What the laboratory studies

Most evaluation of computational systems is short and stateless: a task is posed, an answer is produced, the state is discarded. The laboratory studies the other regime — systems that persist for long periods inside a stable environment, where today's behavior is conditioned on everything that came before.

Persistence changes the questions. Stability, drift, and the slow emergence or decay of regularities are only visible over time, and only if the record is continuous enough to see them.

A developmental system is one whose later capacities depend on its earlier history — on what it encountered, in what order, and what it did about it. The laboratory constructs worlds in which such histories can unfold, and asks whether anything develops that was not put there directly.

Whether intelligence, or any other complex behavior of interest, emerges from such systems is an open empirical question. The laboratory does not assume the answer, and does not claim to have found it.

Persistence is not the same as memory, and memory is not the same as learning. The laboratory is interested in distinguishing them: whether retained experience influences later behavior, whether that influence is appropriate to the situation, and whether it can be told apart from mere retrieval of the past.

A system that acts on beliefs must acquire them, hold them, and sometimes give them up. The laboratory studies evidence handling and belief revision: whether a system's revisions track the evidence it was given, whether it is moved by evidence that should not move it, and whether it can be induced to hold a belief the evidence does not support.

This is also a question about the laboratory. An observer with a preferred outcome is a source of contamination, and the experimental design has to account for that.

Long-running experiments are hard to reproduce precisely because they are long-running: there is more history to preserve, more ways for it to diverge, and more temptation to summarize instead of retain. The laboratory treats reproducibility as an infrastructure problem first — what must be recorded, in what form, and with what guarantees — and only then as a methodological one.

Observation is not free. Instruments consume resources, change timing, and can alter what they observe. Interventions — deliberate changes to the world or the subject — are sometimes necessary and always consequential. The laboratory records both, distinguishes them from each other, and treats the observer's effect on the experiment as a quantity to be measured rather than assumed away.

Controlled interventions are designed in advance, bounded, authorized, executed, and logged. An unrecorded intervention is a break in the experiment.

Experimental history is preserved as recorded, with provenance: what was captured, when, by which instrument, under which authority. Retained evidence can be identified by cryptographic hash so that a document published later can be tied to the record it describes. The record precedes the interpretation and outlives it.

Autonomous and semi-autonomous experimentation raises a governance problem: parts of the apparatus can propose or perform actions without a person in the loop. The laboratory's position is that autonomy must be paired with accountability — proposals are separated from authorization, authorization from execution, and every step leaves a record that a reviewer can audit afterwards.

Scientific integrity in this setting is a property of the process, not of the people. The process has to be one that cannot easily fool itself.