Each essay starts with a real correction, makes one claim, names the evidence that exists today, and leaves the unresolved question visible.
6 min · 2026-07-23
A correction is evidence, not a rule
Why one edit should begin a reviewable learning loop instead of silently rewriting an agent.
6 min · 2026-07-23Why AI needs scoped guidance, not more memory
Recall is useful. It still does not decide what should influence the task in front of an agent.
6 min · 2026-07-23The smallest honest loop for agent learning
A learning claim needs evidence, a change, a comparison, and a way back.
6 min · 2026-07-23Learning needs an undo button
A system should not earn the right to change unless it also knows how to recover.