Narration and the grounding verifier

Two narrators

The deterministic composer builds prose directly from the engine's output. It runs for replayed reads, for signed-out visitors, and whenever the model is unavailable. The second narrator rewrites that material in the house voice.

Each read records which narrator wrote it, so the provenance of the prose is part of the record rather than something to infer from style.

The verifier

The engine produces a set of numbers it actually computed. Any generated sentence citing a number outside that set is rejected before display, and replaced by the deterministic bullet for the same point.

Rejections are disclosed on the read rather than hidden — the uncertainty list states how many generated claims were dropped for citing evidence the chart does not contain. A silent substitution would leave a reader believing the prose was always clean.

What the model is never allowed to do

It never sees the signal and never writes one. Directive language — telling a reader to buy or sell — is banned from generated prose outright.

The division is between measurement and description. The engine measures and decides; the model describes what was decided. A narrator that could nudge the verdict would make the verdict partly a matter of phrasing.

Refusals, not degradations

When the allowance for a read is spent, the request fails rather than silently returning a lesser read. A quietly downgraded analysis is worse than a refused one, because the reader has no way to know which they received.

Frequently asked

Can the narration hallucinate a price level onto my chart?

A hallucinated level is structurally unusable here: the grounding verifier rejects any sentence citing a number the engine did not compute, and a deterministic bullet replaces it. The count of dropped claims is disclosed on the read itself.