About Drake
How Drake performs, grades itself, and works.
Drake is a personal, multi-agent equity research engine. On request it gathers market data, fundamentals, live news, and macro context; runs six specialist analysts in parallel; forces an adversarial bull-versus-bear synthesis; and emits one structured decision — BUY, SELL, HOLD, or ABSTAIN — with an upside probability, a disagreement-adjusted confidence, and a falsifiable "what would change my mind."
It is a research engine proving a calibrated pipeline — not a P&L claim. Small N. We say so everywhere.
How Drake performs
Performance here means one thing: calibration — whether Drake's stated probabilities match what actually happens, graded against the future on each call's resolution date, never backfilled. That is the only honest scoreboard, and today it is deliberately empty.
A hit rate or a Brier score computed from a handful of calls is noise dressed as signal — a lucky week can flatter a coin flip. So we suppress those numbers entirely until at least ten calls have resolved. Until then the scoreboard shows the honest truth: 0 of 147 resolved, and the first resolves Jul 9, 2026.
The reliability curve appears after 10 calls resolve.
It plots predicted probability against realized frequency — a perfectly calibrated engine sits on the diagonal. An empty curve is the correct state today, not an error: we won't plot a diagonal from zero points. First resolutions land Jul 9, 2026 onward.
How grading works →Per-specialist accuracy appears once calls resolve — which of the six lenses actually predicts is itself something we grade, dated to Jul 9, 2026.
Every day, Drake grades itself.
Grade matured calls → learn from confirmed signals → improve the analyst. The mechanism runs at full weight today; the results are dated-empty, because nothing has resolved yet. Learning is keyed strictly to resolved market outcomes — never to whether anyone agreed with a call.
148 calls logged. 0 matured. The first grades Jul 9, 2026 — then daily as calls reach their resolution dates.
No calls have matured yet.
The most-recent graded calls — HIT or MISS, realized vs. predicted, each linked to its run — appear here once the first resolves Jul 9, 2026.
Nothing learned yet — learning is keyed strictly to resolved market outcomes, and none have resolved. When a specialist's directional calls prove out on resolved names, its conviction is weighted up; when they don't, it's weighted down. Outcomes only — a user agreeing with a call never feeds a recommendation.
No adjustments yet.
Every change will be dated, outcome-based, and tied to the resolved calls that triggered it — never to whether anyone agreed with a call.
The method — six views, one adversary, one calibrated verdict
No single model decides. Six specialists each look at the same name through their own lens. A separate bull/bear stage — deliberately run on a different model family for genuine architectural diversity — argues both sides. The synthesizer then blends the views, and where the specialists disagree, confidence is lowered mechanically, never averaged away.
Six specialists
Bull / bear adversary
An adversary argues both sides on a separate model family — so a single model's blind spot can't carry the day.
Disagreement-adjusted verdict
More disagreement → lower confidence, mechanically. Keyed to a resolution date and graded against reality.
The safeguards — the part a skeptic should read first
These are not marketing claims. They are rules enforced in the code, with tests that reject violations. The two that matter most lead.
Degraded or untrustworthy input produces ABSTAIN — never a default BUY or SELL. There is no heuristic fallback path; discipline beats a confident guess.
Drake never shapes an output to please you. It surfaces the bear case on names you love and will tell you when you're miscalibrated. Your stored preference cannot change a recommendation — that's a release-blocking test.
Every displayed figure traces to a source record. The model never originates a number; "not available" is shown as "—", never as a fabricated 0.
Confidence is monotonically decreasing in specialist disagreement, with a hard cap when sources are thin. Conflict is never averaged into false certainty.
Every call is graded against the future, on its resolution date — never backfilled to look better in hindsight. That's why the scoreboard starts empty today, on purpose.
Stop-losses live at the broker, not in a script that can crash. The downside protection survives a process death.
A day down 6% halts all new entries automatically. Plus risk-first sizing (0.75% per trade), position and deployment caps, and an earnings blackout.
Every order's sizing is recomputed independently before placement. A dynamic, boundary-evaluated kill switch can halt trading in-flight, default off. Human-in-the-loop — never full-auto by default.
One name is permanently excluded from analysis, watchlists, and trades for compliance, enforced at the data layer — it cannot appear anywhere on this site.