AI market analysis read against ninety days of history.

Four concrete things: relevance scored, each source classified for what it is, independent corroboration counted, and the finding read against ninety days of the same company’s history. Each is implemented by a named agent.

Because a tool that silently drops things cannot be audited. The log lists what was held back and the score that held it, so the judgement is reviewable rather than taken on trust.

Both, and they use the same chain. A market is researched from agency statistics and the companies in it; a company is researched from its own evidence and independent reporting about it.

A line in a log with the reason and the score that held it - a price page that changed its footer, a reposted requisition, a reprint of a story already reported. The point of showing them is that a reader can disagree with the judgement, which is impossible when a tool drops material silently.

Ninety days is long enough that a quarterly pattern is visible and short enough that the comparison is still about the current state of the company. A finding is read backwards against that window, so a change is reported as a change rather than as news.

A summary shortens what was said. Analysis decides what each source is worth, counts who independently agrees, sets the finding against what was already known, and holds back what changes nothing. A summary of ten articles about one press release is ten paragraphs; the analysis is one finding with a source count of one.