When a number ships into a board deck, someone owns it. Nodal gives your data team the evaluation layer to own it with evidence: every question logged and attributed, every context change regression-tested, every agent held to one governed truth.
Nodal reads your context repo. It never gets write access, and it never touches your database.
Queries run on your warehouse, under your credentials. Row-level data never leaves your infrastructure.
For healthtech, fintech, and insurtech: run the evaluation layer inside your own perimeter.
Context is authored and approved by your data team, reviewed by PR. Nothing is silently auto-generated.
Our test confirms it: a thin Claude Code + Sonnet harness — no custom agent, no fine-tuning — ranked #7 on Spider 2.0-Lite, a widely used text-to-SQL benchmark. General-purpose agents with the right context compete with purpose-built systems. Review the setup ↗
Anthropic's data team found that model-generated documentation encoded the very ambiguities they were trying to eliminate — net-negative on their evals — and that retrieval over thousands of prior queries moved accuracy less than one point. The context has to be owned by a human.
Anthropic found this too — even a working system rots as definitions, tables, and the business change underneath it. Keeping answers right is an engineering problem, and regression evals are how you treat it as one from day one.
“We watched our offline accuracy drift from ~95% at launch to ~65% over a month before we treated this as an engineering problem.”
An executive asks from their phone; the agent interrogates the question before answering; your data team sees exactly who asked what in Slack — captured, attributed, and feeding the eval suite.
Need implementation help? We work with a partner — ask us.