Ratings methodology · beta

Strength, with the reasoning attached.

Every combat card and ability can receive a 0–10 rating. Practical team contribution provides a starting point; ranked win results and observed card use gradually refine it.

What a rating means

Zero represents very poor combat value; ten represents an exceptional, near-perfect result in the evaluated context. A score is not a win probability or a promise that a card fits every team. Units and summoners use different stat treatment. Land cards have no combat score.

Ratings are separate for Modern and Wild, at Bronze, Silver, Gold and Diamond/Champion card-level caps. Wild battle context includes Modern-eligible cards too. This differs from the catalogue’s Wild-exclusive filter, which only shows cards outside Modern.

Practical contribution to a team

Each card uses only the stats and abilities unlocked at the selected league cap. Those caps are retained, including restored archon caps. The model estimates what a card can contribute in a useful role: attacking, absorbing damage, supporting allies or enabling a team. Editorial tiers remain an independent reference and are never used as an automatic scoring input.

Units are evaluated in frontline, backline and, when Reach is available, second-position roles. Attacks depend on position and accuracy. Survival limits useful attack time; Double Strike and Armored Strike contribute through attack output. A mixed melee/ranged card cannot automatically use both attacks from every position. A small allowance for backline melee represents later movement to the front.

Durability considers physical and magic pressure, armor absorbing a whole hit, Shield and Void damage breakpoints, speed, evasion and Forcefield. Heal scales with maximum health; Repair can sustain armor and armor-based attacks. Damage and survival estimates are discounted by curated ability priors and bounded to avoid idealized infinite survival. These are practical estimates, not a full battle simulator.

Mana, matchups and team roles

The model averages ten situations: low mana (15%), mixed opposition (25%), high mana (15%), armored opposition (10%), magic opposition (10%), and Reverse Speed, Noxious Fumes, Melee Mayhem, Close Range and Back to Basics (5% each). Low, mixed and high budgets use 22, 32 and 50 mana. These weights are explicit design assumptions, not measured ruleset frequencies. Matchup estimates also use the selected format’s cap-level catalogue to represent typical speed and attack pressure.

Reverse Speed changes speed’s direction; Noxious Fumes adds poison pressure; Melee Mayhem and Close Range enable attacks from additional positions. Unit abilities are removed in the Back to Basics component. Card pages show each situation’s effect on the card’s own contribution. Multiple simultaneous modifiers and exact active-element combinations are not simulated.

Taunt attracts more incoming pressure, so high health alone does not guarantee sustained attacks. Large physical hits lose overflow against armor without Piercing. Deflect receives more credit for attacking or covering second position than for sheltered support. Retaliate needs melee attackers, and Enrage offers less without melee offense. Fast, fragile attackers still receive an allowance for acting before dying.

Support and affordable partners

Healing, armor repair, team buffs and debuffs scale with relevant recipients and available team space. Stacked support has diminishing returns: several abilities cannot each assume ideal targets. Headwinds counters ranged pressure; Demoralize counters melee pressure. A heal targets one injured ally rather than multiplying with team size. Low-mana cards receive credit for leaving room for allies. Weapons Training needs an eligible, affordable trainer or trainee, with an allowance for adjacency and setup cost, combined mana investment and recipient initiative. Available partners must share a legal archon and element choice; Dragon is not permission to combine arbitrary elements. Gladiator partners require Conscript in ordinary ranked packages.

Bloodlust scales with usable damage and survival. Painforge considers poison and compatible friendly-damage sources, including intermittent resurrection damage; friendly damage still carries a cost. Weary reduces attack and active support availability unless Invigorate is available. Shapeshift receives bounded credit for available bloodline grants. Partner assumptions describe potential teams, not guaranteed buffs in every battle.

Tribute evaluates up to three beneficiaries in a legal 32-mana package, shared attack types and discounted repeat deaths. Its interaction value has diminishing returns and is discounted for narrow bloodlines and dependence on Shapeshift partners. Resurrect support must fit alongside the beneficiaries before an extra repeat is credited. Dispel can remove Tribute buffs; current Final Rest can block an enemy resurrection, rather than suppressing every death-triggered effect. Death Has No Power and other special modifiers may have different behavior.

Archons and calibration

Archons receive credit for team access, including dual-element and Dragon flexibility. Team stat effects account for relevant recipients and have diminishing returns. Tactics use the strongest alternative, recipient count and effect size; alternatives are not added together. Bloodline grants and partnerships remain conditional.

