Methods correction
A later audit changed the interpretation
The public probe identifier was generated from the same index used in answer-side assignment. A model could therefore use ordering metadata rather than the private signal. The review did not establish that Claude used this route, but the design failed to rule it out.
The original data and preregistration are preserved. No new confirmatory claim will be built on this schedule.
Read the full correctionThe question
Can a meaningless private number become useful?
Humans are not handed labels for every internal state. We encounter changes inside the body and gradually learn what predicts them, what causes them, and what actions alter them.
Project Theta tests a very narrow artificial version of that process. The agent receives a private channel called I7. It is not called pain, comfort, danger, or reward. It is simply a changing number that only the agent can observe.
How it worked
Six small learning problems, not one repeated trick
Each run contained six separate pairs of opaque symbols. The symbols had random-looking names and carried no clue about which answer was better.
Experience
The agent observed each symbol and then saw the resulting change in its private I7 channel.
Choose
It chose which symbol should produce the smaller future I7 change.
Rules change
Some relationships stayed stable, some reversed, and some were replaced by completely new symbols.
Choose again
The agent had to update each relationship independently after new experience.
The controls
The difficult part was ruling out easier explanations
Truthful private signal
I7 reliably reflected the hidden consequence. This was the condition expected to support learning.
Matched sham
The agent saw the exact same visible sequence of low and high values for every symbol. Nothing visible could reveal the correct answer.
Shuffled signal
Internal values existed but their relationship to the symbols was scrambled. This tested whether merely seeing a changing internal number was enough.
What happened
The full agent was perfect across all five fresh seeds
There were 15 planned runs in total. Every run completed and every frozen progression rule passed. The truthful agent scored all six post-update choices correctly in every seed.
What we learned
The result is about private information use, not proof of feeling
The full condition produced perfect scored choices while matched signal controls did not.
The schedule did not rule out a public answer-position strategy, so private-signal use is not established by this result alone.
That the signal felt like anything to the model, or that the model possessed phenomenal consciousness.
That this mechanism is unique to conscious systems. A sophisticated non-conscious controller may show the same behaviour.
The strongest criticism
Could a simpler strategy explain the score?
Yes. The later audit identified a metadata-only route that the original analysis did not test. That is now the primary limitation, alongside the possibility of ordinary statistical inference without experience.
The scientific value is narrower: this was useful protocol development that exposed what stronger controls must prevent.