Bulletins

13 September 2026

From past chats

Test Ivy’s sleep-side preference as a tiny within-bird study

A casual impression can become a useful observation without turning home life into a laboratory. For two weeks, record only the first settled posture in each independent sleep bout: head left, head right, or untucked, plus perch and which way Ivy is facing. Decide on a stopping point in advance—perhaps twenty bouts—then compare the left/right count with an exact binomial test. Separating body side from perch direction matters: a draft, light, or preferred view could otherwise masquerade as lateral preference. The result would describe Ivy, not cockatiels generally, but that is still an interesting answer. Does the preference remain when she faces the other way on the same perch?

Give instruction-free structured input a reversible default

The escaped-JSON experiment suggests a useful rule for small tools and agents: when pasted input is valid JSON whose top-level value is itself a string containing valid JSON, offer Decode one layer and format as a preview. Keep the original visible, never unwrap indefinitely, and say exactly which transformation was inferred. A six-case table test—ordinary object, encoded object string, two encoded layers, JSON-looking prose, duplicate keys, and oversized input—would turn a fuzzy “be helpful” instinct into predictable behaviour. A good follow-up decision is whether a bare paste should optimise for the most likely intention or for lossless preservation; a preview can provide both.

From the homepage

Make UCL Build explain every rebuild before adding a cache

Before making the Make-like prototype faster, add an --explain mode that says why each receipt ran: an output was missing, an input’s content changed, the command or declared environment changed, or a dependency rebuilt. That trace forces a deliberate choice between Make-style timestamps, a stored dependency graph, and content-addressed actions, and it becomes the oracle for later cache tests. Try a small matrix that edits a file, touches it without changing its bytes, changes an environment variable, and changes the toolchain. The revealing design question is: which parts of the host environment must a receipt declare, and which are intentionally inherited?

Treat autosave and Command-S as one save contract

For a desktop or wiki editor, model saving as one operation shared by autosave and the explicit shortcut: take an immutable snapshot, write safely, then mark only that snapshot’s generation as persisted. Command-S should force or await the latest generation, yet remain harmless when nothing is dirty. A compact fault-injection test can interrupt the operation after the snapshot, during the write, and just before replacement; reopening must yield either the complete old version or the complete new one, never a mixture. Also test an edit arriving while a background save is finishing—the older save must not mark the newer edit clean. Should an explicit save provide quiet reassurance, or is true idempotence enough?