Grounding is constraining a language model's output to verifiable source data — retrieved documents, tool results, or cited records — rather than its internal memory, so claims can be traced back and checked.
A grounded answer is one where each factual claim is backed by a retrievable source the system actually consulted, ideally with a citation. It is the direct antidote to hallucination: if the model can only speak from supplied, sourced data, it has far less room to invent.
Strong grounding needs more than retrieval — it needs provenance: knowing where each datapoint came from and being able to link back to the origin.
A grounded answer about insider buying cites the specific Form 4 (by accession number) it read, so the user can open the original filing.
Per-datapoint provenance is Arkolith's differentiation — every value links to its origin filing, giving agents the grounding that makes their answers trustworthy.
Arkolith turns this into live, sourced data your agent can query — SEC filings, insider activity, and market data behind one key, every datapoint traceable to its origin.