Veris is simulation infrastructure for building benchmarks. Enterprises use it to put the agents they buy and the agents they build through their own workflows, systems and rules, in a twin of production rather than a demo. The boards below are public benchmarks we built the same way in different domains.
An AI agent is handed one question and a company drive holding 1,435 scanned U.S. Treasury volumes — 118,524 pages of image-only PDFs, with no text layer anywhere. It has to find the right volume through the drive's own MCP server, locate the table through the contents pages, read the right line and column with a hosted OCR server, check the figure against the table's own totals, and reply with one number to the cent. A script grades it, not a model: the figure has to land inside a range built to exclude every known misreading. Finding the page is largely solved. Reading it is not — a third of all failures are a single misread digit.
An AI agent is the second-level anti-money-laundering investigator at a mid-sized US bank, and every row on this board is the same job done by a different harness and model. A colleague hands over an alert and stays on the chat. The agent pulls the ledgers and KYC records, re-screens the parties against a sanctions and PEP watchlist, applies the bank's written procedure, and files a disposition: a structured report, a narrative and the case action. It is graded on the state it leaves the case system in, never on what it says. In 48 of the 153 investigations the deciding record is not on the file — the colleague is holding it, and hands it over when asked. On the 39 of those a guess cannot get right, only 5 of 1,364 attempts that did not ask were solved.
Rory is an AI voice agent on the billing line of a regulated utility, and every row on this board is the same Rory built on a different voice stack. It verifies the caller, explains why a bill went up, takes a payment, opens or changes an instalment plan, asks for a due-date extension and switches an account to paperless, against a billing system and a loan engine that enforce their own rules. Every call chains two to four of those requests, to a calm or an angry caller, over street noise, café babble or a television playing in the room, and the grade comes from what the agent left behind rather than how the conversation read.
Riley is an AI voice agent on the support line of a credit card issuer, and every row on this board is the same Riley built on a different voice stack. It verifies the caller, then freezes or activates a card, issues a replacement, updates where that replacement is sent, and works through a lost card or a reported fraud. Each build runs the same 100 unscripted simulated phone calls, with connect rate, latency, barge-ins and cost on one board.
