Developer tools, simulations, research systems, and whatever else seemed like a good idea at the time. 😈
 Tools for seeing what coding agents and npm scripts are really up to. |
 Agents that do real work, plus a record of how they did it. |
 Simulation and GPU compute that gives the same numbers every time. |
 Point-in-time data, careful evaluation, and the occasional null result. |
 Claims written down before the data shows up, so a study can't quietly change its mind. |
 Seeing what install scripts actually do, and keeping local tooling locked down. |
Some of the more interesting stuff is private, so you get what it does, not the code.
 Keeps tabs on coding agents: what ran, what finished, what got interrupted, and where to pick it back up. 71 tests · 22 live acceptance checks · 4 review rounds |
 Checks that the same computation gives the same answer on different hardware. Harder than it sounds. 2 fixtures · equivalence criteria locked in before any backend evidence |
 Shows what npm install scripts actually touch, spawn, and call. Good to know before you run one. Published on npm · strace-based runtime trace |
 My independent institute. Empirical research on how autonomous agents behave, and technology for understanding difficult systems. 7 public studies |
- neural-field-study JavaScript — A deterministic neural-field model you can replay live in the browser.
- deep-lob Python — A synthetic order-book study of when overlapping windows make a model look smarter than it is.
- release-equivalence Python — When is a new model release actually the same one? Identity, equivalence, rollback.
- clock-dataset Python — Point-in-time vs latest data, and how much hindsight sneaks into historical decisions.
- time-series-experiments Python — Forecasting experiments across volatility, regimes, structural breaks and chaos.
