Independent researcher and programming educator. I read theory papers and measure what they promise.

Out of Time — an out-of-time calibration and stability evaluation of the tabular foundation models TabPFN-3 and TabICLv2 on credit default, under kill criteria written before any foundation model was fitted. Both are read on the origination vintages of Lending Club and Freddie Mac, beside a weight-of-evidence scorecard and a tuned gradient-boosting model. At the softmax temperature of 0.9 the packages ship, both models' mean probability on their own context rows is 0.656 to 0.736 of the realised default rate on Lending Club; at 1.0, observed over expected on those rows holds one. Every number in the paper traces to a recorded run through a claim ledger. Paper DOI 10.5281/zenodo.22999138 · Code DOI 10.5281/zenodo.22994434 · Protocol DOI 10.5281/zenodo.22803068

psi-inference — critical values and p-values for population stability testing in Python: the Yurdakul–Naranjo test of the PSI, the Population Resemblance Statistic, and the effect-size and overlapping tests of du Pisanie et al. The last three had no package in Python or R before it. With no population shift at all, the delta-method z-test published for the PSI rejects in 97.4–97.6% of samples at 10 bins and a nominal 5%; the README shows the simulation. DOI 10.5281/zenodo.22342343 · pip install psi-inference

beyond-dijkstra — the first implementation of the 2026 Duan–Mao–Shu–Yin directed shortest-path algorithm (arXiv:2602.07868), in Rust, gated bit-for-bit against a Python reference. It is not faster than Dijkstra at any size you can store: the faithful port trails by 38–182×, the most engineered variant by 1.8–2.2×. The audit is the result: two lines of the printed pseudocode crash on ordinary graphs, and the completeness clause of Lemma 3.7 is false as stated, with a 70-vertex counterexample and a hand proof that a one-line repair closes the hole. Note DOI 10.5281/zenodo.21478493 · Code DOI 10.5281/zenodo.21473443

logtwothirds — a verified Rust/Python implementation of the 2025 sorting-barrier shortest-path algorithm (Duan–Mao–Mao–Shu–Yin, STOC 2025), benchmarked honestly against Dijkstra. The engineered variant runs within 1.1–1.2× of Dijkstra at n = 10⁶–10⁷ — the smallest published gap — and the study documents why no practical input size crosses over. DOI 10.5281/zenodo.21284462 · pip install logtwothirds

Interests: credit risk modelling, algorithm engineering, AI, quantum computing, computational biology.

Website · ORCID 0009-0005-3612-830X · Medium · nicholaask@gmail.com