GRUNDNORM · SOURCE-OF-LAW TRUTH LAYER
Where an AI tells you what the law says, Grundnorm proves what it meant.
Give it an open identifier and a date. It returns the sealed record of what that norm meant on that date, broken into deontic atoms, each one carrying the literal fragment of text it came from, signed with Ed25519.
It does not answer legal questions. It resolves meaning and proves the resolution.
NOT A PRODUCT YOU SIGN UP FOR
A platform built around your corpus, not a shelf you buy from.
There is no self-service tier and no generic deployment. You bring your specifications, your corpus, your sources and your signers, and the deployment is built around them. How the corpus arrives is yours to decide: an official gazette, a bulk delivery, your own document manager, your own API. What the platform solves is everything that happens after ingestion.
Talk to us about your corpusNOT ONLY LAW
The layer does not know it is looking at law.
It needs three things and no more: texts that state obligations, identifiers that name them, and dates on which versions take effect. Every corpus with those three properties carries the same buried problem. A decision taken years ago is judged today against the text that was in force then, and the text has changed since. Law is where we started because that is where the problem is oldest, not because that is where it stops.
Financial supervision
Which reporting obligation was in force on the trade date?
A review opens years after the fact and reads the rulebook as it stands now. The obligation being judged is the one that stood then.
Tax
Under which wording was this position taken?
An assessment can open long after the return was filed, by which time the article has been amended twice and the consolidated text shows only the last of them.
Pharmaceuticals and medical devices
Which version of the guidance governed this submission?
A dossier is defended against the requirements in force on the day it was filed, not against the ones the agency publishes while the defence is being written.
Standards and certification
The certificate says it conforms to the 2019 edition. Which text is that?
Editions supersede one another on their own schedule while the certificate keeps circulating, attached to a product, in a supply chain, for a decade.
Energy and utilities
Which grid code applied when this plant connected?
Codes, tariff regimes and connection conditions each move on a different calendar, and an asset outlives several rounds of all three.
Internal policy
Which version of the policy did the analyst have in front of them?
This one is not law at all. It is the first question every internal investigation asks, and the honest answer is usually that nobody kept the version.
The rest of this page speaks in legal vocabulary because law is the corpus the demonstrator seals. The machinery underneath is indifferent to what the text is about.
ON ACCURACY FIGURES
We do not publish one, and that is a position rather than a gap.
A precision figure measured over five articles of one regulation says nothing about a commercial-law corpus in another country, in another language, under other institutions. Aggregating unlike corpora produces a number with no referent. Accuracy here is an acceptance criterion per engagement, measured against your corpus and by your jurists. Distrust any vendor who shows you a single precision figure. Including us, if we ever showed you one.
VERIFY IT YOURSELF
Do not take our word for it. The demonstrator ships in the package.
The endpoint and its key are embedded in the published SDK, so this runs as written. The version is pinned on purpose: a page that says check it yourself should not break when we ship.
pip install grundnorm==0.5.0
from grundnorm import GrundnormClient
c = GrundnormClient()
n = c.resolve(id="http://data.europa.eu/eli/reg/2016/679/art_99",
jurisdiction="EU", at="2026-01-01", domain="legislation")
print(n.status, n.verification["verdict"], n.verification["absence"])Read that slowly, because it is the whole argument in three words. The system does not have that article. It says so. And it proves it: the negative answer carries a cryptographic non-inclusion proof against a signed commitment over the set. Then it refuses to call itself authoritative, because the signers are demo identities we control: the response reports key_provenance as pinned_demo and demo_keys as true. A verification layer that cannot tell you when not to trust it is worth nothing.
The proof is bounded and says so. It does not prove that the article does not exist in EU law; it proves that within a declared, signed scope it is not there. The same response states that scope: a demonstrator corpus of five sealed records.
THIRTEEN VERDICTS, NOT A CHECKMARK
It cannot return a bare ok.
Every answer names which of a closed set of verdicts it is in, and the set is exported by the package so you can read it without asking us. Thirteen for a record that exists, ten for an attested absence, ordered from failure to success. Eleven of the thirteen are distinct ways of saying no: content tampered, view inconsistent, outside the sealed window, signature invalid, provenance unsupported, answers a different question. Only the last two are outcomes anyone would call a pass, and one of those still declines to claim authority.
from grundnorm import VERDICTS, ABSENCE_VERDICTSTHREE STATES, NOT TWO
When two competent readings genuinely diverge, it does not pick a winner.
Each deontic atom carries a state: fixed, needs_review, or for_the_court. The last one is the point. Where independent readings disagree in substance, the machine records the divergence and leaves it to a court, rather than choosing one reading to look confident. The state is derived from independent reads agreeing, not from a model's confidence in itself.
This one you cannot reproduce in the demonstrator, and we would rather say so than fake it. Five articles chosen to show the pipeline do not produce real divergence, and sealing a divergent record on purpose would mean fabricating a disagreement the engine never found, which is precisely what this product exists to prevent.
WHAT THE DEMONSTRATOR IS
Five sealed records, demo signers, and a payload that admits it.
The public corpus is a sample of one regulation published as a demonstrator, not an institutional deployment. Its seals are made by demo identities we control, and the verifier declares exactly that rather than leaving you to find out: key_provenance reads pinned_demo and demo_keys reads true. The response carries custody as unverified, the ledger as unanchored, and the classification as not present. None of that is a defect being disclosed. It is the product working: an attestation layer whose first job is to tell you how far it can be trusted.
Bring us your corpus.
Grundnorm is an independent venture of Quantum Nexus Ventures FZCO, a sibling of Nexus Legal rather than a feature of it. Deployments are scoped per engagement, including the acceptance criteria your jurists will measure them against.