Every AI law.
One audited API.

Two corpora, both maintained by us and both built from primary sources. The AI-law corpus holds statutes, regulations and bills collected from official legislature and agency sites. The enforcement corpus holds actions brought by regulators — federal agencies, state attorneys general, and named foreign authorities — each recorded with the agency, the instrument, the AI application, and a link to the announcement it came from. Nothing is entered from a secondary summary without being marked as such.

A profile carries a sector, the states it operates in, and the AI applications it runs. Instruments are resolved against it by rule — jurisdiction, layer, and whether the instrument is in force on a proven date. Enforcement matters are matched on four axes: AI application, the statute cited, the home jurisdiction, and sector. The first three are treated as strong signals and sector as a weak one, because a sector match alone describes an industry rather than a company.

When a brief is produced, the official source of each applicable instrument is read at that moment and is not stored. Provisions that impose an obligation and name the subject matter of the profile are quoted word for word, with the citation and a link to the document the words came from. Whitespace is normalised; nothing else is changed, and nothing is paraphrased.

A direction of travel for enforcement ("rising", "falling") is stated only when four conditions hold at once: enough dated matters, enough of them dated, enough distinct months, and comparable coverage across the two windows. That last condition exists because a collection lane going quiet looks exactly like enforcement declining. A monetary range is given only over matters that actually record an amount, and the count is printed beside it. A penalty recorded in a source that covers several matters is not attributed to any one of them. A provision whose citation is not unique in its document is dropped rather than guessed. Where a source could not be read, the brief says so instead of reporting an empty result.

No language model. Not to classify a matter, not to summarise a statute, not to phrase a finding. Every sentence in a brief is assembled from values the engine computed, by templates held in the codebase, and a test fails the build if a model call appears anywhere in the intelligence engine. This also means the brief contains no assessment of whether an obligation has been met, no prediction of enforcement, and no scoring of any individual person.