MEMORY · ACTION · UNCERTAINTY
Four learned records. One changing world. When the cue stops helping,
is it worth asking a question before choosing a memory?
01 / THE RECORD BANK
Rings show learned operators. The bars show the questioning needle’s belief before ordinary feedback.
03 / DOES ASKING PAY?
| Selector | MSE | Queries | Charged |
|---|
Charged loss = prediction MSE + price × questions / rounds. All four receive the same ordinary feedback after predicting.
WHAT THE NEEDLE DOES
Each record predicts a response to 16 possible probe directions. The needle asks where the records currently disagree most, then updates its belief from the answer.
It asks only when uncertainty about the current output exceeds the question’s price. This is a designed rule, not a learned curiosity system.
WHAT IT HAS TO BEAT
Feedback only tracks which record explains recent outcomes. Fixed probe uses a direction chosen once from calibration. Random probe samples the same direction bank.
All share the record bank, noise assumptions and switching model. Probe policies share the same gate and one-query-per-round ceiling. Actual queries and computation can differ.
WHERE IT CAN FAIL
Try expensive questions, rapid switches, or identical records. Asking can waste resources. Pure rotations offer no preferred direction in the exact model.
This tests retrieval of familiar rules. Learning new rules, protecting memories during writes, and choosing its own goals remain open.
| World | Feedback only | Fixed | Random | Needle |
|---|
Mean charged loss, 1,200 rounds × 20 seeds per world. These are the committed protocol settings; sliders above affect only your live run. See protocol, interpretation and seed-level receipts.