Methodology

For any American price a:

d = 1 + (a > 0 ? a / 100 : 100 / |a|)
q = 1 / d   // raw implied probability (still vigged)

Raw probabilities q_i sum to more than 1 — the excess is the book's vig. De-vig removes it.

Simplest method — divide each side by the overround.

p_i = q_i / Σ q_j

Solve for exponent r ≥ 1 such that Σ (q_i)^r = 1. Bisect on r in [1, 8]. Default for 2-way and 3-way markets — preserves relative pricing between favorites and dogs better than multiplicative on high-vig books.

Σ (q_i)^r = 1   ⇒   p_i = (q_i)^r

Two layers. First, sharp consensus: per-book de-vig the sportsbook prices (Pinnacle, Circa, Betfair Exchange weighted 3×; Novig / ProphetX 2.5×; Bookmaker 2×; every other real book 1×), then sharp-weighted-average across books and renormalize to Σ = 1. Second, prediction-market anchor: Polymarket and Kalshi midpoints, multiplicatively de-vigged when both sides are quoted, then blended as a low-weight check on the sharp consensus.

p_fair = (1 · p_sharp + w · p_anchor) / (1 + w)   then renormalize Σ p_fair = 1

w ∈ [0, 0.25] scales with anchor liquidity (log10(1 + max-liquidity) / 4, capped).
The 0.25 ceiling keeps the anchor a check on sharp consensus, not a 50/50 co-author —
prediction markets carry overnight volume and lag sportsbook close.

Take the minimum fair probability across the two de-vig methods on the raw (vigged) consensus implieds. A pick that still shows +EV under the conservative estimate is the most defensible — fragile picks (positive central EV but negative worst-case) are kept but proportionally penalized and capped at the Lean tier.

p_worst = min(p_mult, p_power)   on the median-across-books raw implieds (Σ > 1)

EV%   = (p_fair × d_offer − 1) × 100
kelly = (p × d − 1) / (d − 1)   // bankroll fraction
Parlae displays 0.25 × Kelly by default (quarter-Kelly).

A single score per pick combining:

• EV% magnitude (most of the weight)
• Market width in cents (tighter → higher)
• Books contributing (more independent prices → higher)
• Anchor inclusion (Pinnacle/Circa present → higher)
• Movement signal — line drifting toward our side

Tiers: S ≥ 80 · A ≥ 65 · B ≥ 50 · C < 50. Thin markets (< 3 books, no sharp anchor) are forced to Lean, capped at conviction 40.

The 10-game strip and "Recent form" bars are sourced from real settled box scores — never modeled or padded. For each player we read the last N official games for the matching stat, compute over/under the offered line, and require N ≥ 3. Fewer than 3 → we say so ("Not enough settled games (n=X)") instead of inventing a number.

Every fire is captured with exact price, line, book, fair odds, and timestamp. At kickoff we re-read the market and record a median closing fair price (requires ≥ 2 books at close). CLV is never re-priced after the fact.

CLV_pp = (p_close − p_fire) × 100   // in probability points

We lead with average CLV and sample size, not binary win-rate. Samples are marked "accruing" until N ≥ 200 — that's where confidence intervals tighten meaningfully.

The no-vig calculator and the no-vig fair-odds tool let you paste any two-way or three-way set and reproduce our power / multiplicative numbers exactly. If a Parlae fair price is ever off by more than 0.5pp from what those tools compute for the same input, that's a bug — file it and we'll fix it publicly.

Every captured +EV pick is graded against the official final at the price we fired at — never re-priced, never backfilled. The score itself comes from a small set of providers. We are explicit about which sport goes through which path so you can audit gaps in the published numbers on /track-record.

• Game-line markets (h2h / spread / total) resolve via game_state.status = "final". SGO publishes finals for the major US books' feed; ESPN's public scoreboard is the fallback for MLB / NBA / NHL / NFL / NCAAF / NCAAB / WNBA / MLS. A cross-provider matcher bridges sgo:* event_keys to espn:* / apisports:* finals by team-name + commence-time so the same pick can be graded from either source.
• Player props (prop:*) resolve via player_game_logs — ingested from ESPN boxscores (US sports) and API-Sports (soccer). The grader looks up the player's actual stat line and compares to the line.
• Coverage gaps we have today. SGO does not always publish finals for tennis, MMA, KBO, NPB, or niche soccer leagues, and ESPN does not cover all of them either. Outright / futures markets (e.g. golf tournament winner) need a dedicated grader that we have not yet written. Picks in these markets sit Pending until coverage catches up.
• Markets we deliberately do not grade. Half-period markets (h2h_1st_half, total_2nd_half, …) and 3-way markets (moneyline_3way, 1x2, draw_no_bet) need either period scores from a richer feed or a dedicated 3-way grader. The full-game grader is anchored on exact h2h / spread / total only — so a half-period market can never be silently graded against the full-game score (which would give the wrong answer; e.g. a 1st-half "Draw" can still be a winner even when the full game ended decisively). Such picks stay Pending rather than misreport a result.

The pledge. We never auto-void or exclude a pending pick to inflate the published win-rate or ROI. The headline numbers on /track-record are computed only from graded picks; the Coverage % badge tells you what fraction that is, and a per-sport Stale count flags any pick whose game started more than 72 hours ago and still hasn't been graded — those are real coverage gaps that bias the published sample, and we want them visible, not hidden.