Charlot × Hiner · energy-native compute
The reproducibility you can afford
Collapse the crossings between the energy source and the compute state and efficiency climbs, but the computation inherits the raw noise of its carrier. You can buy reproducibility back without paying a conversion crossing: let the substrate's own dissipation settle into attractor basins, and commit the basin label, not the microstate. The barrier height you must spend to reject a given supply noise is a Landauer-analog, a thermodynamic price on a reproducible bit.
Energy efficiency
—
Supply noise σ
—
Label invariance
—
Bits / device
—
Free-energy landscape, an ensemble settles into basins · committed label = the basin the ensemble holds
The crossings spectrum: S0 separated → S5 identity
Landauer-analog for reproducibility
The tension in one line: reversible substrates dissipate least but build no basins, so they preserve noise; the basins that reject supply noise for free are made by dissipation, just above the Landauer floor. Illustrative model, the signs and the floor are the point, not the units.