No book is cheapest in the abstract. Offers are comparable only when the player market, threshold, side, payout contract, settlement rules, timestamp, and model probability match.
No current offer feed or calibrated projection artifact is attached. The honest state is no comparable live quote set is attached.
Provenance tier: cross-book comparison methodology; no current line, payout, hit rate, financial return, or product-performance statistic is published.
Match the claim before comparing the payout
Two cards can show the same player name and still represent different bets. The statistic may differ. The threshold may differ. One side may include a push while another grades the same result as a loss, refund, or reduced tier. Participation rules can diverge. Official stat providers can revise the same play differently.
The comparison starts with a canonical market identity: event, player, statistic, threshold, side, period, participation condition, settlement source, and rule version. Offers that do not resolve to the same identity belong in separate rows. A lower threshold is not merely a nicer price; it is a different claim with a different probability.
When the mapping is uncertain, return “market match unresolved — review the product rules and player identity.” Do not force a match because the display labels look close.
A payout card is a contract, not a multiplier
Marketing surfaces compress settlement into one attractive headline. The actual contract may include perfect-entry tiers, partial-hit tiers, refunds, voided selections, reduced payouts, protected entries, dynamic pool shares, or restrictions on correlated combinations. A calculator must encode the whole state tree.
Store the rule version with the offer. For every possible settlement state, record which selections count, how voids alter the entry, what amount is returned, and whether the offer changes when a protected or promotional mode is selected. The expected result comes from the full tree, not from the top advertised return.
A missing branch makes the calculation incomplete. “Void treatment unavailable — load the current rules” is an error with a cure. Assuming the entry simply shrinks or refunds is not.
Exhaustive settlement prevents friendly math
Every terminal state should be explicit and mutually exclusive: full success, each supported partial-success tier, complete loss, void, cancellation, unresolved grading, and any product-specific adjustment preserved from the live rules. The probability assigned across those states must cover the whole contract.
This is where optimistic calculators cheat without looking like they cheat. They model only the headline win and the total loss, ignoring partial returns or adverse rule branches. The output appears precise because the omitted states are invisible.
The flex-versus-power guide explains why settlement shape matters. Any current product comparison still needs a current rule receipt; an evergreen description cannot stand in for one.
Timestamp every offer
Pickem lines move. Products can suspend a selection, change a threshold, alter a payout mode, or remove a player. A comparison built from screenshots captured at different times may combine offers that were never simultaneously executable.
Save provider, event identity, player identity, statistic, threshold, side, payout contract, quote time, receipt time, suspension state, and location or account eligibility. Then define a maximum comparison age. An offer outside that window is stale, even when it still looks plausible.
Do not replace a missing live row with yesterday’s line or a familiar market from another provider. The honest state is “offer unavailable for the declared window.” The cure is a fresh accepted quote, not an estimated one.
The projection must answer the exact market
A mean forecast is not automatically a probability. The calculator needs a distribution for the exact statistic under the information available at decision time. That distribution should account for playing time, role, opponent, team context, and uncertainty around missing or changing inputs.
If one product offers a different threshold, recompute the probability. Copying the same confidence score across several lines hides the most important part of the comparison. A small threshold change can matter when the outcome distribution is concentrated, and matter less when it is wide. The model should know the difference.
Store the model version, feature cutoff, source manifest, and calibration artifact with every probability. A number typed by the user may be useful for exploration, but it should be labeled user-supplied and should not inherit the model’s track record.
Calibration comes before expected value
Expected value magnifies probability error. A model that is systematically overconfident will make the most aggressive entries look the most attractive. Before routing any card, inspect whether predicted probability bands match observed frequencies on an untouched, eligible sample.
Keep the window and counts visible. Sparse bands do not become trustworthy because the curve is smooth. If the model has no relevant calibration slice for the sport, market, or player role, widen the uncertainty or withhold the comparison.
For spread-style selections, public records should use ATS wins and losses, hit percentage, window, and sample size. Player-prop models should publish the metric appropriate to the declared target and the exact settlement ledger. Neither case allows a financial result without real entry terms.
Correlation is not a decorative slider
Selections can share causes. A quarterback passing outcome and a receiver outcome may depend on the same dropbacks and game script. Two teammates can compete for the same opportunities. Weather, injuries, overtime, pace, and blowout risk can move several legs together.
Multiplying marginal probabilities assumes those relationships disappear. Sometimes that understates the joint chance; sometimes it overstates it. The direction cannot be chosen to improve the card. Use a declared joint model, a source-backed historical estimate, or a conservative range that exposes the unresolved dependence.
A slider with no provenance is a storytelling control. If the interface permits manual adjustment, label it as a scenario input and keep it separate from measured correlation. The user should see how sensitive the decision is to that assumption.
Provider restrictions are part of the market
A theoretically attractive combination may not be accepted. Products can block related selections, limit certain players, require a minimum mix, restrict entry types, or change availability by account and jurisdiction. The calculator should not rank an entry that cannot be submitted.
Represent acceptance as a state: eligible, restricted, suspended, unavailable, or unknown. Only an eligible, current offer belongs in the executable comparison. Unknown is not a soft yes.
