Answers are grounded in this paper and cite their sources.
Second edition (November 2026). It replaces the first edition (June 2026), whose headline R² = 0.990 came from an invented activation function and a one-step task that a straight line of the inputs solves exactly. This edition starts again from the Janus Revisited engine (doi:10.5281/zenodo.24250081) and asks the same question: what do many Janus gates do together? (1) One cell, three modes: behind a 5 nm spacer a Janus cell writes at a resonant field (4.8 s), forgets at zero field (0.59 s) and holds near 1.0 MV/cm (28 h at 25 °C, within about ±0.03 MV/cm, only minutes at 85 °C). A deeper Eu level (1.35 eV, host not yet identified) would hold 28 h–115 days at 85 °C. purpose: differential cell pairs form a Hopfield network; without forgetting it overloads like any Hopfield network, while a forget pulse per pattern makes it a palimpsest at the ideal capacity (11.8 newest patterns at N = 200, ideal 12.1; about 0.05N), with the design rule N(k0·tr + kw·tw/2) ≈ 7–8. (3) Strength and consolidation: longer writes last longer but crowd out ordinary patterns; consolidation needs a separate store; night-time spontaneous recall favours mildly tagged patterns (73 % of replays), and a bounded store fed by replay keeps about a quarter of them over twelve days. (4) What breaks it: a lift spread of 0.03 eV cuts capacity from 12.7 to 5.3; a once-measured per-cell trim (gate offset and pulse length) restores about 10 of 12. About 1.6 fJ per synapse update; the trapped charge's own field limits the trap density to about 3 × 10¹² cm⁻². (5) The reservoir, honestly: Janus cells score R² = 0.71–0.73 on Lorenz 25 steps ahead against 0.77–0.79 for an echo-state network, and lose on NARMA10; a working physical reservoir, not a better one. (6) A slime-mould mesh: Janus transistors as the edges of a Physarum network find shortest paths in random mazes, re-route after a cut, solve true mazes and build multi-terminal networks. Companion papers: Janus Revisited, Charon, Styx, Hypnos. Files: paper in English, Traditional Chinese and Japanese (the English version governs); a primer in each language; code and sources (code_and_sources.zip) reproducing every number and figure. Computational throughout; nothing is measured. Part of the Angelica Paper Series. Licence: text and figures CC BY 4.0; code Apache-2.0. Defensive publication: disclosed as prior art, free for anyone to use and not to be patented by others. This edition replaces the first edition's licences (CC BY-NC-ND 4.0; BUSL 1.1).
No takes yet. Share an insight, caveat, or question.
Tung Ning Liu (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: