Building on the extraction pipeline developed by Dimmler (2025), a modular validation service was designed and implemented following a Design Science Research methodology. The service follows the Mixture-of-Rule-based Experts (MoRE-LLM) architecture described by Koebler et al. (2024), in which the LLM serves as the primary reasoning engine and six independent rule-based expert groups perform post-hoc checks on the output. A manual error analysis of 288 incorrect baseline answers identified the dominant error patterns and guided the rule design.The service was evaluated on a corpus of 90 companies using Claude Sonnet 4.5, with each experiment executed three times to account for LLM non-determinism. Mean accuracy increased from 41.1% to 58.8% (+17.6 percentage points), with 157 answers improving per run on average and zero regressions across all 2,670 question–answer pairs. All corrections originated from a single rule group, the units expert, through two deterministic rules: scale expansion and percentage format correction. An ablation study confirmed that disabling this group eliminated all corrections, while disabling any other group had no effect.

These findings demonstrate that a simple, rule-based validation layer targeting predictable format errors can produce a substantial and reproducible accuracy improvement in LLM-based financial information extraction.