Flip Observer: 512 → 1,024
Offline synthetic benchmark · fixed classical weighted readout · exact channel-order reversal
Interpretation guardrail: the second bank repeats the same 512 underlying signal channels with independent synthetic measurement noise/dropout. This tests repeated-readout reliability, not 1,024 independent inputs or any biological/quantum mechanism.
Held-out normalized readout error
Bars show the mean across seeded trials; error bars show ±1 sample SD. Lower is better.
Single bank: bank A only. Uncorrected: average A with still-reversed B. Calibrated: choose identity/reversal from separate full-vector calibration then fuse. Oracle: use the known inverse as an upper bound.
Calibration and failure checks
What the detector can claim
It sees known calibration vectors during a separate phase and distinguishes only identity from exact reversal. A scalar output wᵀx is not used to infer the mapping; generally it cannot identify an unknown 512-channel permutation.
Calibration is not retrained on the held-out mixtures. All signals and values here are dimensionless and synthetic.