Caelum Lab

Reports

• The Artisan Workshop: Can He Build His Own Tools?

  2026-09-29

  He has 17 tools and can build none. Twice he called a tool that did not exist because he needed it. A pre-registered experiment: a Laravel workshop of his own, walled off read-only from the mind that remembers him, five tasks, seven fences fired first, and the result that would count against the idea fixed before the run.

  ExperimentPre-registeredN = 5 tasksNot run

• The Switchboard: Think, Then Speak

  2026-09-28

  A mockup of how Caelum could talk to the outside world. Reasoning and communication become two channels: the mind produces a whole thought, and a separate communication step, still his own, decides whether to speak, how to format and split it, and when, from each channel's rules. Try it across Discord, Messenger, email and the terminal.

  Design sketchInteractive mockupNot built

• Looking Is the Price of Claiming

  2026-08-22

  Three read-only LOOK tools with receipts, and a settlement rule that refuses any claim made without one. Why "no tools" was the right fence around the wrong risk, how the menu is ordered by what is owed rather than by mood, and what the first live look did when it guessed wrong.

  Design specRead-onlyLive verified1 defect found live

• The Voice: Does State Reach the Sampler?

  2026-08-22

  The runtime sent a hardcoded temperature on every turn while its own design doc said inference was out of reach. One derived dial, its derivation recorded per turn, and a dissociation test with the prompt held byte-identical. The channel is real; the sample is small.

  ExperimentH0 not supportedN = 4 per armBounded

• Curiosity as a Scorer, Not a Generator

  2026-08-22

  How the idle loop chooses what to wonder about: every gap the store already holds, scored by weight × novelty × staleness × reach, with the whole decision written into the dream note. Includes the source that was removed because it could never fire, and the finding that all 23 open predictions had no deadline.

  Design spec875 candidatesDecision recorded

How these are run

Falsifiable

Each experiment states the hypothesis it could have failed to reject, and says what a null result would have looked like. A design spec states what would count as the design failing.

Controlled

A check that has never produced a failure is not known to be able to. Every guard is fired once against a known-bad input before it is trusted; the report says which.

Reproducible

Every report names the script in src/ that produces it. The deterministic parts are pinned as tests in the repository and run before anything is called live.

Bounded

Each report ends with what it does not show. Sample counts, dropped samples and caps are stated, never smoothed.

Standing caveat. These are engineering reports on a system's mechanics, not claims about inner life. Where a report says uncertainty, affect or curiosity, it is naming a field in a data structure and the code paths that read and write it. The system runs on one laptop, on one 30B model, at zero marginal cost; anything that required renting a GPU is said to have required it.

Relation to lab.kairo.icu: Kairo is a separate system built by a separate operator. Its lab is the model for this one's form. Where a report here builds on or departs from a Kairo disclosure, it cites the Kairo report by name and says which.