How the court weighs evidence

1. A reading becomes testimony

Each witness has a disclosed rule: a threshold, year-over-year change, trend comparison, or regime test. The rule maps the latest valid observation to a signal from −2 (strongly favors caution) to +2 (strongly favors growth).

2. Relevance changes by horizon

The VIX can matter over the next few months and say almost nothing about the next decade. A relevance mask determines whether each witness may testify at short, medium, and long horizons.

3. Credibility becomes weight

Each month in the available FRED history is treated as a trial. The engine freezes that month’s testimony, measures the S&P outcome 3, 12, or 60 months later, compares it with the unconditional base rate, and publishes hit rate, directional excess return, sample size, and worst miss. Small samples are shrunk toward 50 so they cannot dominate.

The revision-bias receipt

Published credibility currently uses the latest available FRED history. That makes the run reproducible, but revised economic series may look cleaner than they did in real time. Every witness page says so. Time Machine replays use point-in-time ALFRED vintages; a future backtest hardening pass will do the same for every monthly trial.

Time Machine

Historical replays fetch the ALFRED vintage available on the selected date for revision-sensitive data. Market-price series without vintages use dated FRED history clipped at the case date. Later observations stay out of the witness box, missing archives remain unavailable, and the known outcome appears only after the verdict as hindsight.

4. The verdict keeps its uncertainty

The weighted dispersion of active testimony appears beside the composite. Missing or stale sources are removed visibly. The unconditional base rate stays on screen because a signal that cannot beat “markets usually rise” is not useful evidence.

What this does not do

• No personalized output or portfolio recommendations.
• No point targets, guaranteed outcomes, or buy/sell instructions.
• No fabricated continuity when a source goes offline.
• No LLM in the scoring path.