The final route should include the reason. “Higher expected result under the declared assumptions” is different from “only eligible offer” or “other offers stale.” Without the reason, a ranking hides the condition that produced it.
Dynamic pools and fixed contracts need different forecasts
A fixed payout contract can be evaluated from the accepted settlement terms and the model’s joint outcome distribution. A pool-based contract adds another uncertain object: the final allocation of the pool among settlement tiers and winning claims.
Do not compare a provisional pool estimate with a locked fixed contract as though both were guaranteed. The pool side needs a distribution for final settlement, including accepted stake, pool movement, deductions, and the share assigned to the relevant tier. That uncertainty should survive into the comparison.
The parimutuel payout-curve guide covers the pool mechanics. A cross-book tool must label which products are fixed, which are pool-based, and which evidence each branch requires.
Line quality and payout quality are separate
A richer payout does not rescue a much worse threshold. A friendlier threshold does not guarantee value when the settlement contract is harsher. Compare the full expected distribution for each exact offer rather than ranking one visible field.
This also means the “same slip” often does not exist across providers. When lines differ, each card has its own probabilities and may contain different optimal selections. The tool should be willing to say the offers are not directly comparable and present them side by side without forcing a winner.
A sharp surface makes the tradeoff legible: threshold advantage, payout-contract advantage, correlation treatment, rule risk, and quote freshness. It does not collapse all of that into a badge with no receipts.
Promotions live in a separate ledger
A promotion can alter a specific entry’s cash flows, but it should not rewrite the structural quality of the underlying offer. Keep the base contract and promotional overlay separate. Record eligibility, activation, expiration, settlement, withdrawal conditions, and any non-cash credit treatment from the current terms.
When those terms are unavailable, do not estimate a face-value benefit. A banner is not a cash-flow schedule. The comparison can still evaluate the base offer and label the promotion unresolved.
Separating the ledgers also prevents a temporary campaign from contaminating the model’s long-run evaluation. Product quality, projection quality, and promotional value answer different questions.
Staleness can reverse the ranking
A cross-book decision is only as fresh as its oldest required input. One provider may update the player line while another leaves the old threshold visible but suspended. The model may rerun after news while the comparison still holds an earlier probability. Those mismatches create phantom value.
Use one decision timestamp and require every offer and model artifact to fall inside the declared freshness window. When any branch fails, mark the row stale and explain which refresh is required. Do not silently carry forward the previous winner.
The bet-tracking guide is useful because it separates the quote seen, the quote accepted, and the eventual settlement. The calculator needs the same discipline before it can claim a route was executable.
The output should be a typed decision, not a cheer
A useful comparison has a small set of honest outcomes. It can return a matched executable offer with its assumptions, a side-by-side comparison whose uncertainty overlaps, no eligible offer, stale data, unmatched contracts, unresolved correlation, incomplete rules, or model probability unavailable.
Each non-success state needs a cure. Refresh the offer. Load the rule version. Resolve the player identity. Recompute the exact threshold. Supply a calibrated probability. Remove a prohibited combination. Those messages teach the user what evidence is missing instead of nudging them toward a guess.
“No play” is not a product failure. It is the correct result whenever the inputs cannot support a comparison.
Grade the route, not just the entry
After settlement, preserve the candidate offers that were available, the selected route, the reason, the accepted terms, and the final grade. That lets the audit ask whether routing added value beyond the underlying projection. Without the alternatives, a winning ticket cannot prove the route was better.
Keep rejected, stale, and skipped candidates too. Selection bias appears when the ledger retains only the offers that produced an entry. The system should be able to reconstruct the full decision set at the cutoff.
Use the closing-line value guide for market-timing concepts, but do not force a sportsbook close onto a pickem contract that settles differently. Define the later reference appropriate to the product and preserve its rules.
The cross-book audit card
- Canonical event, player, statistic, threshold, side, and period.
- Provider, product mode, rule version, and eligibility state.
- Offer time, receipt time, freshness window, and suspension state.
- Complete settlement tree, including voids and partial outcomes.
- Model version, feature cutoff, and calibrated probability.
- Declared correlation method and uncertainty range.
- Pool forecast when the return is not fixed.
- Promotion overlay kept separate from the base contract.
- Accepted terms, route reason, and final settlement.
- Cure for every stale, missing, unmatched, or unresolved field.
Use Studio to inspect a projection only after the market contract is exact, and use the pickem surface only when it exposes real, current offers rather than filling the card with examples. A quiet surface with no eligible rows is more useful than a confident synthetic entry.
What would make the calculator honest
The calculator earns its name when it can reproduce every candidate offer at the decision time, evaluate the exact settlement contract with calibrated probabilities and declared correlation, and show why one executable route outranked the others. It must also be willing to return no comparison.
What to watch: the first source-linked offer set whose rule versions, timestamps, model artifact, correlation treatment, acceptance state, and settlement ledger all agree. Until that bundle exists, there is no cheapest book—only unmatched marketing cards.
Breakeven win rate at recorded American prices
Breakeven probability is calculated only from American prices that were actually captured in the odds-history table.
Prop hit rate versus recorded line distance
This chart remains empty until a verified source binds a player projection distribution, the offered prop line, and the settled result.
