PROVENANCE RESEARCH
Provenance, not promises.
An AI answer to a regulated question is only as trustworthy as the source behind it. This page sets out an open, reproducible way to measure whether an answer traces to a primary source, and shows the first-party evidence behind every entry in the Bidda registry.
We do not ask you to trust a score. We show you a method you can run yourself, and numbers you can check against the primary source.
SEE THE EVIDENCE
TRY THE LIVE COMPARISON
THE PROBLEM
General-purpose AI models produce fluent regulatory answers that read as authoritative whether or not they are grounded in a real source. A model cannot tell you, on its own, which of its sentences trace to a live instrument and which are plausible reconstruction.
Independent academic research has documented that AI systems used for legal and regulatory work can return unsupported or misgrounded answers on a meaningful share of questions, even when they appear to cite sources. In regulated work, an answer you cannot trace is an answer you cannot defend.
This is a property of ungrounded language models as a category, not a claim about any particular product. Further reading: Stanford HAI / RegLab research on the reliability of AI legal research tools (2024), and the broader literature on hallucination in large language models.
WHAT WE CLAIM, PRECISELY
The honest claim is narrower than "our answers are right", and it is checkable.
What Bidda does claim
Every answer traces to a primary source you can open, carries a content hash you can recompute, and carries a date. That is provenance, and provenance is measurable and independently verifiable.
What Bidda does not claim
We do not claim an answer is legally correct, nor that following it makes you compliant. Bidda is reference intelligence that supports your judgement; you and your advisers draw the legal conclusion.
THE EVIDENCE
Provenance coverage of the registry
These are properties of Bidda's own corpus, measured deterministically across all 10,090 nodes on 18 July 2026. They describe source-traceability, not legal correctness, and each is something you can confirm for any node yourself.
100%
resolve to a primary-source URL
Every node links to the official instrument you can open.
100%
carry a content hash
A SHA-256 fingerprint you can recompute to detect any change.
98.8%
carry five or more citations
Averaging 6.79 primary citations per node across the registry.
Weekly
source re-check cadence
Every source is re-fingerprinted on a weekly cadence to catch drift.
VERIFY A NODE YOURSELF
HOW EACH NODE IS BUILT
Coverage figures are a dated snapshot measured over the live registry. They are a measurement of traceability and integrity, not a determination that any node is correct or that any obligation is met.
THE OPEN METHOD
Grounded versus ungrounded, measured in the open
The point of a benchmark is that anyone can reproduce it. Here is the exact method we use to measure whether an answer traces to a primary source. It compares an ungrounded model answer with a source-grounded one, and scores the share of claims that trace to an openable source.
1
Fix a question set
Take a representative set of real regulatory questions across pillars and jurisdictions, and publish it in full so the test is reproducible.
2
Answer without sources
Ask a general-purpose model each question with no supporting material. Record the answer and any sources it volunteers.
3
Answer with the source supplied
Ask the same question again with the relevant Bidda node(s) provided as context. Record that answer and its citations.
4
Grounding check
For every factual claim in each answer, check whether it traces to an openable primary source. A claim with no source it can point to is ungrounded.
5
Misgrounding check
Where an answer does cite a source, check whether that source actually supports the claim. A citation that does not support the claim is misgrounded.
6
Report the share that traces to source
Publish, for each condition, the share of claims that trace to a primary source. This measures provenance, not legal correctness.
RESULTS
STATUS: IN PROGRESS
We are running this method across a representative question set and will publish the full question set alongside the results here, so the figures can be reproduced end to end. The number only means something because the method is open; a score with no reproducible method behind it is marketing, not evidence.
WHY THIS MATTERS
For compliance teams
You can put an answer in front of an auditor with the source, the date, and an integrity check attached, instead of a paraphrase you have to defend.
For AI builders
A grounding layer with published provenance gives your agents answers they can cite, and gives you a record of what they relied on.
For everyone else
Provenance is the one property of an AI answer you can check without trusting the model. This page shows how to check it.
SEE THE LIVE COMPARISON
BROWSE THE REGISTRY
This page measures provenance and integrity: whether an answer traces to a primary source, and whether that source is unaltered and current. It is not legal advice and not a determination that any node is correct or that any obligation is met. Every rule should be reviewed by a qualified professional before it is relied upon.