Six specialists in parallel
Every recommendation begins with a parallel fan-out across six specialist analysts — Fundamental, Technical, News, Macro, Geopolitical, and Industry. Each is a separate prompt to a frontier model, run with a tightly scoped role and the same point-in-time snapshot of the company. Specialists don’t see each other’s outputs; their opinions are formed independently.
Adversarial bull/bear synthesis
A separate orchestrator inspects the specialist verdicts and forces an explicit bull-versus-bear argument on an alternate model family. The synthesizer reads both sides and produces a single Recommendation with a probability of upside, a confidence, and a horizon. Confidence is then mechanically lowered by the disagreement spread — a 6/6 unanimous BUY carries more confidence than a 4/2 BUY even if the model said otherwise.
No backtests, no hindsight
Drake’s calibration record is forward only. Every call carries a resolution timestamp; when that date arrives the realized price is fetched and the prediction is graded against it. Backtests over web-search-trained models leak hindsight, so we don’t publish them. The only honest metric is hit rate and Brier score on calls that have matured since the model last saw the world.
Abstain is a first-class output
When the specialists can’t form a quorum, when the synthesizer fails its
retries, when the data quality gate fails, the recommendation is ABSTAIN,
not a defaulted HOLD. Abstain pages distinguish between model
abstentions (Drake chose discipline) and infrastructure abstentions (an
API call failed) so an outage week doesn’t read as discipline.
What each page contains
The per-run page renders the synthesizer’s editorial reasoning — opening, setup, conviction, risk, verdict — alongside the underlying market snapshot, the per-specialist verdict, the catalyst calendar, and a citations block listing every external source the live web search returned. The JSON-LD payload includes the underlying ratings and reviews so AI search engines can attribute statements without scraping the prose.
For AI crawlers
See /llms.txt for a curated index. Tagged
claims (FACT / INFERENCE / SPECULATION) are
embedded in the JSON-LD so a citation can be marked appropriately. Drake is not a
recommender system; it is a research engine whose accuracy is publicly tracked.