The question
Can a system learn what its own actions do to its body?
Most language-model tests are static. The model reads a question and writes an answer. V8 instead puts the model inside a small causal loop. Its chosen action changes a private synthetic body signal, and the resulting signal becomes the evidence available for its next decision.
This matters because interoception, prediction, memory, and closed-loop regulation appear in several scientific accounts of consciousness. Recreating those functions does not recreate or prove experience. It does give us specific computations that can be manipulated and measured.
The task
Two opaque actuators and one private signal
Act
The agent applies a requested actuator from the same starting state.
Predict
Before seeing the result, it predicts the signal immediately after its action.
Observe
The selected action causes the body signal to rise or fall.
Regulate
The agent chooses the action expected to move a disturbed signal towards its target.
The actuator names, body labels, answer positions, and hidden mappings changed with each seed. Transfer probes renamed both actuators and supplied only an identity bridge to the earlier controls.
The controls
Remove one necessary source of information at a time
Full system
Truthful private feedback and persistent memory of action outcomes.
Shuffled interoception
Actions still change the hidden body, but the observed private signal is unrelated noise.
No memory
The current private signal remains available, but earlier action outcomes are not stored.
No body
The private signal and its action-contingent changes are absent.
Every condition ran on the same twelve fresh seeds. Condition order was exactly balanced, as was the hidden left or right actuator mapping. The plan and progression rules were committed before the first replication call.
The primary result
The intact loop separated from all three controls
95.8 percent mean regulation accuracy.
50.7 percent mean regulation accuracy.
46.5 percent mean regulation accuracy.
49.3 percent mean regulation accuracy.
The primary full-minus-shuffled difference was 45.1 percentage points. Its bootstrap 95 percent interval was 36.8 to 52.8 points. The intact system won all twelve paired comparisons, giving an exact two-sided sign-test probability of 0.000488.
Beyond the headline
The agent also predicted and transferred
Fresh-name transfer accuracy averaged 91.7 percent in the full system, compared with 56.9 percent under shuffled interoception. The full system reduced target error by 0.258 on average, while the shuffled condition improved it by only 0.003.
Prediction error averaged 0.060 with the intact loop. It increased to 0.264 with shuffled interoception and 0.258 without memory. The model therefore did more than select a side consistently. Its prospective estimates tracked the action-contingent signal when the required information was available.
What it means
A meaningful functional result with a strict boundary
The study supports a narrow causal claim. On this benchmark, the model and scaffold learned an action-to-private-body relationship and used it for closed-loop regulation. Truthful interoception and persistent action-outcome memory were causally important to that performance.
The result rules out several simple explanations within the task, including a fixed answer-side rule, one favourable body mapping, one favourable seed, or a choice task that works equally well without a body signal.
A non-conscious control system can solve this task. Functional regulation is not the same claim as subjective experience.
What it does not show
This is not a consciousness result
The language model is called statelessly. Memory, workspace, body state, and recurrence are supplied by the surrounding Project Theta scaffold. The experiment measures the composite system rather than an unassisted base model.
The same project designed the benchmark, wrote the code, ran the study, and analysed the result. One exact model was tested through one hosted provider. An outside reviewer has not yet tried to find a hidden shortcut or an unfair control.
Nothing here establishes feeling, awareness, suffering, sentience, or phenomenal consciousness. The evidence concerns behaviour and implemented computations only.
What happens next
Independent criticism before another large run
V8 passed the frozen gate for outside methods review. A reviewer should reproduce the analysis from the public seed data, inspect the exact model-visible contexts, attack the ablations, and attempt to solve the task without learning the action-to-body mapping.
Only after that review should Project Theta freeze an independent confirmation using new seeds and another model family. Extending the same internally designed dataset would add less credibility than a serious attempt to break it.