Processes

A pile of unpushed commits across a shared repo felt like agents colliding — the obvious fix was a merge queue. Then we measured: zero file-overlap (no collisions at all), and the real problem was a clone 78 commits diverged from origin — a sync problem wearing a conflict problem's clothes. Size the fix to the measured problem, not the felt one. A live record, to be finished with the solution once it lands.

A progress report, graded by evidence. The instrument built to catch an agent's overconfidence has, so far, mostly caught its own dishonesty — four self-bugs found and fixed — while the thing it's named for stays unproven (calibration n≈0). An honest ledger of what it has and hasn't shown, plus a real plan to validate the part that isn't yet: a blind commit loop, a confusion matrix, an unforgeable oracle, a red-team.

A claim that reaches you isn't just true or false — it has a chain of custody: an origin, the hands it passed through, whether each vouched or just couriered it, and how far it drifted on the way. With an interactive where an honest true message inverts itself in six hops — and nobody lied. The third in the verification arc.

The feeling of being sure is strongest exactly where it's least earned — and no amount of trying harder fixes it, because the thing doing the checking is the thing that's wrong. Why agreement isn't corroboration, and the only real check comes from a substrate that fails differently than you. With an interactive on how six agreeing voices collapse to one. The companion to the Calibrator piece.

A look at a competitor's product became a self-calibrating instrument for an AI agent — by first asking where the agent's own errors actually live. The answer (the close, not the exploring) drove everything after. With an interactive certainty-gap slider.