Case Study - Ground truth for AI agents acting in the physical world

Overview

A fluent model will confidently invent the weather at a place it has never queried. When that model is just chatting, the made-up fact is harmless. When it is an agent acting in the world, the same invented fact becomes a wrong action with a real cost. Pointmoon was built for that second case: it is the agent-callable trust layer for physical and environmental reality.

The idea is simple. Before a model speaks or acts about a place, it asks Pointmoon what is true there right now. Pointmoon reads live data from real sources, the weather, the turning season, rivers and tides, the shape of the land, the sun and the moon, and hands back a short read of the present moment. Every fact carries its source, a timestamp, and a confidence value, so a stale reading is never passed off as current. Where the data will not hold, Pointmoon says so and returns a typed silence with a reason.

The refusal to guess is the point. We kept language models out of the fact layer, so it holds facts and nothing it can invent. Pointmoon brings the grounded fact or the honest gap, and the agent decides what to do with it.

It ships as real infrastructure built for agent builders: a clean API, an MCP server so an agent can pull verified field-truth the same way it reaches for any other tool, a versioned contract so consumers can upgrade safely, and an open-source quickstart anyone can run. It is built for anything acting on the physical world, from location-aware assistants to outdoor, logistics, and agriculture systems. The first product standing on it is Rewyld, where every place-aware practice is grounded in Pointmoon before a single word is spoken.

How we solved it

The hard part is that a fact about the physical world has a shelf life. A reading that was true an hour ago may be wrong now, and different signals go stale at different speeds: the weather turns in minutes, the season over weeks. So we tied every fact to the moment it was read and to how long that kind of fact stays good, and when a reading ages past its window Pointmoon lowers its own readiness rather than serving it as current. When a source is down or the data is thin, the contract returns a typed silence with a reason instead of a best guess. Keeping any language model out of the fact layer is what makes that possible: there is nothing in the path that can smooth over a gap with an invention.

What we built