Why we built ARTHR.

Every team that has tried to build on regulatory data hits the same wall in the same week. You point a model at a federal register PDF. You get back a paragraph that is almost right and quietly wrong about the deadline. You add retries, system prompts, fine-tuned classifiers. You still get drift on edge cases that matter, and no two teams misclassify them the same way.

Three failure modes show up in every off-the-shelf approach we tested early on. The model invents a deadline because the source document buried it in a footnote. The model conflates two regulators with similar names. The model hallucinates a sector tag because the closest training example was tagged that way. None of these are fixable with a better prompt. They are fixable with a better pipeline.

ARTHR is the engine we built when we accepted that regulatory data needed deterministic structure on top of generative summarisation. Four named components. GAWAIN crawls the regulator web on a scheduled rotation, with regional redundancy and backoff. PERCIVAL extracts and classifies against a controlled vocabulary that is curated by humans and queryable by everyone else. LADI ties alerts to regulators, regulations, and other alerts, so the dataset is a graph and not a list. SWORD is the delivery layer that fronts MCP, REST, webhooks, and flat-file outputs, so the same record reaches whichever stack you build on.

The output of ARTHR is a typed alert record with fourteen high-signal fields, full provenance, and a stable schema. A product team can query it. An audit can reproduce it. A compliance practitioner can read the human summary, click the source, and trust the deadline. This is what we couldn’t get from a generic LLM, and what every team building on regulatory data has needed since the day they started.