Notes
Notes on retrieval, evals, observability, and the engineering that starts once the demo is the easy part.
Naming a tell is what kills it, so any list of AI words starts dying the day it ships. Underneath sits the failure no list touches: the draft commits to nothing. I built an editing skill for that one.
Why I leave the future to astrology and reach for reference classes, premortems, and calibration logs instead. Disciplined uncertainty in plain text.
A short note from Israel on what repeated alarms do to attention, engineering judgment, and team habits, and which working practices make interruption easier to absorb.
Repair loops, small diffs, test trust, and how to get CI back to green without trashing the codebase.
SYNAPSE was a 2025 framework for AI agents that adapt their own success criteria via MCDM. The deterministic-control-plane idea later shipped as Bernstein.