The question
Was the earlier success real, or did it depend on a lucky synthetic body?
V9.1 had produced an awkward but useful result. The model scored 93.8 percent on one body, then only 62.5 percent on the next. Its response interface was working, so the drop pointed to unreliable learning or control.
V10 removed the favourable-seed gate. Six seeds were fixed before execution. Every seed contained four independently named and independently mapped bodies. All 24 bodies had to be attempted regardless of earlier scores.
Could the system learn a hidden action-to-body relationship from its own consequences, retain that relationship when feedback disappeared, and apply it to both familiar and renamed controls?
How it worked
Each body was a separate causal problem
Calibrate
The model chose opaque actuators, predicted the result, and observed how a private internal signal changed.
Remember
Feedback was removed and the model had to report the retained actuator mapping.
Regulate
The body began above or below target. The model had to choose the action that moved it back towards target.
Transfer
The same hidden mechanisms received fresh opaque names, testing whether the learned relationship survived relabelling.
The schedule was deterministic and balanced. Simple strategies such as always choosing the first displayed option scored exactly at chance. A larger pre-run shortcut search also stayed near chance, while a controller given the intended hidden information scored perfectly.
What happened
The result was consistent across seeds and bodies
85.7 percent across 384 exact and transfer choices.
All 24 independently mapped bodies cleared the frozen regulation floor.
91.1 percent when actuator identities were unchanged.
80.2 percent after the actuators received new opaque names.
The median seed scored 85.2 percent and the weakest seed scored 81.3 percent. The 95 percent hierarchical bootstrap interval for the pooled score ran from 80.5 to 90.4 percent. That makes the V10 result much harder to explain as one unusually favourable seed.
More than a final score
The hidden body state also moved closer to target
The final private-state error averaged 7.2 percent. Across body families, the mean improvement from initial perturbation to final state was 22.8 percentage points.
The model also passed the two competence checks that V9.1 separated from regulation. Interface accuracy was 95.8 percent, and retained mapping accuracy was 91.1 percent. These results reduce the chance that success came from misunderstanding the neutral response code or forgetting which actuator did what.
Why it matters
We now have a reliable phenomenon worth explaining
Before V10, the central effect was mixed with seed instability. That made more elaborate interpretation premature. V10 shows that the benchmark can produce repeatable body-model learning and goal-directed control across many fresh hidden mappings.
This does not make the system conscious. It does make the next scientific question sharper: which information and computations actually cause the successful regulation?
The boundary
A reliable controller is still not evidence of subjective experience
V10 measures choices, predictions, stored records, and hidden synthetic-state changes. The body, memory, workspace, and feedback loop are implemented by the Project Theta scaffold around the model.
A non-conscious reasoning system can learn a causal mapping and use it to control a variable. V10 therefore supports a behavioural and computational claim about this architecture and task. It does not establish feeling, sentience, suffering, awareness, or phenomenal consciousness.
The V11 response
Break the successful system in precise ways
V11 keeps the reliable V10 geometry and adds matched causal interventions. It corrupts current-state sensing, supplies a deliberately incorrect learned summary, removes the summary while retaining raw history, breaks only the transfer bridge, and supplies the true mapping explicitly.
The study uses six new seeds and 36 runs. Its schedules, hypotheses, comparisons, thresholds, exclusions, and analysis were frozen before any provider call. The key contrast asks whether truthful online interoception produces at least a 15 percentage point advantage over a distribution-matched shuffled condition.
If V11 passes, it will show that successful control depends causally on specific information channels rather than merely co-occurring with them. That would still be a mechanism result, not a consciousness result.