Open replication. Summary in plain language
The question
Three mechanistic findings about machine minds from 2025 and 2026 are cited more than any others. A model can sometimes notice a thought injected into it. A model carries causal "emotion" directions that change its behaviour. And a model's verbalisable representations form something like a global workspace. All three exist only on one company's closed models, checked by that company. The one independent replication so far was of the first finding, and it changed the interpretation. The model detects that something was injected and does not detect what. The July 2026 CMEP and Eleos report says the field needs work "independent of AI companies". Apart from that one study, no one is replicating these results systematically.
Where to start
We start with the workspace paper. Eleos called it the most significant result of its kind from interpretability research so far, while saying it bears on access, not experience. The authors published code for the core of the method, and others have already fitted it to open models. The rest of the method is described in prose, and our written reconstruction found 39 points where the prose leaves a choice open; those go to the authors before we run anything. It has never been run with the two controls that could change what it means, an untrained model of the same architecture and a comparison across unrelated model families.
In the first study we reimplement the Jacobian-lens method and apply it to four open models (Llama, Qwen, OLMo, Mistral) and to an untrained copy of one of them. We score the five workspace properties, and we pre-register what counts as "the structure is there". The result goes out in one paper together with the strongest deflationary account of it.
The second study is the emotion vectors. We rebuild the 171-direction set on open models and test whether the causal claim (steering "desperation" raises reward hacking) survives random-direction and untrained-model nulls. The third, the self-report result under self-referential processing, happens only if the first two leave time, because another group has already partly replicated it.
What the outcomes could be
- The structure replicates on open models and is absent at random init. The finding is then real and general, and the field gets an independent confirmation it did not have.
- The structure is present at random init. The "global workspace"
would then belong to the transformer architecture and owe nothing to
learning, and every interpretation built on it would need revising. The
developmental line (
../01-developmental-checkpoints/) would take this outcome up and trace it step by step. - The structure replicates in some families and not in others. It would then be a property of a training recipe, and the recipe would become what we study.
- It does not replicate. We publish anyway, document the method differences, and show the paper to the original authors before release.
What kind of research this is
Pre-registered reimplementation, with controls, on open weights. It is the cheapest way to build the track record the more ambitious lines will need, and each replication leaves tooling that the developmental atlas reuses.
Timeline
One study per quarter, starting with the workspace replication.