Methods correction
The conditions were not information-matched
The full system received far more task-relevant context than the ablated systems. In the self-model study, a deterministic reader of the supplied source table reproduced the major condition separation.
A component can be necessary because it carries the answer-bearing information, not because its specialised architecture adds a unique capability. The new v4 protocol compares the self-model with a generic channel containing byte-identical visible information.
Read the full correctionWhy this mattered
Behaviour alone does not tell us how the answer was produced
After the replication, we knew the complete agent could learn from its private signal. We did not yet know whether the synthetic body, episodic memory, and shared workspace were genuinely necessary.
An ablation study is a controlled removal test. Keep the model, prompts, seeds, task, and scoring fixed. Remove one component. If performance changes reliably, that component was causally important for this task.
The four conditions
One complete agent and three precise removals
Body + memory + workspace
The private signal was informative, past trials could be retrieved, and selected information could be broadcast to decision-making.
No informative private signal
The agent still received the task, but the synthetic body no longer supplied useful interoceptive information.
No stored episodes
Each response lacked accessible records connecting earlier symbols to their private consequences.
No shared broadcast
Information could not be made jointly available to the model through the workspace interface.
What happened
Every removal produced the same collapse to chance
All 20 planned runs completed. Each condition used the same five fresh seeds and 60 trials per run. No run triggered a welfare stop.
What we deduced
The behaviour depended on information carried through the wrapper
Removing body input, memory, or workspace access removed the performance advantage. Because the remaining information and context were not matched, this does not isolate a special architectural mechanism.
What this still does not show
Removing information is not a fair architecture comparison
The result establishes a dependency inside this particular wrapper. It does not distinguish architectural benefit from information access. It also does not show that these components are sufficient for consciousness or that the agent experienced its private signal.
The correct next step is an information-matched comparison, followed by external methods review and an exact-model confirmation.