What is the observed / inferred / unknown evidence model?
It is how AI Control labels every statement about machine communication. Observed means directly captured or verified, such as raw payloads, identities and system events. Inferred means an analytical hypothesis about meaning, always with a confidence and supporting evidence. Unknown means what cannot currently be established, such as encrypted content or an unmapped token. The three are stored separately and rendered differently so an inference never looks like a fact.
- Observed: verified.
- Inferred: analytical, with confidence.
- Unknown: unresolved, kept visible.
- Inferences are never stored as raw evidence.
How it works
- Raw evidence — Communication is captured and preserved unaltered with an integrity hash.
- Observed — Verifiable facts are extracted: identities, protocol metadata, timestamps, system events.
- Analysis — Structure, patterns and correlations are computed from the evidence.
- Inference — Hypotheses about meaning are proposed, each with a confidence and its evidence.
- Unknown remains visible — Whatever cannot be established is listed explicitly — never silently dropped.
Frequently asked questions
Why not just show the most likely meaning?
Because operators act on what they see. Showing a guess as fact leads to wrong decisions; showing confidence and unknowns leads to proportionate ones.
Where is inferred data stored?
In separate interpretation and correlation records that reference the evidence, never inside the raw evidence store.