FINITE PHYSICAL TEACHER → TRAINED HYBRID → MEASURED EXPLORATION
Tokens, concepts, trajectories.
Seventeen graph tokens pass through three-qubit feature circuits and learned attention. Inspect heldout predictions, learned steering, an actual branching search, continuous GFlowNet trajectories and an error-correcting channel.
Verifying the saved experiment…
RECOVERY INSIDE THE SAVED INFERENCE PIPELINE
Ideal quantum feature-memory layer
All 40 heldout examples, unchanged saved heads. Independent X memory noise acts after feature preparation and before measurement. Level 0 uses 3 physical qubits/token; ideal recovery levels 1 and 2 use 9 and 27. No noisy encoding gates, syndrome/correction errors, phase-error protection or fault-tolerance claim.
The literal tensor panel replays the saved quantum-hybrid heldout pass. The comparison selector changes the measured model metrics. Padding remains masked even when circuit feature values are nonzero.
Reproduction, source scope and measured limits
Run .tools/book-quantum-cpu/bin/python examples/quantum_tokengt/check.py and .tools/book-quantum-cpu/bin/python examples/quantum_tokengt/train.py --run NEW_UNIQUE_RUN. New runs preserve prior generations. Primary sources and exact read scope: source manifest. Methods and reproducibility.