Kickstarter Serve

Bet = YES is the frozen rule and nothing else: edge ≥ 2.5 points and the move filter reading toward, inside the 0.20–0.70 band (0.80 until 30 September — everything at 0.70 and above is the favourites book's now, see below), overround ≤ 1.06, in universe. Need is how many points inside fair you have to rest to make a side worth taking — negative means fair itself is already good enough. Stake is 1u flat: the frozen spec's halvings (ATP Challenger/125 ×0.5 and the 0.20–0.30 price slice ×0.5) were retired on 4 October — no money ever ran them, the trader bets these in full — and the record was restated to match. Qualifying is staked like anything else. missed is a one-off and will not recur: the row clears every rule as it stands today, but it was scored before the rule changed on 16 September, and a match that has already started is never re-scored, so its verdict can never be revised. Fourteen sides on 14–16 September are in that state — Leme Da Silva v Sharma among them, at +2.65 with the move toward. They are deliberately NOT counted as picks anywhere in the P&L or the line-move panel: the book did not fire on them at the time, and backdating a bet nobody placed is how a track record stops meaning anything.

The favourites selector is a second, separate rule on the same scores, frozen 14 September. It reads the Price Edge column — the same edge ÷ fair shown there — core at ≥ 2.5%, ext at 1.5–2.5%, both needing the move toward and the row in universe, in the price band 0.70–0.95. The column shows the number up to 1.00 because that is a favourite's price either way, but nothing above 0.95 is ever a pick — the 0.95 cap is part of the frozen rule, and every attempt to re-cut the band failed to transfer between periods (rank correlation −0.81 design to test). Stakes are its own: core 1u, ext 0.5u — qualifying is not halved (lifted 30 September) and Challenger/125 is full size, as it now is in the general book too. Bet = YES(Fav) marks a row this rule took and the general rule did not. These picks are not in the P&L by default — the Profit page has a Regular Model / All Models switch for that.

Since 30 September the two rules divide the price line rather than overlapping. The general band stops at 0.70 and the favourites book owns 0.70–0.95 outright, so YES(Fav) is now every favourite the book bets and not only the ones above 0.80. Nothing is given up in the strip that moved: a general pick below 0.80 has edge ≥ 2.5 points, hence ≥ 3.1% relative, hence a favourites core pick already — 263 of 263 on the backtest rows. What changes is the stake and the execution. Stake starts at the flat tier — 1u core, 0.5u ext — and is then sized: the handoff's tercile multiplier, terciles of edge × terciles of mvhat, each 0.5/1.0/1.5, normalised over this book's trailing 24 months so its total planned stake is unchanged. Clips run 0.05u to about 2u. Sizing was switched on 30 September after the pipeline was reconciled against the package's own scorer: run as-if over five recent dates, 814 sides, every side agrees exactly on fair, cal, edge and mvhat once our own Pinnacle-close fill is disabled, and with it enabled median |Δmvhat| is still 0.000000 (p90 0.0069). The favourites rule here selects exactly what the package's fav_rule selects on those dates, with no exceptions. Two cells are withheld entirely, WTA 1000 main draw and ATP Challenger qualifying; they also read Fav(hold) and come back at half stake if their fills show win-vs-close ≥ +1.0 over 100. And these orders take the offer at or within a cent or two of fair instead of resting a cent under it: at 77c a cent is 1.4 ROI points against a ~7% book, resting filled 23% of favourites picks and filled them adversely, and the live order log is blunter still — every band under 0.70 filled 100% of its picks and 0.70+ filled 0%.

Price Edge is edge ÷ fair — the frozen first-sighting edge as a share of what the contract costs, i.e. the model's expected ROI on stake if bought at the open. It is context, not a gate: the rule reads the point edge. Set Current rules and the Bet column re-decides every row against today's per-level edge floors, read live from the same config the Kalshi autotrader trades from — so this view and the bot cannot disagree about what is a pick today. Nothing frozen is rewritten: the verdict each row was scored with stays the graded book, and where the two views differ the Bet cell's tooltip says so and names the floor for that row's level. The differences run both ways here, because the slate shows every side: a level whose floor came down surfaces sides the old flat floor said NO to, and a level whose floor went up drops sides it said YES to. Floor (as scored) stays the default and remains the graded record. The split rule that used to occupy this switch is parked for 2027, paper only — its numbers live in FINAL_RULESET §1, not here.

