What is
behind the
number

Objective

The draft intelligence engine recomputes a roster-building comparison after every recorded pick: the highest-scoring candidate under its assumptions, two fallback plans, and the modeled cost of waiting. The scores compare whole-roster fit across a draft; they do not establish that a candidate produces the best real outcome.

Where it runs

One engine serves every draft surface. Sleeper, ESPN, and Fleaflicker public drafts normalize into the same board state, and manual recovery produces the same decisions when a live feed is unavailable. The war room, draft assistant, and manual assistant all render output from this model version.

• Live public drafts on Sleeper, ESPN, and Fleaflicker.
• Manual and pasted draft entry for league hosts DraftADynasty does not read.
• Arrival-phase guidance while you wait and on-clock guidance for the pick in front of you.

Inputs

Every input is visible with the recommendation that used it. The engine starts from the current board: which picks are recorded, which players remain, your roster, and your league's exact scoring and lineup rules.

• Current market value for each remaining player, source-labeled from the live provider consensus.
• Roster fit: whether a candidate fills an open starter slot in your league's lineup shape.
• Availability urgency: one minus the modeled probability the player survives to your next turn.
• Positional scarcity: remaining supply in the player's position tier against league demand.
• Modeled opponent demand: what the intervening seats are forecast to take, from aggregate market behavior - never from any individual manager's private information.
• Recent position runs and tier supply are simulation context, rather than a sixth independently weighted score term.

How a score is built

The engine runs a seeded simulation of the picks between now and your next turn. The following-turn score weights current value 36%, roster fit 23%, availability urgency 18%, positional scarcity 13%, and modeled opponent demand 10%. Other decision phases omit urgency from base utility and renormalize the remaining weights by 0.82, then apply phase-specific availability and roster penalties. These weights describe mechanics, not validated predictive strength.

Deterministic replay requires identical player-value rows, league format, board state, model version, seed and simulation configuration. A board revision alone does not pin a changing value feed. The browser ledger records audit context and hashes; exact replay also requires retaining the original input snapshots, which the ledger does not supply.

While you are waiting, the model discounts candidates unlikely to survive and promotes realistic next-turn targets. While you are on the clock, it compares taking a player now with the modeled cost of waiting one more pick.

• Plan A/B/C ordering with a transparent 0-100 score for each candidate.
• Survival probability to the forecast target pick for every candidate.
• Per-pick position forecasts for each intervening seat with a confidence grade.
• Position run forecasts: expected picks, probability of one, probability of two or more, and early/steady/late timing.

What you see

Every recommendation card carries its model version, board revision, seed, simulation count, and confidence grade. Plans name a top-scoring option, a fallback, and the modeled drop if you wait. The input list explains the score, and a drafted player is never recommended again.

Assumptions

The opponent model treats the rest of the room as aggregate market behavior: it assumes seats draft near current market order, leaning toward need where the board pressures them. It assumes recorded values are fresh, your league settings are correct, and the recorded board is complete. When a value feed is stale or unavailable, the affected surfaces fail visibly rather than substitute a fabricated number.

Limits

Trades, sleepers, and reaches happen. The opponent forecast describes what the market is likely to do, not what any specific manager will do. Values move with injury and news faster than any cache refresh. A high score is a roster-building argument, not an outcome promise: injuries, coaching, and luck dominate single-season results.

• Survival probabilities degrade quickly with long horizons and small boards.
• K and DST coverage lacks clock alignment in some manual entry flows and is marked when affected.
• The model does not see another manager's queue, chat, or intentions.

What this model does not claim

• No public accuracy claim: this draft model has no graded draft-specific holdout evaluation for its current version. Weekly ranking-board grades on the track-record surface do not validate draft decisions, survival probabilities or roster outcomes.
• No mind-reading: opponent forecasts are aggregate market simulations, never knowledge of another manager's intent.
• No outcome guarantee: a recommendation is a second opinion, and the pick stays yours.
• No fabricated fallback: when inputs are unavailable the engine degrades visibly instead of inventing values.

How it is checked

• Deterministic output, opponent-roster sensitivity, and duplicate-free full mock behavior are locked by the draft intelligence contract suite on every change.
• Live recommendations are recorded in a browser decision ledger with model version, input context and state hashes for auditing. The ledger does not retain every original value snapshot and therefore does not by itself enable exact historical replay.
• Graded draft-specific accuracy evaluation is pending. The separate track-record policy snapshots and grades weekly ranking boards; those results do not transfer to this draft model.

Changes

• 2026-09-26 - draft-model-card/v1: clarify phase-specific score composition, input requirements for replay and the draft-specific evaluation boundary for draft-intelligence/v6-phase-aware-utility.