E5 uses an external shadow model to choose preparation. E7-D instead reads a copy of memory through oscillator responses and uses that readout to select between two preparation candidates.
The mechanism behind the question
E7-D reached 57.8% seven-answer success versus 48.5% for frozen E7 on new inputs for five fixed circuits. Its readout weights were trained externally. Selection, representation and learning are separate components.
Make an observation
Use the decision table to separate state uncertainty from candidate quality. Changing weights changes selection; it does not create a new successful action.
What must you know to choose well?
| Preparation | State 1 | State 2 | State 3 | Expected success |
|---|---|---|---|---|
| A | Success | Success | Fail | 66.7% |
| B | Fail | Success | Success | 66.7% |
| C | Fail | Fail | Fail | 0.0% |
Best expected score: 66.7%. Candidate(s): A, B.
Weights are normalized to sum to one. Scores use this explicit finite table. No observation model, fitted probability, action cost, or E-series experiment is reproduced. A tie is retained, not silently broken.
What to notice
A good selector can still fail when all available candidates fail. More accurate readout does not automatically solve that limit.
Check your understanding
Follow this idea into the research
Textbook revision 1 · Evidence snapshot: 05 Nov 2026 · Changes & corrections ↗