Methodology
“Accurate” is an evaluation target, not a marketing promise. Values are uncertain, contextual, and manager-dependent.
Layer 01
Current dynasty market rankings come from Dynasty Dealer and map directly to Sleeper player IDs. Source, retrieval time, and stale states stay visible; market value is not presented as an Atlas projection.
Layer 02
Projected scoring and weekly outcomes require a separate, permitted projection source. Dynasty Atlas never turns market rank into a fake football projection.
Layer 03
Format scarcity, TE premium, starter impact, usable depth, replacement level, roster-space cost, and risk adjust the decision — not the player's universal identity.
Layer 04
Target season and explicit constraints change recommendation priority. Factual roster classification and universal market values do not change when an objective changes.
Why four quarters do not always equal a dollar
Raw value is strictly additive. Effective value preserves the best asset, applies diminishing weight to assets above replacement, adds a visible elite premium, and caps roster-space penalties so adding positive value never creates an unexplained negative marginal contribution.
Accuracy cannot agree with itself
Chronological train/validation/test splits and historical snapshots prevent hindsight leakage.
Track market rank correlation and held-out completed-trade error by format and tier.
Measure interval calibration, responsiveness, stability, and missing-data confidence.
First-party Dynasty Atlas data
The community board is built only from head-to-head comparisons cast by signed-in Dynasty Atlas members inside Dynasty Atlas. No outside ranking site is scraped, imported, or republished, and no vote is ever fabricated to make the board look populated.
“Which asset would you rather roster in this dynasty format?” — shown above every comparison.
Online Elo-style pairwise updates, replayed deterministically from the stored vote log. Version atlas-community-elo-1.
Superflex and 1QB votes are stored and ranked separately and are never pooled.
Top-100/250/300 controls limit the relevant asset pool, and the matchup-margin control keeps choices near one another on that board. Permitted market observations seed matchmaking only; they never enter the community score.
An asset joins the public board at 3 comparisons. Below that its rating is still the untouched neutral prior, so publishing it would present noise as a ranking. This is a usefulness threshold, not a privacy one.
An asset is “Established” only at 25 votes across at least 5 distinct accounts. Between 3 and that point it is published but labelled Provisional.
Every rating is pulled toward a neutral prior in proportion to how few votes it has, so a three-vote asset cannot outrank a three-thousand-vote asset on noise.
Voting requires an account. One account’s influence on a single asset is capped, submissions are idempotent, and per-minute and per-day limits are enforced in the database. No IP address is stored.
Seven- and thirty-day movement comes only from stored snapshots. Until snapshots exist the column stays blank — history is never back-filled.
Raw votes are private to the account that cast them and protected by forced row-level security. Published aggregates contain no account id, email, league, or voter identity, and head-to-head records for a specific pair are withheld until at least 5 distinct accounts have voted on it. That cohort is a privacy threshold and is deliberately separate from the 3-comparison threshold for appearing on the board; it is never lowered to make the board look busier.
A community score is a preference signal, not a statistical forecast and not a market price. It is never blended with market values.