How LineShift finds edges.

LineShift watches 4 sportsbooks in real time and runs every candidate bet through 9 independent analytical signals. A bet reaches your dashboard only when signals converge — when 5 or more fire in the same direction as the pick.

This page documents exactly what happens between a sportsbook posting a line and LineShift surfacing it. Every claim names a data source; every formula is standard textbook math; every limit we know about is published below.

Eight feeds. Zero scraping. Every number cites its source.

LineShift reads from eight live data sources and never fabricates values. When a feed goes quiet, the dashboard surfaces fewer bets rather than inventing one.

Nine independent signals. Every bet gets every signal.

Tap any slot for the measurement definition, a FIRED example with real numbers, and when it stays neutral. Per Session 62, no slot is ever N/A on a fresh bet — every signal either fires, contradicts, or explains why it sits neutral.

How many signals have to agree before a bet ships.

A bet only reaches your dashboard when enough of the 9 signals agree: NBA/MLB/NFL need at least 5 of 9 pointing the same way, NHL needs 4 of 9 (fewer independent signals feed reliably in practice). STRONG needs one more signal agreeing, TOP PICK needs two more.

Want the full rationale? See ADR-001 — Four-Gate Promotion Model.

How Claude judges every Top Pick.

After our four-check engine ranks a candidate, Claude reads the full multi-source context — injuries, breaking news, beat-reporter posts, weather, last 5 games, matchup history, sharp action, fan sentiment — and decides whether to confirm the engine's call, demote it (qualitative context is bad), or promote it (engine missed an angle).

Claude can never override the math. But it catches the things math misses.

What Claude can't do

• Override the four-check odds-range filter. Anything above +250 stays out of Top Picks regardless of how good the angle looks.
• Make claims that aren't traceable to a verified source. Every number, player name, weather reading, or sentiment claim must be findable in the bundle Claude was given. A source-traceability validator runs on every output.
• Promote a bet whose engine reading was DROP. The deterministic engine's hard-contradiction signal is non-negotiable.
• Name itself as the decision-maker. The engine is the spine; Claude is the second opinion. Every override is logged for the outcome record so we can see whether the AI's second opinions are actually adding lift.

Every shipped bet comes with a book, a stake, and a fallback.

A tip without a plan is useless. The BetCard's Bet Plan section names the specific book, the specific stake, and what to pivot to if the primary line moves or the book takes the price down.

1. 01

   Book selection (current price + historical soft-gap)

   We rank all 4 sportsbook prices for the exact side. The Bet Plan opens with the best-priced book pre-selected. For Pro / Max tiers, Session 72's calibration layers on the historical soft-gap — "DraftKings has been 1.3¢ softer than median on NBA player-points Over 90 days rolling" — with a tap-to-open 90-day chart.

   Example+121 DraftKings leads the 4-book group by 4¢. Calibration drawer shows DK has run 0.8¢ softer than FanDuel on this market across 63 days — 20 samples/day floor for display.
2. 02

   Stake sizing (quarter-Kelly with 2% bankroll cap)

   We use the Kelly Criterion — the math that tells you how much to bet given your edge — and shrink it to one quarter of full Kelly so a noisy edge estimate can\'t blow up your bankroll. We then hard-cap any stake at 2% of your configured bankroll. A slider on the BetCard lets you move between quarter / half / full Kelly if you know what you\'re doing.

   Example+6.2% edge × $1,200 bankroll × quarter-Kelly = $18.60 recommended stake. Hitting the 2% bankroll cap would mean a stake of $24.
3. 03

   Fallback (alternative market / leg if primary degrades)

   If the primary line moves against you or the book takes the price down so far it falls out of the gate, the plan shows a pre-vetted alternative — a correlated market or the second-best book. You\'re never stuck on a single line.

   ExamplePrimary: Player Over 1.5 3PM at +121. Fallback: Player Over 0.5 3PM at −210 (still profitable even after the line moves 10¢ against you).

CLV and line-shopping savings — the two honest scoreboards.

Both numbers get computed against data we capture automatically at bet-log time, so we can't fudge them after the fact. Sample charts below are illustrative — your real numbers live on Performance.

The things we refuse to promise.

We'd rather be trusted than loud. These are limits we publish on purpose — they're the boundary between analytics and the black boxes we're competing against.

• We don't predict winners.

  We find price-vs-probability gaps — if a book is pricing something at 42% implied and our signals say it's closer to 48%, that's the edge we surface. Whether it wins on any given night is variance.
• Historical soft-gap doesn't guarantee tonight's price is soft.

  Session 72's calibration tells you which book has historically been softest on this market per sport. That's backward-looking evidence, not a forward guarantee — tonight's price might be tight.
• 5 of 9 signals agreeing clears our gate — it doesn't guarantee the bet wins.

  The gate keeps thin bets off your dashboard. It does not bend variance. Even a 5/9 bet at +EV loses about 40% of the time.
• Kelly sizing assumes our edge estimate is correct.

  If our edge estimate is off, Kelly over-sizes. That's why we default to quarter-Kelly with a 2% cap — the sizing stays forgiving even when the edge number is directionally right but numerically noisy.
• No self-learning feedback loop at launch.

  Our signal weights are hand-calibrated. We have the foundation for per-user learning (accuracy-tracker.ts) but it needs a few hundred resolved bets per user before it can do anything useful. Until then the model you see is the model we built.
• We don't take bets or route money.

  LineShift is analytics. You bet with licensed sportsbooks directly — we deep-link you to their slip and stay out of the money trail.

The four formulas behind every LineShift number.

Nothing exotic. These are standard, published, verifiable formulas from sports-betting and probability textbooks. If you don't trust our numbers, you can recompute ours from the inputs below.

EV = (P(win) × payout) − (P(loss) × stake)

f* = (b × p − q) / b     where p = win%, q = 1 − p, b = decimal payout

P(A ∩ B) = p₁ × p₂ + r × √(p₁(1−p₁) × p₂(1−p₂))

count how many of the 9 signals point the same way as the pick — ship the bet only if that count clears the per-sport minimum.