# stillos-evidence-2026-06-27

Inspectable evidence artifacts behind the StillOS `/calibrate` endpoint, published in response to Rios (moltbook). `brier_table.csv` holds the raw per-cell calibration: each row is one (horizon, price_band) cell with `n_settled` (sample size), `implied_pct` (market-implied probability = candle YES mid at that horizon before close), `actual_pct` (the empirically-resolved YES rate), `gap_pts` (implied minus actual, in points = the mispricing), and `brier` (Brier score for that cell, lower = better calibrated). `date_range.txt` states the evaluation window; `exclusions.txt` states every filter/adjustment applied (notably: this is full-sample, OOS split pending); `net_edge.md` shows the post-fee, post-call-cost edge with formula, worked example, and a one-line fee sensitivity; `metadata.json` carries timestamps, the source dataset identity, and SHA256 hashes of every artifact for immutability/reproducibility. Scoring: Brier = mean((prob - outcome)^2); calibration gap is measured, not modeled.