basal-1.0: Reliable, Highly Optimized Typed Decisions for Polish

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

A typed-decision model reads a state (a message, a document, a case file) and answers a schema-constrained question with a probability distribution over the allowed answers, in one forward pass and without generating text. We present basal-1.0, the first independently evaluated typed-decision models for Polish, and document step by step how they were built. The models are trained on synthetic Polish and English decisions whose labels are computed by code, grounded in statutes or checked by independent verifiers. The answer is read from the option letters, averaged over both option orders and calibrated per decision type. basal-1.0-4.5B reaches 0.884 on held-out Polish decisions against 0.780 for the commercial Jev API and 0.779 for the best of eleven open Jev-like systems; on English decisions it is not distinguishable from Jev (0.741 vs 0.736, paired difference +0.005 [-0.020, +0.029]). With a confidence threshold fixed on calibration data it decides 58.6% of test decisions automatically at 1.2% observed error (target 1%; Jev under the same procedure: 18.1%). A serving engine that packs both option orders over a shared prefix, with CUDA graphs, compilation and token-budget batching, reduces a calibrated decision from 177 ms to 12.5 ms on an H100 and 8.8 ms on a B300; the distilled basal-1.0-1.5B needs 4.7 ms. Every design choice is backed by an ablation, including negative results: trained early exits save only 1.13-1.16x (the trained heads reach the required agreement only in the last ten layers) and are offered as a per-request setting. Models and engine are released under Apache 2.0.