How TrueSplits works

• Play-by-play & snap counts — via nflverse (CC-BY-4.0). The source for every computed fantasy point, the QB on the field, game script, and the full-game flag.
• Teammate on-field participation — FTN Data via nflverse (CC-BY-SA-4.0, credited to FTN Data). This dataset only refreshes after a season, so any teammate-on-field split is prior seasons only and is labeled as such — never presented as live. We keep it segregated from the CC-BY-4.0 play-by-play.
• ADP — via Fantasy Football Calculator (free API, 12-team, all three scoring formats). Where the market drafts each player.
• Current team, depth chart & QB1 — via nflverse depth charts (CC-BY-4.0). The most-recent depth-chart snapshot that fills each board player’s 2026 team, role, and suggested starter for the situation overlay (hand-curated facts always win).
• Next Gen Stats — via nflverse (CC-BY-4.0). Player-tracking aggregates behind the “why” — target depth (aDOT), separation, air-yards share, YAC over expected, rush yards over expected, stacked boxes, completion % over expected — shown as prior-season context beside the split.
• Play-action charting — FTN Data via nflverse (CC-BY-SA-4.0, credited to FTN Data). We surface a derived team play-action rate as context; we never redistribute the charting rows.
• Market value — via FantasyCalc (free API). Redraft trade value, positional rank, and 30-day trend distilled from thousands of real league trades — a market reference shown beside the split, never folded into the computed TrueValue.

We use team and player names as public factual references only — no NFL/team logos, uniforms, or player photos. Analytics only; no wagering.

A split is just a per-game average recomputed under an honest condition, straight from the game logs. You can toggle any of these in the Explorer:

• With / without a given QB — keyed on the QB actually on the field, never silently blended across QBs.
• Game script — as favorite vs as underdog (from the pre-game spread), and in wins.
• Venue — home vs on the road.
• Full games only — drops games below a snap-share threshold (early exits / token appearances) that quietly drag the raw average down.
• Recency — last 5 games, or “since week N” when a real in-season change is detected.
• Playoff weeks — weeks 15+.

Every split shows its game count, and a split under 4 games is flagged small sample so a three-game number reads honestly. The week-by-week chart stays visible so you can judge confounds yourself — a with/without-QB gap is sometimes really a role or injury change.

A gap between two splits isn’t automatically real. TrueSplits compares the gap in a player’s averages to its margin of error — the standard error of the difference, √(s₁²/n₁ + s₂²/n₂) — and to his normal week-to-week swing, then labels it: signal when the gap is at least twice that margin, likely noisewhen it’s within one, uncertain in between, or small sample when a side has fewer than 4 games.

This is a transparent rule of thumb, not a formal significance test— we don’t claim a p-value, and when a split is compared to the season average the two samples overlap, so that read is deliberately informal. The point is honesty about confidence, not false precision: every badge sits next to the game count and reflects the gap-versus-margin, so you can see exactly why a difference is called real or not. On the draft board, a confident Target/Fade is only made where the gap reads as signal.

The board prices each player off his go-forward baseline — his value under normal conditions, not a fluky tail. The rules:

• Bona-fide-starter baseline. We price off his most-recent season in which his primary QB was a genuine full-season starter — playing at least 65% of his games and at least 8games. A short backup / injury-fill-in stretch is shown as a caveat, never as the headline number (e.g. a stud isn’t priced off the five games after his QB got hurt).
• Same-QB short samples are combined.If the established starter simply missed a few games, we combine both seasons’ with-that-QB games into one fuller, current sample rather than softening the call.
• Volatility gate. When the baseline is situation-dependent — a discounted backup stretch, or representative seasons spread by 12+ PPG (cross-season dispersion) — we show both levels and decline a confident call.
• Durable role changes count. A role step-up that persisted (e.g. a committee back becoming the lead back) is his new normal and is priced as such.

The 2026 situation overlay only picks or labels which real split is the baseline and shows context beside it — it never changes a computed number. A confirmed 2026 QB who matches the baseline makes the call informed; a confirmed new QB with no games together keeps the established number as a labeled projection (verdict stays Fair — we never make a confident call off a pairing with zero games).

• Has splits — enough real games to price and make a call.
• New situation — his baseline comes from a different team/QB than his 2026 spot; softened to Fair and labeled, because his splits are from the old situation.
• Rookie / no history — no NFL games yet, so no fabricated number — an honest empty state, never a made-up TrueValue.