LEARNED FIELDS · EXACT PHYSICAL TEACHERS · COVERING GEOMETRY
Follow a learned flow back to its physics.
Thirteen trained GPU runs, exact reduced Gibbs states, a specified width limit, constructive cover calculations and cuQuantum circuits. The finite physical models are tied to actual clean dessins; they do not replace their infinite locally gentle algebras or prove GRH.
Loading preserved numerical records…
The same 128 points are tracked through saved model evaluations. Progress is sampler time, distinct from the RG scale. Diffusion paths are stochastic; flow paths use Heun integration. The 3D view adds time as a coordinate.
Quantum diffusion and a learned coherent state flow
The seven-spin teacher is a quantum Gibbs state on the genus-two K₂,₅ dessin incidence graph. Exact partial trace over five white spins leaves a two-qubit density matrix. Classical samplers above learn its Z-measurement law; the separate four-qubit purification model below learns a coherent state-preparation curve using full target-state access.
The channel preserves positivity and trace. Its inverse is not a positive map on all states. The trained purification uses two system and two ancilla qubits; this changes subsystem entropy without violating unitarity. Grid fidelity is measured, not a uniform theorem.
Inspect all coherent-flow circuit states · Complete coherent-flow circuit and training code · Diffusion and flow models · Exact reverse-mixture control
A width limit with specified assumptions
Independent one-qubit features φ(x)=√2⟨Z⟩=√2 cos(w·x+b), Gaussian w and uniform phase b give K(x,y)=exp(−‖x−y‖²/2). Only the readout is trained in this experiment. This is a different limit from enlarging the four-qubit generative networks.
Max-plus arithmetic gives a piecewise-linear operation. A tropical variety or HMS comparison additionally needs a valued algebra, geometric data and a categorical construction. Over a valued field, leading coefficients can cancel: val((1+T)+(−1+T))=1 although both summands have valuation zero. In characteristic two the sum is zero. Finite-field algebraic coefficients do not supply ordered probabilities or ordinary gradient calculus.
Exact coefficient search and reversible oracle
Executable width/RG experiment · Three tested tropical/hybrid accelerator references
Free-group covers and actual dessin modules
F₂ is a free group; 𝔽₂ is a coefficient field. The theta graph has fundamental group F₂ and a degree-three tree cover. Its contracted Cayley graph is degree four. The genus-two/four selectors instead give the degree-five/nine cover trees of their contracted dessins. Clean white bivalent vertices are suppressed in these tree drawings. Filling the dessin faces adds surface relations: the universal cover of a compact genus-two surface is not this tree.
Module calculations use the genuine infinite genus-two polynomial dessin algebra. B is its one-arrow string at vertices 0 and 4; it has Ext¹(B,B)=1 and an explicit nonsplit self-extension. The graph-only deck-Hom example is separately labelled and is not substituted for this algebra.
Pants gluing and quantum games
The dual-number Frobenius algebra gives exact oriented 2D TQFT gluing. Multiplication is not a unitary gate; its physical realization is heralded. Conformal/string sewing needs further analytic data. CHSH is a finite two-qubit game, not the correlation separation in MIP*=RE. The latter’s negative answer to Connes embedding is an external theorem.
Executed cuQuantum surface circuits
cuStateVec and cuTensorNet ran on the RTX4090 in the approved isolated Linux environment. These larger circuits are untrained finite matter circuits, distinct from the trained four-qubit feature models. Small cases have independent NumPy reference checks; large cases have normalization and inverse-circuit echo checks.
GPU circuit runner · Resolved conda lock · Complete numerical and exact record · Remaining implementation and research work
Arithmetic frontier: inspect the refined cube-factor sector ↗
Fresh adaptive theta/Mellin sampling and certified 10⁻²⁴ readout training ↗
The new laboratory offers a fresh adaptive sampler and a preserved fixed-label replay as separate certificate sets. Both explicitly budget the Fourier midpoint-target perturbation. The older high-precision demonstration below retains its original conditional status.
HOLomorphic READOUTS · APPROXIMATION VS ZERO LOCATION
Train an analytic family with an explicit error budget.
These are actual completed primitive Dirichlet L-function readouts at conductors 3, 4 and 5, with 120 gradient updates each. The accompanying proof separates optimization, sample error, Fourier aliasing and Taylor tails. The numerical demonstration uses high precision, not a validated interval oracle: its displayed uniform budget is conditional on the assumed evaluation/rounding bounds.
Arbitrary accuracy is proved for fixed compact analytic targets with independently certified constants and oracle error. The infinite-feature population readout converges under an ℓ² initialization hypothesis, with a source-condition polynomial rate. Its slow eigenvalues tend to zero: there is no width-independent exponential guarantee. These results do not supply real-rooted arithmetic approximants or prove GRH.
Full derivation and proposed objectives · Exact rational growth/tail certificates · Executed family experiment
12 ACTUAL TRAINED RUNS · LOCAL STRATEGIES · CHECKABLE MOMENTS
Learn a quantum game, then inspect its certificate boundary.
CHSH, two tilted CHSH objectives and the magic-square game. Three seeds each; one tilted run gets stuck at the classical optimum. Select its preserved checkpoint rather than assuming every training run succeeds. All players act locally during play.
What this says about Connes embedding
Every displayed strategy is finite-dimensional and belongs to both Cqa and Cqc. Its word Gram matrix is PSD by construction; a vector state need not be tracial. This does not witness their separation. The negative embedding result is the external theorem MIP*=RE, not a conclusion of these simulations.
Useful training tasks include finite strategy optimization, NPA constraint proposals, exact sum-of-squares certificate search and separately constrained tracial microstates. Finite successful fits imply neither general game decidability nor arithmetic positivity for GRH.
MIP*=RE, pinned v4 · NPA hierarchy · Proofs and limits · Actual training source
HOMOTOPY GLUING · RIBBON DATA · ACTUAL DESSIN FAMILIES
Which loops disappear when a face is filled?
HoTT’s hubs-and-spokes construction attaches a disk with a fresh hub and a continuous circle-indexed family of paths. The finite drawn rays below sample a two-dimensional region. A graph star alone is not the disk; deleting its vertex is not a punctured disk.
The punctured genus-two dessin surface has free group F₄. Capping its face adds [a₁,b₁][a₂,b₂]=1 and changes the covering problem. Graph covers are trees; filled-surface covers include lifted cells. Homology alone misses this: H₁ is Z⁴ in both cases.
Cyclic order, Belyi data, gentle zero relations and polynomial/formal/operator completions remain additional data. Homotopy gluing alone does not produce a Hom-controlled module functor, an AR mesh, HMS or a spectral determinant identity. No HoTT/univalence axiom has been added to Lean4.
Official HoTT book, pinned edition: §§6.6–6.8 · Dessin calculation and proof · Exact dart checks