Methodology
Built against PRD v8.0. The north star is a decision-grade answer with evidence, not a recommendation slogan. OPEN financial, licensing and regulatory parameters stay OPEN until an owner decision.
A historical decision may only use information available at its decision timestamp. Effective, available and capture times are distinguished in the data layer — never only in a prompt. Agents cannot bypass this by querying a raw table.
Numerical models produce forecast numbers on versioned targets. Language models interpret evidence, classify events, explain, identify risks and synthesise. They never invent a stock price or numeric forecast.
Specialist agents run without seeing each other when independence is required. The supervisor receives structured results only after they are recorded. Disagreement is shown on purpose.
Public research is approved by a human. The research supervisor prepares a decision brief. It cannot publish.
A published snapshot is never overwritten. A changed view is a new snapshot with lineage. Misses stay on the ledger.
Critical data, model or governance failures block publication. The system may return insufficient evidence or not calibrated instead of forcing a prediction.
Accuracy is reported per task class, horizon, coverage and universe, always with a benchmark. There is no single ‘AI accuracy’ number.
Every strong bullish conclusion is challenged. Bull, base, bear and contrarian cases carry assumptions, evidence, and invalidation — not decorative probabilities.
DRAFT → REVIEWED → APPROVED → PUBLISHED. Published is ACTIVE until the horizon resolves, then RESOLVED against the benchmark. A new forecast marks the prior SUPERSEDED (retained). Failed governance is BLOCKED (retained).