Experiment stories
The results, including where they failed
Project Theta publishes positive outcomes, failed gates, discovered shortcuts, and repaired protocols together. That is how a development project becomes a credible research programme.
See the loop
A dot with a body, a memory, and a next decision
This is a diagram of the experiment, not a recording or a claim that the model literally looks like this. Watch information move from an event, through the synthetic body, into memory and the next choice.
The embodied agent approaches an unfamiliar cue.
The body-learning result held across 24 fresh synthetic bodies
V10 tested six fixed seeds, four independently mapped bodies per seed, and 1,056 model decisions without favourable-seed replacement or early stopping.
The interface worked. The result did not hold.
V9.1 fixed a genuine response-interface problem, passed on one synthetic body, then failed its frozen reliability rules on the next fresh body.
Can an AI regulate a private synthetic body signal?
The intact agent learned how its actions changed an opaque internal signal, then beat three ablated controls across twelve fresh paired bodies.
Why we stopped the strongest-looking experiment
A review found a perfect answer-position shortcut and showed that the full system received useful information that the controls did not. Here is what that changes.
Could an AI learn from a private signal it could not name?
The original private-signal result, its controls, and the later schedule limitation that prevents confirmatory interpretation.
What happened when we removed memory, workspace, or the body?
The ablation changed performance, but it also removed task-relevant information. The article now explains why that matters.
Next stage
Find which mechanisms cause the reliable result
V10 passed every frozen reliability rule across six seeds and 24 independently mapped bodies. V11 now intervenes on the information channels that might explain that success.
- Matched state corruptionTest whether truthful current-state sensing is causally necessary.
- Summary interventionsCompare truthful, incorrect, and absent learned summaries while preserving raw history.
- Transfer bridge controlBreak only the identity link used by renamed-actuator probes.
- Explicit mapping controlCheck whether the task remains solvable when the hidden mapping is supplied.
The present result is behavioural and computational evidence from an internally designed benchmark. It is not evidence of phenomenal consciousness.