A typical chatbot
Gives a fluent answer from memory. It may be right or wrong, and you cannot tell which.
- No source to open
- No date on the information
- Guesses when it does not know
Groundtruth Data
AI makes things up.
We check it against official records.
Chatbots sometimes state facts that are simply false. Groundtruth looks each fact up in official sources such as SEC filings, FDA drug labels and clinical-trial registries, then gives you a receipt you can verify yourself. If it cannot prove an answer, it says so instead of guessing.
SEC EDGARFDA labelsClinicalTrials.govNo signup to try
The problem
A confident answer is not a correct answer.
When a chatbot gets a number wrong, nothing tells you. There is no source to open and no way to check.
Groundtruth
Copies the answer from the official record and attaches proof of where it came from.
- Link to the original source
- Date the record was captured
- Says “I could not verify that” when it cannot prove it
What it is for
Three ways people use it.
Get a fact you can trust
Ask about a public company’s financials, a drug label, or a clinical trial. You get the answer, the source, and a receipt.
Try Core →Test your AI before customers do
Building with AI? We measure how often your model gets facts wrong, using answer keys checked against official sources.
See Evaluate →Fix the mistakes you found
Ready-made, source-checked datasets to test your model and teach it the corrections.
Browse the data →How it works
Collect. Answer. Prove.
- 1
Collect
We store facts from official records with their source link and capture date. Records are never overwritten. A change is added as a new entry.
- 2
Answer
The answer is copied from the stored record by code, not written from memory. No record means no answer.
- 3
Prove
Every record is sealed into a signed public ledger. Your receipt can be checked in your own browser, without trusting us.
Measured, not claimed
Results from our own test, published as they are.
We ask questions generated from recent records, plus made-up questions that should be refused. Read the method and limits.
—answered correctly
—answered (the rest it declined)
—made-up questions it answered
—verified facts stored
Packaged dataset
500-case Drug Indication Proof Evaluation
A historical MED-RT source-grounding evaluation: 500 unique questions, verified reference answers, source provenance, and recorded model grades.
- JSONL evaluation and provenance files
- Historical model responses and grades
- SHA-256 manifest and internal-use license
This is separate from the retired 338-row diagnostic export. It is not clinical advice, an independently blinded holdout, or evidence of model improvement.
Read the evidence
What 500 SEC filing questions revealed.
Our recorded evaluation returned 136 correct answers out of 500. Read the task, the failure patterns and the limits behind that result.
Read the study →Work with us
Have a specific use case?
We agree on sources, scope, price and turnaround before any work begins. Your team keeps control of its own model.
Talk to usQuestions
Quick answers.
Is this a chatbot?
No. It is an answer engine for official records. It does not chat, and it does not guess.
What can I ask?
Annual financials of public companies, facts on FDA drug labels, and clinical trials by their NCT number. Coverage is limited to the records we have collected.
What does the receipt prove?
That the record was captured from the named source at a specific time and has not been altered since. It does not prove the source itself is correct.
What happens when it does not know?
It tells you it could not verify the answer, rather than inventing one.
See it for yourself.
Ask one question and open the receipt.
Ask it a question