AI that assists, and knows its limits
The model decides how to help, within boundaries the platform enforces. It never decides what it is allowed to do.
Bounded by design
The research agent and the AI Assistant both work as tool loops. The model chooses among tools with strict schemas; VERIDEX validates every argument, enforces source policy and budgets, and asks you before any external disclosure.
- You
You ask a question about the case, or start a research run.
- Model decides
The model picks a tool and fills in its arguments.
- VERIDEX enforces
VERIDEX checks the call against the tool’s strict schema and confirms the tool is allowed here.
- VERIDEX enforces
It enforces source policy, budgets and pacing, and stops to ask you before anything is disclosed to an external destination.
- VERIDEX enforces
The tool runs over the case’s own data and returns a bounded result.
- Model decides
The model reads the result, then calls another tool or drafts an answer.
- VERIDEX enforces
The answer is checked against the context the tools built. Citations are attached, and caveats are added when it leans on unreviewed or rejected items.
- You
You read a cited answer and make the judgment.
Cited and graded
Findings, synthesis and person-analysis claims cite the exact evidence behind them and carry an explicit support grade. A claim without valid citations does not pass.
Support grades describe citations. They are not a probability, and they are not a measure of confidence that you have the right person.
- Supported
- The cited evidence backs the claim.
- Partially supported
- The cited evidence backs part of the claim, not all of it.
- Unsupported
- The cited evidence does not back the claim. It stays visible, marked as such.
- Rejected
- The claim did not pass validation.
An assistant that shows its sources
Ask about the case in plain language. The assistant gathers context only through its tools: the structured profile, findings, synthesis, person analysis, visual evidence, source coverage, conflicts, research progress and review status.
It can see unreviewed leads, but it labels them. When an answer cites something unreviewed or rejected, a caveat is added. Searching the web needs your approval of a specific disclosure first.
A research agent you can stop
A research run lets a model plan what to search, open and follow for one case. The platform separates permission to discover a source from permission to fetch it, applies budgets and pacing, checks that quoted text really appears in the source, and tracks novelty and coverage.
You can pause, resume or stop a run, and approve or decline each disclosure to a new destination. What it finds becomes candidates for review, never accepted evidence.
Needs a configured providera deterministic demo agent is built in
Safe by default
Out of the box, every AI capability runs on a built-in deterministic demo provider, so the system works end to end without contacting any AI service. Connecting a real model is a deliberate configuration step for your deployment.
-
Analysis, assistant and AI report
Needs a configured providerdemo provider until configured -
Visual analysis of images
Off by default -
Assistant web search
Off by defaultand approved per message
Read about security and privacy
The safeguards that are built in today, and the ones that are still recommended next steps.