Measurement records

What we measured, and what we got wrong first

Every organ we ship by default has to earn it — in public, against a specification frozen before the run.

Each record below contains the pre-registration and its amendments, the corrections we had to make, the data exactly as it came out, the instrument that produced it, and the verdict. The mistakes are in there on purpose. A record that hides them is not a record.

Records

Physics E1 — does recall earn its keep?

Pre-registered 2026-08-01 · battery run 2026-08-02 · verdict 2026-08-03

A creature is given a handful of experiments in a physics environment and must state the law. Four conditions differ only in how the observations reach it: not at all, carried in the prompt, carried and also recalled, or recalled alone across a session boundary. The question is whether the memory organ adds anything.

  • Specification, amendments, and three corrections — unedited
  • Screening census over the whole identifiable environment population
  • Battery data, instrument telemetry, gate and drift stamps
  • Driver and analysis code, so the numbers can be recomputed rather than trusted

Open the record →

Uptake E2 — is it the surface?

Specification published before firing 2026-08-03 · battery 2026-08-03 · verdict 2026-08-04

E1 found a memory organ delivering its payload in full while the creature recovered nothing from it. E2 holds the content fixed to the byte and moves only where it sits: the same text is unusable in the system prompt and fully usable in the user message.

First record specified in public before the run

How to read these

Where a record names a specific model or command-line tool, it is reporting one surface, one injection shape, and a small number of runs. It is not a ranking of models, and it will not support being read as one.

Pre-registration means the specification was frozen before the measurement ran, and every later change was argued in writing against people obliged to push back. For this first record the ordering rests on our working repository's history, so it is pre-registered but not independently verifiable. From the next record onward the specification is published before the run, which fixes that at no cost.