Palantir's ontology assumes a fundamental property about its domain: data has a correct structure.
An airport is an airport. A shipment is a shipment. A financial transaction has a sender, a receiver, an amount, and a timestamp. These aren't matters of opinion. They're facts with defined schemas. Enterprise data is largely objective, and Palantir's two-layer ontology — semantic elements for the nouns, kinetic elements for the verbs — works because the underlying reality is stable.
Knowledge work doesn't have that luxury.
The Messy Middle
When you're analyzing a health NGO's strategic landscape, "community wellbeing" doesn't have an EPA-mandated definition the way "dissolved oxygen in waterways" does. When you're mapping a creative agency's competitive positioning, "brand authenticity" means something different to every stakeholder in the room.
Most knowledge work operates in what we call the messy middle — concepts that aren't purely objective facts and aren't purely subjective opinions. They sit on a spectrum, and their position on that spectrum determines how much confidence you should place in any analysis built on them.
An ontology that ignores this spectrum produces misleading intelligence. It treats "community resilience" with the same confidence as "quarterly revenue." One is a measurement with a standardized definition. The other is a framework that three experts would define three different ways.
Consensus Scoring
We added a layer that Palantir's architecture doesn't need: consensus elements.
Every concept in our ontology gets scored on a 0.0 to 1.0 subjectiveness spectrum. An EPA dissolved oxygen standard scores 0.95 — near-universal consensus on its definition. A community wellbeing framework scores 0.15 — significant variance in how different experts define and measure it.
This scoring changes everything about how the system generates analysis.
When the agents encounter a high-consensus concept, they report with confidence. When they encounter a low-consensus concept, they qualify. They surface the competing definitions. They flag where domain experts disagree.
This isn't hedging — it's precision. An analysis that presents "community resilience has declined 12%" without acknowledging that five experts would measure it five different ways is not rigorous. It's misleading.
Palantir's ontology doesn't need this because enterprise data doesn't have this problem. Ours does, because knowledge work is where the consensus problems live. And pretending they don't exist doesn't make the analysis better — it makes it wrong in ways that are hard to detect.