Decision Log
Ten independent agents score every covered name from real price history and combine into one BUY / SELL / HOLD call. When that call changes on a name, the flip is logged — the prior call, the new call, the conviction behind it, and the one-line reason drawn from the agents' own votes. The log is persisted: it survives restarts and is never back-filled.
How the agent weighting adapts
The ten agents' voting weights move with how well each agent's past calls scored against real outcomes, within a published range. Hit-rates carry their 95% confidence range and the count of distinct graded calls behind them; excess edge is measured vs the market benchmark over the same window. A figure not yet earned from real outcomes shows as “—”.
The learning disclosure is momentarily unavailable — nothing is estimated in its place.
Method: bounded, disclosed online reweighting — every change is graded against real outcomes and replayable · Permanent equal-weight baseline: every agent at 1.0× — the unlearned comparison never goes away.
A steady desk adds no rows by design.