Adversarial registry. Summary in plain language

The question

Cogitate is the best-known test of rival theories of consciousness, and what made it useful was its format. In the Cogitate collaboration two rival theories wrote down in advance what they predicted, a neutral team ran the experiment, and the result was published whichever way it fell. We know of no one running that format for AI systems. Claims about machine minds come from the labs that build the models. The people who made the claims evaluate them, and everyone else argues about them afterwards on social media.

A small institution can supply the format. It needs a registry (a public list of pre-registered claims and their predictions) and a protocol saying who predicts, who runs and who checks. It also needs two camps willing to commit, and open models to run on.

Where to start

We start with the protocol document and one pairing, and leave software for later. The first pairing should be a live disagreement in which both sides have already published on open models and would each accept a result from the other's method.

The candidate pairs two groups who both worked on open models and read the same experiment in opposite ways. Anthropic's Jack Lindsey showed that a model can sometimes notice a concept planted in its internal activity and name it. Pearson-Vogel and colleagues, working on an open Qwen model, could read the planted concept out of the model's internal activity when it was asked about it, even though the plant was removed before the question. Lederman and Mahowald found that open models notice that something was planted far more often than they can say what, and argue that the content lingers without the model having introspective access to it. The test that separates them asks the model two kinds of question over the same planted state, one that asks it to look inward and one that only asks for an associated word. If looking inward recovers more of the concept, introspection reaches the content. If not, the content was only lingering. It is a single experiment, and it can be pre-registered.

The protocol came first, adapted from Templeton's adversarial collaboration rules, then the registry entry template (claim, theory, models, prediction from each camp, analysis, threshold, executor, date). Both were drafted on 2026-09-30, and the first-pairing proposal on 2026-10-30. Next we approach Eleos as neutral host or co-host, and after that the two camps. The experiment runs on open models, and the entry and the result are published together.

What the outcomes could be

• One entry is completed. Both camps signed up to a result, on open models, in public. The registry then exists, and the second pairing is easier to recruit.
• The camps decline. The registry then publishes the protocol and the invitation, and pre-registers the institution's own claims (from the other lines) as its first entries. A standard is on record even so.
• The format produces a result neither side likes, as it did in Cogitate.

What kind of research this is

Meta-science and institution-building, with almost no compute. The output is a protocol, a registry and completed entries. Of everything the group does, this is the most likely to make its name mean something. It also protects every other line from the "you evaluated your own claim" objection.

Timeline

The protocol and template take the first month, and the first pairing should be recruited within a quarter. The AMCS Models of Consciousness meeting in Copenhagen (12 to 16 October 2026) came before the protocol was ready, so the conversations happen at the Eleos conferences and the next AMCS meeting.