Sports projections made simple.

• Line · the number the bookmaker is offering Over/Under.
• Projection · the model's estimate for that player tonight.
• Gap / edge · how much higher or lower the projection is vs. the line, in percentage points.
• Side · Over if the projection is above the line, Under if it's below.

• Category A · model and market differed by 5pp or more. Settled at 49.3%.
• Category B · differed by 2.5–5pp. Settled at 50.0%.
• Category C · differed by under 2.5pp, or the row was anomaly-flagged. Settled at 51.0%.

Those rates are measured over 21,192 settled outcomes, and they run in the opposite direction to what the old labels implied. A larger disagreement with the market has historically settled worse, not better — which is why these are neutral letters now, and why no category is promoted above another anywhere on the site.

• Parlay-slip persistence so candidate slips can be graded with a real hit rate after games settle.
• Wider market coverage on NBA/MLB game lines (moneyline, spreads, totals already shipped for NBA playoff games).

• NBA rebounds — the strongest cohort on record. The model has stable signal on REB projections.
• NBA points + assists — barely above coin flip on a large sample. We surface these projections but treat them as watch-list calls, not high-confidence reads.
• MLB strikeouts — smallest sample of any market we cover and below coin flip so far. The variance profile of pitcher hooks + manager decisions makes this an honest weak spot.
• MLB confidence climbed back into "watch" on the May 22 settlement — as of that date High was 49.7% on 396 settled rows, Medium 50.4% on 141, Low 53.3% on 435. These are the May 22 figures, not current ones; the live rates are in the category captions above. Low is still the best MLB cohort, but only ONE rival now beats High by ≥1.5pp (was both before May 22). The calibration overlay auto-promotes MLB High from a "Needs more tracking" downgrade back to its raw label. The decision rule is pinned by tests — we only invert when ≥ 2 rivals beat by ≥ 1.5pp, so a single best-tier (Low) can't trigger inversion.
• Monte Carlo internal validation — two-date check — May 22 looked promising (MC Strong 62.5%, Watch 65.2% on 311 joins). May 23 reverted to roughly coin flip (Strong 50.0%, Watch 29.4% on 287 joins). Two dates is too small to draw a conclusion. Across both: Strong 56.7% (17-13), Watch 50.0% (20-20), High-variance 50.2% (259-257). Promotion to production scoring is on hold until ≥ 5 dates show consistent separation.
• Curated rail is outperforming parlays meaningfully — across the first 2 days of tracking, single-leg curated picks are 8-4 (66.7% on 12)while multi-leg saved parlays are 6-44 (12.0% on 50). MLB-only curated picks are 5-1 (83.3%). The honest read: selectivity over volume is working; correlation risk is brutal on 4-5 leg slips. We surface both tracks but expect users to weight the curated rail more heavily.
• Curated rail prefers selectivity over volume — the homepage "Tonight's curated projections" rail picks up to six leans per slate by edge × calibration-adjusted confidence × market strength. Inverted (sport, tier) combos are excluded. Better to see six trustworthy reads than 300 of mixed quality. The picks are saved before games via pipeline.snapshot_curated and graded after settlement via pipeline.grade_curated — so the curated rail will eventually carry a real, auditable hit rate of its own.
• Calibration is now derived from the live audit — the confidence overlay reads model_audit.json every render. When the nightly settle adds more data, labels adjust automatically. We fail closed: thin samples stay informational and inverted tiers are flagged. No category is ever promoted above another: on settled data the categories run in the opposite order to what their old names implied, so none of them earns priority.

Numbers are pulled from Results. Sample sizes are still small in absolute terms — anything you read here is a record, not a forecast. No 80%-accuracy claim is made anywhere on the site, and won't be until we run a real out-of-sample backtest.