Every value flag the model raises, logged before the match and graded in public. Paper record — no real money.
Flag-time entry. Each simulated bet is priced at the odds available when the model first flagged the edge — not the hindsight-friendly closing line. Stakes are a flat 1 unit, always.
Fair probabilities. Bookmaker margin is removed with the Shin method; an edge is the model's win probability minus that fair market probability, and only edges inside the working band become flags.
Why CLV is the referee. Closing line value asks: after we flagged a side, did the market itself move toward it by close? It needs no match result, so it converges 5–10x faster than ROI and is much harder to get lucky on. Positive median CLV is the fastest honest evidence of edge; ROI follows only at large samples.
Flags appear in the pending table before their matches start and are graded automatically after. Nothing is ever edited or deleted.
All flagged candidates regardless of band: 213 flags, 177 settled, P/L -22.21u, ROI -12.6%.
Ledger rows unavailable right now — refreshes after the next morning update.
Ledger rows unavailable right now — refreshes after the next morning update.
In-band flags with a valid entry→close pair: 79, median +0.00pp, mean -0.04pp. Beat the close on 29, tied 18, worse on 32 — beat-rate among movers 47.5%.
n = 47 · median CLV -0.05pp
Entry graded against a real captured closing line (close capture live since ~2026-06-22).
n = 32 · median CLV +0.00pp
Older flags graded against the last hourly capture (~T-30min) — a stand-in, not a true close.
The two referees are reported separately and never pooled — an hourly-era “close” is not a true close, and mixing them would flatter (or smear) the record silently.
| book | pairs | median clv |
|---|---|---|
| bet365 | 29 | +0.00pp |
| pinnacle | 50 | +0.00pp |
Across all flagged candidates regardless of band: n = 188, median +0.00pp.