Writing
Usually on evaluation and how to know whether a system actually works.
- Weighing Unseen Planets: Calibrated, Amortized Inference from Transit Timing2026-06-18
When planets tug on each other, their transits arrive early or late. That timing signal encodes their masses, even for planets that never transit. This is what I learned trying to invert it with a neural network: the model stays honest where everyone feared it would lie, and the real wall is the data, not the algorithm.
- Breaking the TTV Degeneracy: I Said the Fix Was a Second Observable. Here Is the Test.2026-06-18
A follow-up experiment. In the last post I argued that the wide, near-resonant posteriors in transit-timing inference are physical, a limit of the data and not the model, so the lever is more observables rather than a fancier network. Here I add transit durations to the forward model and watch the mass-eccentricity degeneracy collapse.
- Two Agents, the Same Score, Different Failures2026-06-09
Aggregate success rate tells you an agent failed. It will not tell you that two agents with the same score fail in completely different ways. A short look at why per-axis failure profiles are the more useful number.
- Writing a Judge Rubric Two Models Can Agree On2026-06-09
When two LLM judges from different families disagree at high confidence, the problem is usually the rubric, not the judges. Here is the three-iteration loop that got them to converge, and the one change that actually did it.
- Which Capability Actually Broke? A Calibrated 5-Axis Judge for Agent Tool Use2026-06-04
End-task success tells you an agent failed, not why. Here is a 5-axis decomposition of tool-use correctness, graded by cross-family LLM judges, with the calibration discipline that makes the verdicts mean something.