Built to be trusted
Every conclusion has a receipt.
The biggest worry institutions have about AI is "where did that answer come from?" LyraMind answers it by construction: click any conclusion and see what came from the market, what LyraMind calculated, and what it merely inferred, with sources, dates, and both sides of the evidence.
A Trust Card, on every major insight
AMD thesis assumption 'hyperscaler capex stays strong' has weakenedINFERRED · conf MEDIUM
Observed, reported by an authoritative source
Microsoft datacenter spending rangelowered (FY)SEC · 2026-08-15
LyraMind inferred, interpretation, not fact
AMD options positioningmildly defensiveinference · conf 0.5
Supporting
- Microsoft lowered its datacenter spending range (primary source, dated)
Contradicting (kept visible)
- Google maintained its capex guidance
Thesis status and 'reads' are LyraMind interpretations, not facts. Every observed value shows its source and date. Interpretation, not investment advice.
Four evidence classes. Every statement is labelled
OBSERVED
Directly reported by an authoritative source (an SEC filing, a BLS release). A fact.
CALCULATED
Deterministic math on sourced values, with the formula shown. Auditable.
INFERRED
LyraMind's interpretation of the evidence. Clearly not a fact.
SPECULATIVE
Low-confidence causal / forward-looking read. Uncertainty made very visible.
Five rules we never break
1. No unsourced number. Every figure carries its provider, timestamp and as-of date.
2. No fabricated source, citations come from retrieved documents, never invented by a model.
3. No causal claim without classification: "coincided with," not "caused," unless the evidence is strong.
4. Stale data is visibly stale: "13F: quarterly disclosure, not current"; "COT: weekly, as of Tuesday."
5. Contradicting evidence stays visible, the model can't cherry-pick only what supports its answer.
Licensed by construction
Every field references a source in a licence registry with its rights (display, redistribution, storage, retention). A production gate blocks any data from a customer-facing answer unless it's approved and its rights are reviewed, so nothing unlicensed or unverified ever reaches you. And a minimum-sample rule means a "hit rate" is never shown on too little history.