Published ability priors remain editable. Their Standard/non-standard blend is currently 20% /80%. Quantitatively modeled effects replace their former flat bonuses. Other abilities retain prior-based contributions. Overlapping pair and stat bonuses are suppressed for modeled effects. Battle-adjusted ability ratings are not fed back into card ratings.

A fixed smooth curve maps model points to 0–10, with a unit center of four and an archon center of 2.5 by default. Explained curator adjustments then apply directly, limited to −2 to +2 per card and scope. Adding cards can change partner access and matchup estimates, but does not force ratings into a percentile distribution.

How battle evidence changes a score

Completed, non-surrender ranked matches with usable lineups and pre-battle player ratings contribute daily statistics. Shared or manually selected replays are excluded. Level caps come from explicit match settings or conservative lineup inference. Underlevelled cards do not supply cap-level strength evidence. Features appearing on both teams are excluded from strength evidence.

A draw counts as half a win. Results are compared with an Elo-style expectation based on pre-battle player ratings. We retain today and the previous 89 UTC days, applying a 30-day half-life by day. Detailed replays and player identities are not retained as scoring evidence. We do not measure distinct players or apply per-player volume caps; repeated sampling of the same accounts can bias results.

Cards need 20 appearances, 15 effective samples and two sampled days. Abilities need 30 appearances, 20 effective samples and two days. Five percentage points above or below expected results imply one potential rating point. Card influence is limited to 0.8 × effective / (effective + 75) × min(1, days / 3). Ability influence uses 0.65, a denominator offset of 150 and time breadth up to seven days. No usable evidence leaves the practical baseline unchanged.

Use rate as a utility signal

Use rate is the fraction of qualifying matches in the same format and cap containing a card. Mirrors count once. Underlevelled appearances count for selection frequency. Both numerator and denominator use the same daily time decay. A bonus needs 20 card matches, 15 effective card matches, 50 effective scope matches and two sampled days.

The bonus is at most one point before confidence weighting, reaching its ceiling at 20% weighted match share. Its weight is totalEffective / (totalEffective + 100) × cardEffective / (cardEffective + 30) × min(1, days / 3). Unseen cards receive no popularity penalty. Final scores combine the practical baseline, win-result adjustment and usage bonus, bounded to 0–10.

Card uses and win rates

Lifetime counters preserve collected ranked uses, wins, losses and draws per card, format and cap. Each team contributes a use, including underlevelled cards. Mirrors contribute two uses. Win rate is wins / uses, with draws in the denominator. These totals describe our sample, not all game battles, and survive expiry of daily rating evidence.

Confidence and limitations

“Curated only” means neither evidence threshold is met. Usage alone can establish “Early” confidence. “Moderate” requires 150 effective strength samples and seven days; “Strong” requires 500 and 14 days. These labels measure sample size and time coverage, not causal certainty or player diversity. Bots, ownership, teammates and collection bias can affect both win results and popularity.

Summaries allow changes to weights, decay and confidence thresholds. Changes requiring discarded replay details apply to newly collected matches. Historical replays cannot be reprocessed.

Collection status

9,336 summarized ranked battles · 9,336 ranked battles · 9,336 eligible for strength · Latest battle 2026-11-08.

Latest run: 2026-11-10 · 0 new battles summarized. Some upstream requests were unavailable; completed ingestions are retained for the next run.

A job scheduled every hour at minute seven discovers players from public leaderboards and opponents, fetches authenticated ranked history when credentials are available, and samples public queues and replays. Account/format requests prioritize caps with fewer qualifying battles in the last 90 days, using each account’s recently observed cap. Every fifth planned request explores the longest-unchecked account/format so existing coverage does not prevent discovering new scopes. API leaderboard labels do not prove cap coverage, and missing scopes cannot be guaranteed each run. Ratings refresh daily; collection runs update daily summaries between refreshes. Scheduled runs may be delayed by the hosting provider. Shared or manually submitted replays are ignored, because players selectively share their results. Bulk history access depends on the upstream API’s current access rules. Publicly discoverable matches are a sample, not a complete or random game-wide archive.

Publishing and accountability

Authorized curators edit a draft, preview its results, then publish a named version with a reason. Draft changes do not alter public scores. Publishing atomically replaces the score snapshot, records the profile and author, and refreshes cached ratings. Battle collection refreshes the published profile’s evidence without publishing draft weights.

Current engine: practical-3.1. Profile: 542bcb8c-10bd-439b-ae5e-ec5d67aa6115. Every card’s breakdown exposes its components and current evidence.

Sources