Essays, notes, and working models for AI safety, agentic systems, and trustworthy infrastructure.
OpenAI cracked Navier-Stokes with 10,000 agents in 88 hours. The direction came from a 2021 thesis written without a computer. Compute scales the search. Taste still picks it. Here is how to extract taste to steer an autoresearch loop, and three UX to try.
Shortening two caliper skills, grill-skill and evaluate-skill, by about 45% using a non-inferiority eval bar: prove the shorter version is no worse before shipping the cut.
A case study in using Caliper to evaluate blader/humanizer, tighten voice calibration, and turn the improvement into an upstream contribution with regression coverage. With a postscript on what the maintainer did instead.
Tribal knowledge encoded as an AI skill is still just text until you evaluate it. Ablation baselines, routing regression tests, trajectory autoraters, and the gotchas flywheel keep encoded knowledge from rotting.
Everyone is worried about AI reading things it shouldn't. That's the wrong threat model. The problem starts after the agent reads, and forty years of access control has nothing to say about it.
Skills bundle instructions, scripts, and MCP servers into a single installable package. That convenience is also the attack surface, and in February 2026 someone used it 335 times in three days.
What this blog is for, and what it is not.
