Grounding Measurement — Early Partner Program

Of the values your AI agents produce for your books and reports, how many can be traced back to a source document? We measure that number — the grounding rate — so your pilot decisions rest on evidence, not impressions.

What you get

• Grounding-rate measurement of your AI agent outputs, with a written report
• A breakdown of caught ungrounded values, reported monthly
• Material for an evidence-based go/no-go on independent output verification

Why an independent layer

Controls that live inside the system they are supposed to check cannot serve as evidence about it. Okugaki runs as an independent verification layer in front of the agent's output path — the same separation your auditors already expect from any control worth relying on.

Scope, stated plainly

This program provides measurement. To be precise, here is what it does not include:

• No correctness verification. Okugaki verifies provenance — whether a value traces back to an observed source — not whether the value is right. If the source is wrong, a grounded value is faithfully wrong. That boundary is what keeps the verification mechanical and reproducible.
• Not an assurance engagement. The report is a measurement report, not an audit opinion or an evaluation of your internal controls.
• No operational guarantees or support commitments. The proven environment is an OpenAI-compatible API with MCP tool integrations; for other configurations, ask us first. A QuickBooks Online pack exists at demo quality; we do not make support commitments for it at this stage.
• Data handling. The ledger of observations and verdicts is written to a destination you control; Okugaki itself stores nothing. What is shared with us for the report is agreed before the engagement starts.

Measurement runs for up to one month (start date and extensions by agreement). The precise definition of the guarantee is published as the canonical claims scope document in our public repository.

Get in touch

We can share a 15-minute demo recording, or run a 30-minute live detection demo against your workflow.