FME: Finance Media Intelligence Engine

I wrote a 68-page specification for how a finance documentary channel should think, and FME is that spec running as software. A single database holds channels, snapshots, signals, candidates, opportunities, projects, claims, sources, lessons, costs and logs. The command line walks a topic from ingest and signals through gaps and outliers, demand probing, scoring and briefing to a project checkpoint and export, and a final lesson step feeds what was learned back in. Intelligence is never reset between videos.

Judgement stages (topic debate, audience twin, script review) are executed by a session model against role cards, and a verdict that does not cite engine rows is invalid. Demand is measured with search result pools rather than keyword APIs, which were verified to be plan-limited. The outlier detector surfaced channel videos performing 47 to 132 times their baseline, which became the queue.

A second specification added the Visual Intelligence Engine: a shot and attempt ledger, a visual bible, metaphor and purple-cow registries, hero-frame-first generation from a still-image engine into image-to-video, and a generation law enforced in code (one generation plus at most one diagnosed retry per shot; a third attempt raises an error). Two episodes are registered with 129 and 17 cited claims respectively, and a card renderer gained animated money-flow diagrams.

How it works

1. Ingest watchlist channels and snapshot their metrics
2. Detect outliers and probe demand with search pools
3. Score candidates and write a brief with cited claims
4. Create a project with its claim and source ledgers
5. Plan shots in the visual engine under the generation law
6. Record lessons and costs back into the database