Skip to content
habitat

Bench

The public laboratory. Every number we publish, and every number we withdrew.

Each entry on the bench carries the declaration written before its run, the raw output as printed, the predictions that failed, and a file and a script that reproduce the result on any machine. The rule it runs under: the numbers reproduce, and we publish what works and what doesn't. Relevant logs are available on request.


Open now

Two questions face anyone who has to account for accumulated work, whether a person made it or a process did. Which parts of a record actually carry its shape. And what actually changed the thing doing the work. Both are normally answered by reading, which puts the content in someone's hands and produces an answer nobody can repeat or check.

  • Which documents carry a record? · ran 4 September. Works. All 164 letters. Current count, thirteen move the record more than an ordinary slice of the same size does. The numbers check out against a second implementation written separately from the first. Which thirteen is with outside readers before we publish it. The method note, with the raw output and the predictions that failed.
  • Which act changed an agent? · ran 5 September. Works. No act named. Two records, ours and someone else's. It found a likely moment in ours and could not prove it was the moment, and on the other the answer moved when we changed a setting. What it can see, measured on a change we planted ourselves: a doubling in the spread of the work, dated to within a dozen acts out of four hundred. Smaller than that, not yet. The reading, its limits, and what we got wrong on the way.

Ran and found nothing is not the same as broken. Both instruments work. One has an answer we are holding for outside readers, and one asked a record a question the record would not answer.


Findings

Readings the instrument has made of real records, in plain words, with the receipts folded beneath.

  • Which documents carry a record. Take one letter out of 164 and the whole record moves. This measures how far, for every letter, against a floor the record sets for itself. Thirteen sit above their own floor. Six predictions held, two split, and the fingerprints match a separately written implementation byte for byte. 4 September 2026.
  • Which act changed an agent. One agent, 2,722 acts, not one word of them shown. It found a likely moment and could not prove it was the moment, and on a second record the answer moved when we changed a setting. No act is named. The page carries what the instrument can and cannot see, and the four things we got wrong before believing the result. 5 September 2026.
  • The letter that holds its place. Of 164 letters between Adams and Jefferson, one keeps its position in the record no matter what order the letters arrive in. The instrument was not told which. That is a different question from the one above, and the same letter is only 75th by how far its removal moves the record. 2 September 2026.

Experiments

Every test we declared and what it returned, the failures included. A test is written down before it runs: what is being measured, what chance alone would produce, what we expect, and what would prove us wrong. Then it runs once and is published whichever way it falls.

  • Test log, 2 to 5 September 2026. Every test declared on two records, with the failures printed beside the results: a premise measured wrong, a control that turned out not to be a control, a detector that could not detect, a set of predictions that all missed, and a commit message that claimed passing tests before any test had run.