The failure mode
Language models produce citations that look right - a plausible statute, a quote that reads like contract language. Some do not exist; some do not say what the model claims. For a firm, the risk is a fabricated reference travelling into a memo or a filing before anyone catches it. PONS closes that risk mechanically: nothing ships ungrounded, and every claim arrives with its evidence attached, so the reviewer's check takes a glance instead of a re-read.
That is what enforced verification buys the firm: the machine proves a source exists and where it sits, and review gets faster instead of longer. Telling a model to only cite real sources is a request, and requests get ignored under pressure. PONS enforces grounding at three layers - and none of them relies on the generating model grading its own work.
Layer 1: research answers keep only the citations the model used
When PONS answers a research question, retrieval runs first and builds a numbered source list; the model must cite inline against that list. After generation, only the citations the model actually used survive to the answer. Every reference shown maps to a real retrieved source - the model cannot invent source number 14 if only 9 exist. The filter is trivially simple, and worth more than any paragraph of prompt instructions about honesty.
Layer 2: extraction and review require verbatim spans
Extraction and contract review must return the evidence itself, copied verbatim from the document. In tabular extraction - one document per row, one question per column - every cell carries the exact text that supports it; a cell without a supporting span is recorded as not found, and the reviewer sees that honestly. Contract review works the same way: a finding the model cannot pin to actual clause text does not become a finding.
A verbatim span turns verification from a re-read into a glance.
If the system claims "the liability cap is 12 months of fees" and hands you the exact sentence it read, checking is a search: the sentence exists or it does not. A paraphrase sends you back to reading the whole clause.
Layer 3: generated documents pass a weighted evaluation gate
Drafted documents get the heaviest treatment, because a draft can be wrong in ways a citation check will not catch. Every draft is scored on four dimensions:
| Dimension | What it checks |
|---|---|
| Legal accuracy | Citation correctness, source usage, jurisdiction fit |
| Context coverage | Every question, note, and client fact addressed |
| Language quality | Correct legal language for the target locale |
| Structure and format | Template sections present, formatting sound |
A draft must clear the overall weighted score and every individual dimension. Beautiful prose with shaky law fails; so does sound law that ignored half the matter context. The legal accuracy check is the strictest: cited law must be real, relevant, and from the right jurisdiction, and a document that lists law after law without connecting it to the facts fails on its own. When a draft fails, a targeted fix pass repairs the specific problems rather than regenerating the whole document and hoping the second roll lands better.
What this adds up to
Research answers whose every citation is real. Extractions where every cell carries its evidence - and a cell without evidence says so. Drafts that must clear a legal-accuracy gate before anyone sees them, and get repaired precisely when they fail. The check that used to be a re-read becomes a glance, and that is where the time comes back. One detail worth knowing: confidence values on extracted cells are the model's own high, medium, or low self-assessment - there to sort what gets checked first.