2 · AI Guru and League Value

The composite (see the overview) answers "what does the blend of every source think?". It cannot answer "in this league, with these rules and this lineup, who is mispriced?" — because every consensus board assumes standard scoring and a standard lineup. Three pieces of the advisor exist to close that gap: the league-value board, the AI Guru, and the Pure FPPG board. All three are computed from real per-game production, and all three pass through the same injury gate.

Previous: 1 · Narratives and point adjustments

League value: points over replacement, under your rules

Every player's per-game production over the last three seasons is stored as a bonus-aware profile — not just yards and touchdowns but the things exotic leagues pay for: first downs, 40-yard plays, 100-yard games, completion tiers, distance-tiered receptions, TE reception bonuses. Scoring that profile under your league's exact Sleeper settings gives a projected points per game under your rules.

Points per game alone over-values quarterbacks in every format, so the board subtracts a replacement level: the points per game of the last player at each position who would start in your league. That count comes from your real roster shape and team count — a 12-team superflex league starts far more QBs than an 8-team 1QB league, and FLEX and SUPER_FLEX slots are apportioned across positions by how they are actually filled (superflex slots go mostly to QBs; flex slots split between RB and WR with a sliver for TE).

Value over replacement is then discounted for sample size: a per-game average from 5 games is not the same evidence as one from 55. Only players who beat replacement are discounted (a five-game sample keeps about 38% of its edge, a full season about 68%, a long-tenured veteran nearly 90%). Players below replacement are left where they are — riding the bench is itself evidence, and shrinking a negative number toward zero would promote exactly the players with no case.

Three gates keep the board honest:

• retired or inactive players are dropped;
• players with a season-ending designation are dropped (the injury gate, below);
• players under 0.1 points per game are dropped as residue, not production — a kneel-down here and a garbage-time carry there once put a third-string QB at rank 113.

Unsigned free agents stay on the board but sort below every signed player.

How much the league-value board is allowed to say

The board's share of the composite is not fixed. It scales with scoring deviation — a measure of how far your rules move the player-value board away from standard PPR. Deviation combines two things: how much individual ordering churns at the top of the board, and how many of the top-60 value seats change position between the two boards. The second term matters because a league where 22 of the top 30 values are running backs can still correlate almost perfectly with standard scoring when each position's internal order barely moves.

A vanilla league sits at the floor (25% of the composite goes to league value); the most warped themed leagues reach the 60% ceiling. This was rebuilt after four "Grand Slam" themed leagues exposed the failure mode: the RB-themed league once received a recommendation list with zero running backs, and the WR-themed league twelve straight quarterbacks, because the consensus superflex sources out-voted the theme.

Two further guards from the same fix:

• Displacement guard. Percentile blending compresses the top of a 550-player value board, so even a well-weighted league-value board can be out-shouted. In any league with meaningful deviation, each candidate's score is scaled by his standing on the available league-value board — the best available value gets a nudge, a player ranked 120th or worse takes a hard haircut. Worst top-8 displacement across the four themed drafts fell from 108 spots to 37.
• Need saturation. The superflex QB scarcity bonus scales fully with your remaining QB need. Once the QB room is stocked the bonus is zero, so it can break ties but can never build a wall of quarterbacks on its own.

In a chopped (weekly-elimination) league the league-value board is switched off entirely, because the chopped survival board already contains this league's scoring projection and letting both vote would count it twice.

The AI Guru: the composite, nudged by two signals

The AI Guru is a derived board. It starts from each player's composite rank for your league and moves him up or down by a bounded number of spots based on two signals, then re-ranks the pool. It is never loaded from a file and it never feeds back into the composite.

Signal 1 — scoring fit and positional premium

This is where the league's exact scoring lives. Both terms convert points to rank spots using the candidate pool's own points-per-rank density, so a cosmetic 0.05 bonus cannot move a player as far as a 1.5-point premium can.

• Scoring fit (within position, ±10 spots). How much more your ruleset pays this player's real production profile than it pays the typical player at his position. It measures how many extra points per game your rules give him beyond what they give the average player at his position, converted to board spots by how densely points are packed at that depth, so a vanilla league moves nobody while a 40-yard-bonus league lifts big-play receivers and a completion-bonus league lifts volume passers.
• Positional premium (across positions, ±8 spots). How much your ruleset grows a position's starter-versus-replacement spread relative to other positions. Replacement-adjusted on purpose: a bonus every quarterback earns lifts replacement quarterbacks too and cancels out; only genuine positional scarcity moves the board. This is what puts an elite tight end ahead of quarterbacks in a 2.5-PPR TE-premium league with nerfed passing touchdowns — verified live, where a themed TE league's Guru board ranked Trey McBride #8, one spot ahead of Josh Allen.

The two scoring terms share a ±16 budget. Under standard scoring both are exactly zero, so a vanilla league sees the unmodified composite order.

Signal 2 — role security

Rankings are slow to reflect who actually has the job. This signal combines the Sleeper depth chart with a proven-production test (roughly 24 weighted games at starter-level output for the position) and moves a player accordingly:

Redraft magnitudes are doubled because current-year role certainty matters far more when there is no dynasty upside to wait on. The role term is capped at ±12 spots.

One owner per signal. The role-security term is skipped for any player who already carries an active role, injury or legal adjustment from the news layer this cycle (see doc 1). A demoted running back has already been moved -5 on the Squatch board, and that move already reached his composite rank; penalising him again here for being a backup would count the same fact twice.

What the Guru no longer does

Earlier versions added a third "AI sentiment" term: the raw adjustment delta from the news layer, applied a second time on top of a composite that already contained it. That term has been removed. The news layer reaches the Guru through the composite's adjusted Squatch boards and nowhere else. Callers may still pass the old sentiment fields for compatibility; they are ignored.

Budgets

Scoring terms share ±16, the role term has ±12, and no player moves more than ±20 spots in total from his composite position. Movement versus the composite order is reported per player, which is what the up/down badge and its tooltip show.

Pure FPPG

The simplest board: raw projected points per game under your scoring, with no replacement level subtracted and no blending. It answers "who scores the most points in this league?" — useful for best-ball and for sanity-checking the other two boards. Unsigned free agents sort below every signed player so a released star does not top the board on stale production.

Rookies and other players with no NFL production do not have a profile and so cannot appear on the league-value or Pure FPPG boards directly. Rather than leaving them rankless, they are spliced in at their league-tuned Guru position and flagged as estimates.

The injury gate

Every board the advisor computes — league value, Pure FPPG, AI Guru, and the pick recommendations themselves — drops any player whose Sleeper injury status is a season-ending designation (IR, Out). Last season's production is worth zero points this season, and a rank built from it would be a trap. This is a separate mechanism from the news layer's injury adjustments: those shape how far a playable player moves; the gate removes players who cannot play at all. The gate reads Sleeper's status directly and degrades to "no gate" rather than failing if Sleeper is unreachable.

Where each board shows up

For the prediction engine on the game side of the site, see How Our Model Works.