A well-written market brief can be wrong in a way that is hard to see. The prose reads well, the reasoning sounds sound, and one input is quietly incorrect. For an AI system that reports on markets, the usual quality checks, which focus on the text, miss the real risk. So NOVAQ checked the inputs instead of the prose.
The system
The intraday analysis agent builds a multi-layer picture of the market and writes a structured brief, with about 93 scheduled slots each trading day. Each brief draws on a large set of computed fields covering price, levels, regime, session state and calendar.
The check
All 253 fields passed to the brief were recomputed independently from raw data. The first pass confirmed 202. It found 17 mismatches, 28 semantic issues and 6 values that could not be verified. Three rounds of fixes followed, and a live audit then checked 60 bundles during real sessions.
The semantic lesson
The 28 semantic issues were the most instructive result. A number could be correct and still mislead. A level below price cannot be a target above it, yet a model handed a bare number has to guess what it means, and a capable model guesses plausibly. The lasting fix went into the data layer: each field now carries its direction and role. Meaning belongs in the data.
The result
A reader can follow each statement back to a field and each field back to its source. The agent has no order rights, and its reliability rests on verified inputs.
Why it matters
Most AI quality work focuses on prompts. On this project the larger gains came from the data the model receives. The same principle runs through the knowledge-based expert, whose inputs are versioned and hashed.
Limits
Field verification covers correctness of inputs, not the quality of the conclusions drawn from them. The brief has no order rights.