Get API keyQ3 2026 · METHODOLOGY RELEASE
The Company Data Grounding Index.
A quarterly benchmark of whether leading AI systems identify the right company, state current facts and support them with official evidence.
40 / 100 POINTS
Is the material company fact correct?
Compare the stated value with the evidence at the benchmark cut-off.
THE QUESTION
A confident answer is not necessarily a grounded answer.
The index tests the parts of company intelligence that language models routinely blur: similarly named entities, current versus historical facts, ownership paths and financial periods.
The model may decide how to investigate. It does not decide what counts as proof.
- Correct legal entity
- Correct material fact
- Correct at the stated date
- Supported by official evidence
WHAT IS AI GROUNDING?
Connect a generated company answer to evidence.
AI grounding connects a model’s response to external information used to support the answer. For company data, that means resolving the legal entity, checking each material fact against a relevant record and preserving the time and source context.
Retrieval provides context to a model. Verification evaluates a particular claim against evidence. A useful company-data benchmark tests both the answer and whether its supporting record establishes that claim for the correct entity and date.
This index focuses on company facts. It is not a general ranking of model intelligence, and its proposed claim families are broader than the six identity fields supported by the current CompanyProof verification endpoint.
ILLUSTRATIVE EVALUATION
How a plausible answer can fail a grounding check.
Suppose an answer says that “Example Holdings is active and generated £12 million last year.” The entity may be a namesake; the status may come from an old snapshot; the revenue may describe the whole group and an earlier accounting period. Each claim needs its own adjudication.
| Dimension | Question for the reviewer |
|---|---|
| Entity resolution | Does the registration number and jurisdiction match the intended entity? |
| Factual accuracy | Does the supplied value match the evidence for that specific attribute? |
| Evidence fidelity | Does the cited record actually support the claim? |
| Temporal accuracy | Is the statement correct at the requested date and reporting period? |
This is an example of the review process, not a scored model run. No sample size, model score or comparative performance result is implied.
For implementation, start with the company verification checklist, financial statement checks and supported verification fields.
REPRODUCIBLE SOURCE EXAMPLE
Inspect a real record before scoring an answer.
The public Monzo example identifies one legal entity, preserves three registry observations and links the source with a check date. Download the snapshot and inspect how the registration number, claim, observed value and source remain connected.
Review the dated evidence example → · Download the source snapshot
This is a manually reviewed reference example. It is not a model run or comparative score. The full benchmark continues to require a frozen test set, recorded model versions, repeated runs and adjudication.
THE 100-POINT SCORE
Four dimensions. No credit for confident invention.
A model can answer correctly and still lose points when it selects the wrong entity, cites weak evidence or presents an old fact as current.
Factual accuracy
Is each material company fact correct at the benchmark cut-off?
Evidence fidelity
Does the answer rely on the right official record without inventing support?
Temporal accuracy
Does it distinguish the current state from an older but once-correct fact?
Entity resolution
Did the model select the correct legal entity, identifier and jurisdiction?
HARD FAILURESWrong legal entityFabricated sourceUnsupported ownership pathStale fact presented as current
CLAIM DIFFICULTY MAP
Company facts fail in different ways.
The founding edition starts with the UK and European Union and separates simple identity claims from time-sensitive ownership and financial claims.
Registered name, number, legal form
Official entity record
Brand or namesake confusionActive state, dissolution, incorporation
Registry status and event history
Stale current-state claimCurrent appointment and cessation
Officer record or filed notice
Former director presented as currentShareholder, UBO and ownership path
Filed ownership evidence
Unsupported percentage or broken chainPeriod, currency and filed value
Filed accounts or official statement
Mixed periods or silent currency changeFROM PROMPT TO PUBLISHED SCORE
Every result must survive the same evidence path.
Raw outputs are retained. Scoring happens only after the intended company and official record are independently resolved.
Register the test set
Freeze the company, claim, jurisdiction and evidence cut-off before model runs begin.
Lock model versions
Record the exact model, access date, settings and system instructions used.
Run independently
Repeat prompts without carrying context between tests and retain every raw answer.
Resolve the entity
Match each answer to the intended legal entity before any fact receives credit.
Verify the evidence
Compare claims with official registry records and preserve the adjudication trail.
Publish the denominator
Release aggregate scores only with exclusions, failed items and sample size visible.
PUBLICATION GATE
No leaderboard before the evidence audit.
This release establishes the scoring contract. Named model rankings will appear only after the complete founding test set has been run, adjudicated and checked for reproducibility.
- Exact model versions and access dates
- Prompt and scoring rules disclosed
- Repeat runs retained
- Exclusions and denominator visible
- Material corrections dated
- Methodology
- Published
- Test set
- Evidence review
- Model runs
- Not published
- Leaderboard
- Withheld
Withholding an incomplete ranking is part of the method—not a missing result.
HELP SHAPE THE FOUNDING EDITION
Contribute one company claim that models get wrong.
FREQUENTLY ASKED QUESTIONS
Company Data Grounding Index FAQs
What is the Company Data Grounding Index?
It is a CompanyProof Research benchmark methodology for evaluating AI-generated company facts. It considers whether a model selects the right entity, states accurate facts and supports them with appropriate evidence.
How is the Company Data Grounding Index scored?
The published methodology allocates 40 points to factual accuracy, 25 to evidence fidelity, 20 to temporal accuracy and 15 to entity resolution, for a total of 100 points.
What is factual accuracy in a company-data benchmark?
It asks whether each material company fact is correct at the benchmark’s evidence cut-off. The intended legal entity and the relevant period must be established before the answer can be assessed.
Why does the benchmark assess evidence fidelity?
An answer can state a correct value while citing irrelevant or invented support. Evidence fidelity tests whether the referenced record actually supports the company claim being made.
What does temporal accuracy measure?
It measures whether the answer distinguishes current facts from historical ones. A once-correct company status, director appointment or financial figure can be misleading if presented without its relevant date.
Why is entity resolution scored separately?
Facts about the wrong legal entity do not answer the question correctly. The benchmark checks company identifiers and jurisdiction to detect confusion between brands, subsidiaries and similarly named businesses.
Which company claims are included in the methodology?
The claim families include legal identity, status, directors, ownership and financials. Each family has different evidence requirements, such as appointment dates, ownership paths or reporting periods.
Does the current index publish an AI model leaderboard?
The published release establishes the methodology. Named model rankings are withheld until the founding test set has been run, adjudicated and checked for reproducibility; no ranking should be inferred from the methodology alone.
How will benchmark results be made reproducible?
The protocol records the test set, model versions, prompts, settings, raw outputs and evidence cut-off. Published aggregates must include sample size, exclusions and failed items alongside the scoring process.
Can I contribute difficult company-data questions?
Yes. Submit a company claim that models get wrong with the intended entity and evidence needed to assess it. Useful examples include namesake confusion, outdated officer records and mismatched financial periods.
Contribute a benchmark question