Who May Sign? Typed, Calibrated Decisions on Company Representation Rules in the Austrian Commercial Register

Know-your-business onboarding requires deciding who may legally bind a company. Commercial registers publish this as free text, but only a few also publish a machine interpretation; the Austrian Firmenbuch does not. We present SignRule-Decide, a 4B-parameter typed decision model that reads a register's representation text together with the registered roles and answers fixed questions (can a managing director act alone, which coalitions of office holders can bind the company, how many signatures are needed) with calibrated probabilities and an abstention option, without generating text. All training labels come from the registers themselves: the Norwegian register's rule interpreter and the Austrian courts' per-person representation codes, extended for joint powers by a reviewed table of standard register wording; no synthetic examples or model-generated training labels are used. On 400 held-out Austrian free-text extracts, the model agrees with a reference consensus of two independent AI assistants on 99.3% of coalition questions, 97.7% of minimum-signer questions and all managing-director-alone questions, against 80.2% minimum-signer agreement for the phrase table; a fine-tuned 300M-parameter encoder comes within one point, while our model ranks its own confidence better. On Norwegian texts the official interpreter could not read it reaches 93.0% on the rule type, and on Danish texts never seen in training 92.5% on coalitions, although its abstention thresholds do not transfer to the unseen register. We report pre-registered hypotheses including those not met, describe the AI-assisted construction of the reference sets, and release code, weights and aggregate results. Code and aggregate results: https://github.com/aliildan/signrule-decideModel weights: https://huggingface.co/aildan/signrule-decide-4bSoftware DOI: https://doi.org/10.5281/zenodo.23224891