Publications

Technical papers, system working papers and reference documentation produced by the HFThot Research Lab. Every publication ships with the open-source code and notebooks needed to reproduce the figures.

September 2026 · Preprint · 40 pages

ThotBook-AI: CPU-Only Language-Model Training with Physics-Informed Verification, Membership Auditing and Versioned Corpus Provenance

ThotBook Research · Independent Researcher

A Rust system that trains, serves, verifies and audits a personal language model entirely on CPU, no GPU and no external API. Every pipeline stage reads and writes a versioned Delta Lake table, so the corpus that produced a model is recoverable at a numbered version with a commit timestamp; a self-contained reverse-mode autodiff engine solves the differential equations appearing in ingested documents rather than only quoting them.

ThotBook-AI Rust CPU-only Delta Lake PINN Membership Inference
September 2026 · Preprint · 16 pages

Polarway: Railway-Oriented Data Processing on a Lakehouse Substrate

ThotBook Research

Applies Railway-Oriented Programming to analytical pipelines built on Polars: every transformation returns an explicit success/failure value, so error handling composes rather than being bolted on through exceptions. Seven original crates add a Delta-Lake-backed lakehouse with ACID transactions and time travel, network-native streaming sources, and a Spark DataFrame compatibility layer. On 5M rows, 3.4–8.7× faster than PySpark on scan/aggregation/join/windowing.

Polarway Rust Polars Delta Lake Railway-Oriented
pysic-rs
September 2026 · Preprint · 9 pages

pysic-rs: A Validated Mathematical-Physics Engine in Rust, and What Grading It Against Closed Forms Revealed

Mel Alvarez · HFThot Research Lab

A CPU-only mathematical-physics library covering special functions, ODE/PDE solvers, quantum mechanics, general relativity, gauge theory and topology, built around one commitment: every routine is graded against a closed form or tabulated value rather than demonstrated on a plot that looks plausible. Applying that commitment exposed eight incorrect routines in our own code — a Dirac propagator whose norm reached $3.6\times10^{57}$, an eigensolver returning the matrix diagonal, Bessel functions returning $+50.6$ for a quantity bounded by 1, and a vacuum spacetime reporting non-zero Ricci curvature. The paper analyses why the existing tests could not have failed, most starkly a Ricci-flatness test that passed a constant metric to a routine that differentiates its argument.

Pysic-rs Rust PyO3 Dirac ADM Wheeler–DeWitt Validation
September 2026 · Course · 30 pages · 34 solved exercises

Mathematical Prerequisites for Computational Physics

HFThot Research Lab · 10,99 €

Complex analysis and special functions, floating-point cancellation, spectral linear algebra, ODE stability and stiffness, Fourier and Strang splitting, asymptotics and turning points, Clifford algebra, and constrained systems — every chapter ending on the invariant that would have caught the corresponding error. In French, with all 34 exercises fully solved. The thread running through it is one question: how would I know this result is wrong?

Course Analysis Numerics Asymptotics Clifford ADM
May 2026 · System paper · 18 pages

Optimiz-rs: A System for Composable Numerical Primitives in Quantitative Finance

Mel Alvarez · HFThot Research Lab

A JAX-style system paper covering the seven primitive families exposed by Optimiz-rs v2: hidden Markov models, backward stochastic differential equations, McKean–Vlasov diffusions and mean-field games, path signatures, Hawkes processes, robust drift estimation and persistent topology. Includes complete derivations, theorems, pseudocode, honest single-threaded benchmarks, and an empirical validation of Sznitman's $\mathcal{O}(N^{-1/2})$ propagation-of-chaos rate.

Optimiz-rs Rust PyO3 BSDE McKean–Vlasov Signatures Hawkes TDA