, field notes, and the occasional essay. Written by the people building it.
研究室から
Tokyo · San Francisco
Tokyo · San Francisco
Papers, evals, field notes, and the occasional essay. Written by the people building it.
A correct optimizer can look impressive when its verifier has already revealed the answer. Evidence must earn the right to become a score.
A strategic agent can hit the business KPI while failing to preserve the behavior that makes it deployable.
An agent can hit the outcome metric while learning the wrong behavior. Better evidence follows the trace.
A benchmark can have gold labels and still produce unstable evidence if the rubric, metric, or ranking rule moves.
Some AI evaluation claims become checkable only later. ValueBlindBench asks which claims are safe to cite now.
AI search does not only answer a query. It decides which sources become visible.