1. Pick your market. Click only facts you've observed — the tool never predicts your conditions, you tell it what the night did.
2. Every row is a question about the rest of the day. The big number = how often it happened on days exactly like this one since 2018.
3. Read the receipts before trusting a row: ±X vs base = how much your conditions moved the odds off normal (fat move = your conditions matter). n = how many real days back the number. Badge: DURABLE = the edge pointed the same way in 2018-20, 21-23 AND 24-26; MIXED = eras disagree — treat as weather, not physics; THIN = not enough days, don't trade it.
The ± on every number is its 95% wobble. "72% ±4" means the true odds live somewhere in 68-76 — trustworthy. "65% ±12" means the sample is too small to pin down, and the badge flips to THIN automatically. If the ± is bigger than the edge, there is no edge — there's a guess.
Rule: act on DURABLE rows where the delta is bigger than the ±. And go easy on day-of-week: it slices the sample 5 ways before you've clicked anything else — use it alone or with ONE other condition, never stacked deep.
Calibrated. We graded 108,339 of our own walk-forward estimates against what actually happened. The numbers came back honest overall, so they are shown as computed — except four outcomes that were off by 3+ points in the same direction in 8 of 8 years, which carry a fixed correction. Ten others wobbled year to year and were deliberately left alone: correcting a wobble is fitting noise. The full record is on the Scoreboard.
The data: our own tape — NT minute bars 2006→today (GC from 2008; overnight coverage gets dense from 2009), clean continuous build (dominant contract, back-adjusted, spike-filtered). Nothing here is fitted or optimized; every number is a counted frequency, and the DURABLE badge is the out-of-sample test: three eras counted independently must agree. Size labels (gap/Asia range) are relative to the trailing 20-day average — "LARGE" always means large for current conditions.