Fair 0.512* — the asterisk means valuebettennis had no open for this match yet, so fair and quote are the first price Pinnacle showed instead, frozen at that first sighting and not re-read as the line moves. These rows are picked and bet like any other, and have been since 16 September. Until then they could not be: the 2.5-point floor, the price band and the per-segment calibration were all fitted against the feed's opens, so firing the rule on a substitute read a threshold off the wrong scale. That objection died with the thing it protected. The feed's open broke on 14 September — its median overround went from 1.0566 across the 7,111 matches the model was calibrated on to 1.0928 after, and 61% of what it now publishes fails the 1.06 overround cap outright. Our own first sighting sits at 1.0477, which is nearer the calibration regime than the feed's current number, so the asterisk had been excluding the better of the two prices. The residual objection is real and is not hidden: a first sighting is not literally an open, and it differs from the feed's published open by a median 2.08 points. What makes it tolerable is that the gap does not shrink as the sighting gets earlier, which is what it would do if the gap were our lateness rather than the feed's error. The mark stays on the page as provenance, not as a warning — the two price sources must stay tellable apart, because comparing them is the only way this change is ever judged. Use Open above to see only one kind or the other.

Open to close is the market's opinion changing, not the model's. Fair already is the open — the feed carries Pinnacle's opening price and nothing re-quotes it, so every run of a day stores the same fair for a side and there is no intraday move to show. close is the closing line and Line move is close − settled open in points (since 4 Oct 2026, the backtest's definitions on both ends): on a feed-sourced row the settled open is the fair itself, and on a Pinnacle-sourced row it is the feed's reading of the open — usually the same number as our first tick (median gap 0.000 pts, measured 4 Oct), but where the two disagree it is the backtest's column that counts. A row with no settled open at all stays blank rather than grading on the feeler; fair is still on the row, so nothing raw is lost. Positive means the market came towards this side after it opened and the open was the better price. That is closing line movement, and it is deliberately not called CLV: closing line value is the close minus what was actually paid, a different and flattering number, because §2 rests a cent under fair and that discount alone is worth about a point. CLV in that sense lives on the Kalshi page, split from the discount. Both are blank until the match is graded and the close arrives, usually a day later, and blank for a void: a price a market was suspended at is not its last word. Line move is the honest scoreboard while the settled sample is still too small to read — it says whether you were on the right side of a moving line, on every bet rather than on the handful that have finished. A close marked * came from our own Pinnacle poller rather than the feed, because the feed never published that match — on 2026-09-14 that was 107 of 221. It is the last price Pinnacle showed before the match left the pre-match board, which is the same thing the feed's close is; the mark is there so the two are never silently mixed. † is the same source but a weaker claim: the poller lost sight of that match more than ten minutes before the advertised start, so the number is the last price it caught, not the last price on the board. Its line move is greyed, because the gap between those two is not a real move.

ROI is model ÷ price − 1, the expected return per unit at that price. quote is Pinnacle's raw vigged price, fair is the devigged open, and −1c/−2c are one and two points inside fair — the prices you would actually type into Novig. Placing an order moves the row onto the Book, which is where the venue, stake and fill get recorded.

Kalshi take is the same ROI if you cross the spread at fair on Kalshi instead of resting on Novig: model ÷ (fair + fee) − 1, with the taker fee 0.07 × p × (1−p) added to what you pay. It is the answer to "is it still worth it if I just take it now". Hover a cell for the fee in cents. The fee is a curve, not a haircut — it peaks at 1.75c at a coin-flip price and falls to about 1.1c at 0.20 or 0.80, so it eats a far larger share of a thin edge in the middle of the board than at the edges. Against the 2.5-point floor that is roughly 70% of the edge gone at 0.50 and 45% at 0.20. The maker fee is a quarter of it, 0.0175 × p × (1−p), which is why the frozen rule posts rather than takes.