01 The receipts
The single rule we won’t bend: every featured pick is written to our database the moment it fires and graded after the game ends. There is no “quietly delete the misses” option in our pipeline. The track-record card on the homepage reads from the same database that fires Telegram alerts to paid members — if it shows up there, it shows up here.
What this rules out. No retroactive deletions. No selectively-reported windows. No “lock of the day” we cherry-pick at the end of the week. If we sent a pick on Tuesday and lost, that loss is on the homepage Wednesday.
02 How the record is counted
The headline record and every window of the track-record ledger use the same public accounting rules:
- Tracked since March 26, 2026. That’s the first featured pick in the database. There is no earlier “backfilled” record.
- One bet counts once. Several of our models can fire on the same bet — same game, same side, same market. That’s useful confluence information (we show it on the card), but it is one bet, so it counts once in the record, graded from the earliest alert we sent. Before v2, each co-firing signal counted as its own row, which inflated the sample; v2 removed that inflation retroactively across the whole history.
- Captured price where available. Results use the price recorded when a pick was published. A documented standard-price fallback is used only when no price was captured.
- Governance exclusions stay auditable. A pick suppressed by a safety control, validation failure, or unsuitable market regime remains in the internal audit trail. Public performance includes only picks that were actually released as featured picks under the rules in effect at publication time.
- Pushes, voids, and pending picks never count toward the win rate — a push is not a win, and an ungraded game is not a result. Split games (our two feeds on opposite sides of the same market) are excluded from the combined pool, since claiming a win there would be cherry-picking.
The same rules apply to every published window and sport breakdown. If an accounting rule changes, the displayed history is recomputed consistently and the change is documented.
03 Model governance
A Sharp Signal is a model-generated candidate that has passed our current validation and live-performance controls. The exact features, transformations, thresholds, vendors, and model definitions are proprietary and are not shipped to the browser or published in static feeds.
- Historical validation checks that a candidate has enough evidence to be evaluated.
- Holdout and stability checks reduce the risk of promoting a pattern that only fits the data used to discover it.
- Price-aware live monitoring compares released performance with the expectations that supported activation.
- Fail-closed controls pause or withhold candidates when evidence, freshness, or data quality is insufficient.
Models are shown to members under pseudonymous labels. A label identifies a consistent model for product and performance reporting without revealing its internal taxonomy or recipe.
Historical performance is evidence, not proof of future performance. Model selection itself can introduce bias, so live results, sample size, and uncertainty remain part of the evaluation.
04 Game Reports
A Game Report is a separate single-game analysis stream presented in plain language. Its proprietary inputs and generation logic remain server-side.
When two published views agree, the product may mark the agreement. When they disagree, both released views remain visible rather than silently discarding one after the outcome.
Every Game Report carries one of six outcomes once the game is in: hit, miss, push, pass (the analyst recommended no bet), postponed, or pending (game hasn’t completed yet). Only hit and miss count toward the win rate. pass picks track separately so they don’t inflate the pending count.
05 ROI math
ROI is computed from released picks using the price captured at publication time wherever available:
- Captured-price result: a winner earns the profit implied by its recorded American price; a loss loses the recorded stake.
- Documented fallback: if no publication-time price exists, the record uses a consistent standard-price assumption and labels estimated figures where shown.
Pushes, voids, unresolved picks, and contradictory duplicate positions are excluded from win-rate math. Aggregate records count a released bet once even if more than one model supported it.
Known limitation. Small samples and missing price captures can make an individual model’s return estimate unstable. Treat model-level figures as descriptive, not as a guarantee or a substitute for the aggregate released record.
06 Sample size disclosure
Bigger samples generally produce more stable estimates, but no historical sample removes regime risk or guarantees future results. Published aggregate figures include the number of graded picks that produced them.
Short-window and model-level results can move sharply. Treat them as descriptive context, not predictive certainty.
07 Pauses & reactivations
Any feed on this site — a single signal, a sport, or an entire stream — can be paused or reactivated at any time. We treat that as normal operating procedure for a quantitative system, not an exception. Some of the reasons we’ve paused things in the past or could pause them in the future:
- Regime shifts — regular season → playoffs, rule changes, a new ball, a coaching-strategy wave that the training data didn’t see.
- Edge decay — released performance diverging materially from validated expectations.
- Data-quality disruption — a required input becomes stale, unavailable, or changes shape.
- Bookmaker model updates — when the market re-prices around a pattern we’ve been exploiting, that pattern can stop paying.
- Sample-size guardrails — a new or low-volume signal can be benched until it has enough live data to evaluate honestly.
When something is paused, picks stop firing publicly but still write to the database for forensics. When we reactivate, the historical record stays attached — we don’t reset the books to make the comeback look cleaner.
When evidence deteriorates or conditions change, affected picks are withheld while the model is reviewed. A later reactivation does not erase the earlier released record.
Why this disclosure exists. Most pick services bury their pauses and quietly tweak the model until the public stats come back to baseline. We’d rather tell you up front that regime risk is real, that we may pull or restore a feed without warning when our monitors say to, and that the historical record stays intact either way.
08 What you see on each pick
Pick cards are tier-aware. We pack different layers of context onto the card depending on what your plan unlocks — the underlying pick (side, price, score) is the same across paid tiers; the analytics layer expands as you move up.
- Edge sees, on every pick: the bet, the captured American odds at fire time, the best-of-book price across our 5 major US sportsbooks, live score as the game plays, and the resolved final score.
- Pro adds, on each pick from an active-roster model: the model name, its historical win rate, ROI (real-odds where verifiable), p-value, sample size, and the confluence / dispute chips that mark when two independent models agree (or disagree) on the same call. Pro also unlocks the cross-tier per-model leaderboard.
- Sharp gets everything in Pro plus each signal three minutes before it’s released to Edge and Pro, the daily model deep-dive page, and unlimited AI Betting Analyst queries.
See the full feature comparison for the per-tier breakdown.
What stays proprietary
To preserve whatever edge the signals have, we don’t publish:
- The exact computational definition of any model or signal.
- Feature sets, transformations, weights, thresholds, training data, and vendor or source mappings.
- Internal model names, pipeline topology, operational telemetry, and unreleased candidates.
If we published the exact recipe, the books would price around it inside a week. The track record, the methodology, the grading rules, and the per-pick metadata your tier unlocks are all on the table — the recipe is what stays in the kitchen.
09 Monitoring and pauses
Released models are monitored for performance drift, price-aware results, sample adequacy, and input health. Internal limits are deliberately not published because they are part of the proprietary decision system. Controls can pause a model when:
- live performance materially diverges from validated expectations;
- the available sample or captured-price coverage is insufficient;
- an input is stale, incomplete, or inconsistent; or
- the market regime no longer resembles the conditions under which the model was validated.
Paused models stop releasing featured picks. Their prior released record remains intact, and reactivation requires a fresh internal review.
Questions about a published aggregate can be sent to hello@sportsgods.com. We can explain the accounting basis without disclosing proprietary model internals or private data.