Defensible by default.

Environmental work ends up in front of regulators, boards, buyers and courts. So every answer in Statvis traces to the exact passage it came from, every site stays private, and your professionals make the calls.

It reads the whole record, not the first ten matches.

A typical site file is about 20 times bigger than anything an AI model can hold at once. Most chat tools search for the top 10 to 50 passages and answer from those. If the answer is spread across 200 pages, most of it is never seen, and the one report that disagrees may never be read.

Statvis is built to work across the complete document set. It combines exact-word and meaning-based search, separates documents about the site from documents that only mention it, and keeps searching until it has the evidence.

Top-k retrieval

Answers from the ten best-looking passages

ReadRelevant, never readRelevant and read

Statvis

Searches until every relevant passage is covered

ReadRelevant and read

Contradictions, surfaced

Site files disagree all the time: two reports, two dates, two depths. A good answer shows both sources and says which one is the primary record, instead of silently picking one.

Near-identical IDs, kept apart

To a meaning-based search, MW-3610 and MW-3601 look 99% the same. Statvis also searches for the exact words, and every answer links to the page, so you can confirm it names the same well, lot or address.

Your sites stay yours. Confidential by design.

A vault per site

Each site is its own secure space, open only to the users you authorize. Access is set per site, not per firm.

Never used for training

Your documents and data are never used to train AI models. That's the default, not a setting you have to find.

No carryover between sites

Nothing learned on one site carries over to another. Unlike consumer chat apps with "memory", one client's details can't leak into another client's answer.

AI finds the facts. You own the conclusions.

We think of Statvis as an extremely thorough librarian. It reads everything, finds the facts and cites them. The judgement about what they mean, and every number you sign, stays with your professionals.

Calculations run as repeatable code, not as model guesses: the same inputs give the same answer every time.
When a report updates, only the affected sentences change, each one cited and marked for review.
Hard-to-read cells in old scans are checked by independent reads and settled by a person, not by the model's own confidence.

We measure it on real site documents.

EnviroBench is our benchmark for environmental document tasks: real problems, answers verified by people, and automatic scoring. It's re-run whenever a model or method changes, and a change that introduces a new error doesn't ship.

The private set holds 5,603 reviewed production cases. The tasks, scoring and results are on envirobench.com.

Well placement

Wells: find each labelled sample location on a site plan and place its point.

1,567

Table mapping

Tables: label every cell as sample, chemical, unit, value or method.

2,266

OCR auto-accept

Cells: decide which can skip review without accepting a wrong value.

722

Search reranking

Queries: order retrieved passages so the relevant ones come first.

603

Table separators

Crops: place the row and column lines of ruled or borderless tables.

300

Table containers

Pages: box every table and straighten crooked scans.

Table cell OCR

Hard cells: read one cell from a crop and count wrong values that would skip review.

The gate

Any new error against the baseline fails the run.

Ask us what you'd ask any AI vendor.

From our RemTech 2026 workshop on AI in contaminated sites. Use these with any vendor, or your own IT team, before client data goes into a new tool. A vague answer is an answer.

Ask	Red flag	Statvis
Is our data used to train models?	An opt-out buried in the settings.	No.
How are answers cited?	A document name, or nothing at all.	Page and passage, boxed on the source page, so you can check the exact text.
What evaluations do you have?	"We tested it and it looks great."	EnviroBench, described above.
What happens when it's wrong?	"It doesn't make mistakes."	It shows its sources, surfaces conflicting records and routes uncertain reads to a person.
Which models do you use, and can that change?	Silence.	We'll name them.
Where does our data go, and for how long?	"It's in the cloud."	Covered in writing in our security questionnaire, available on request.

Request the security questionnaire Workshop deck

Take control of your contaminated sites.

Bring us one site's documents. We'll show you what's in them.