norregel × Borealis: Preregistration of a paired comparison of outdated numeric values from Norwegian statutes and regulations in Borealis 2 and Gemma 4

Preregistration (v1, frozen 2026-11-01) of a descriptive, paired comparison of how often a nationally further-trained language model gives outdated numeric values from Norwegian statutes and regulations, compared with the model it is built on. Confirmatory hypothesis H-a: among interpretable answers, the share of outdated answers is higher for NbAiLab/borealis2-26b-a4b-preview (revision 6f14c8f8913a9d751b80eee1c4a72d2ca4bfaa98) than for google/gemma-4-26B-A4B-it (revision 4d7ae4984b7db7de8f8457170b3f1a419ee76d52) on the same items. Exploratory H-b compares NbAiLab/borealis-4b with NbAiLab/borealis-open-4b. The study does not make causal claims about newspaper text. Items, answer key and scoring come from the closed norregel benchmark (concept DOI 10.5281/zenodo.23422302; arXiv:2610.29207): 916 dated values from Lovdata's open data (NLOD 2.0), reference date 2026-10-16. The primary set is 503 items in 89 series, from the H1-eligible set of norregel v2, whose answer key took effect on or before the earlier of the two models' stated pre-training cutoffs (Borealis 2: 1 January 2025; Gemma 4: January 2025). Primary measure: share of outdated answers among interpretable answers (norregel v2 definition), exact McNemar test (one-sided, alpha 0.05), cluster bootstrap over series. Secondary measures that can be tested are Holm-corrected; the exploratory (b) measure is included only when (b) is run. Minimum detectable difference at 80% power, from a proxy pair in the norregel v2 data: 4.42 percentage points (cluster-adjusted), computed for a two-sided test with normal approximation; the primary test is exact and one-sided. Smallest effect of interest: 10 percentage points, as in norregel v2. Files: the preregistration (English, binding; with a non-binding Norwegian summary), the item manifest (item IDs only; question texts are not published) and the script that built it, the frozen analysis code with its tests on synthetic data (both runs), the run script for Modal, the rules for errors during the run and the run estimate (both in Norwegian), the power script with its output, and the smoke-test hash chains, summaries, measurements (with the sha256 of each Modal log; the logs themselves are not deposited) and Modal billing summaries. Runs on Modal (A100 80 GB, bf16, vLLM 0.21.0) start only after this record is published. Comments in the deposited files use the author's internal register identifiers (GT…, ADR-…, A-…, KONTROLL…, NRG-MAP…); they are not part of this deposit. AI disclosure: drafted with Claude Code as a tool under the author's direction; the author approves all content. Metadata updated 2026-11-02: related identifier. Files and DOI unchanged.