Most AI assistants know what their model happened to learn, and nobody can say exactly what that is. For a system that interprets markets, this is a real problem: when its view changes, you want to know whether the market changed or its knowledge did. NOVAQ's market context expert was built so that this question has an exact answer.
In August 2026 the knowledge base met the running system. The expert reads live data, session statistics and the economic calendar, together with a digest of market states from the vault, and it classifies the current state of the market.
From notes to runtime
Twelve universal states and seven session states were distilled from execution-state notes, a note type designed in May. Those notes rest in turn on attributed atoms, so a statement in the digest can be traced back through its atoms to authors and sources. The digest can be regenerated only from vault notes, and its hash enters the expert's input record. Any change in what the expert knows is deliberate and auditable.
This is the point where the knowledge pipeline stops being a library and starts working. The 145,000-note knowledge base becomes a small, curated, versioned input to a running AI system.
Discipline by design
The expert runs every 15 minutes during open markets, up to 480 runs a day. It can decline to give a view, and its first run did so because the data was not fresh. An expert that says it does not know is more useful than one that always answers.
Keeping authors apart
Early on, a concept from one author appeared in an analysis of another, and it sounded entirely credible. An instruction did not stop it, because a model cannot reliably ignore what it already sees. Provenance filtering did: a model cannot use a note that never reaches it. Control over what an AI system retrieves turned out to be stronger than instructions about how to behave.
Why it matters
The design gives three properties that are rare in applied AI: inputs that can be inspected, knowledge that is versioned and answers that can be reproduced from their record. It complements the verification of the analyst's data and fits the same discipline as the rest of the platform.
Limits
The expert runs in shadow and does not affect trades. Other experts in the same design remain disabled.