vanKonijnenburg is a large language model trained end-to-end for grounded answers. Evaluated against Google's official FACTS Grounding methodology on 643 independently fact-checked answers: 0 fabricated facts, only 4 with a reasoning error.
estimated score under Google's official FACTS Grounding methodology, on our internal 643-answer fact-checking benchmark (checked claim-by-claim against source documents)
answers with a residual reasoning error — the only issues found; not one fabricated fact across the whole benchmark
Under our own stricter internal measure — which also counts reasoning errors and output artifacts, not just literal source-matching — the score is 95.3%. We report both rather than only the flattering one.
Grounding is trained in, not bolted on afterward with a separate checking step.
Every training example follows a Chain-of-Evidence: each claim traces back to a verifiable source, not just plausible-sounding text.
Direct Preference Optimization specifically targets the reasoning errors that generic LLM judges miss — scope mismatches, category errors, misapplied facts.
Every answer in our benchmark was checked claim-by-claim against its source document, by hand — not a single automated pass/fail score.
Usage-based, like the LLM APIs you already integrate with.
No monthly minimum. Waitlist members get preferential early pricing and priority onboarding.
vanKonijnenburg — the grounded LLM. We're onboarding a small group of early users first. Leave your details and we'll reach out.