Methodology

Score and tier answer different questions

Score is a versioned index for the source snapshot shown in the footer. A score of 50 is the average hero in that rarity for that snapshot, and 15 points is one snapshot standard deviation.

Tier is a live standing among the heroes available today. A hero can keep an 84 score while falling to A- if newer heroes push it outside the top percentile bands.

Each letter splits into three steps, and every step is a fixed share of the heroes in that rarity band. Sub-steps are only meaningful where the population is large enough to fill them: at 5★ each step holds dozens of heroes, while the handful of 1★ heroes leave some steps empty entirely.

Tier percentile ranges

A tier is a live percentile standing within the same rarity and selected mode. It is not a score range.

• S+top 2.7%
• S2.7–5.3%
• S-5.3–8.0%
• A+8.0–17.0%
• A17.0–26.0%
• A-26.0–35.0%
• B+35.0–43.3%
• B43.3–51.7%
• B-51.7–60.0%
• C+60.0–68.3%
• C68.3–76.7%
• C-76.7–85.0%
• D+85.0–90.0%
• D90.0–95.0%
• D-95.0–100.0%

Heroes sharing the same displayed score stay together, so a tie crossing a boundary moves into the lower tier.

What the model reads

The pipeline imports Heroplan hero data, parses mechanical clauses into a typed effect tree, prices those typed effects, and blends special value with attack, defense, and health.

• Stats use LB2 and one balanced 25-talent path selected from all 512 legal maximum paths.
• A Classic hero's base form uses no costume bonus. Its costume forms use the same all-available-costumes-maxed stat context and 5% mana-generation bonus. Other families retain the stat context supplied by Heroplan.
• The same talent build is kept in offense, defense, and titan modes.
• Offense, defense, and titan modes are scored separately.
• Overall is a weighted blend of those modes.
• Rank is derived from score inside each rarity band.
• Tier is derived from that live rank, not from a fixed score threshold.

Why disagreement is expected

Community rankings measure player judgement across real teams, habits, and metagame context. This model measures the parsed mechanics it can price. Where those disagree, the disagreement is a useful signal rather than something to hide.

Descriptions are generated from typed mechanics

Hero descriptions are written from the parsed effect tree in this project's own words. Untyped clauses are omitted instead of copying the game's printed skill or passive text, so a short description usually means the model does not fully understand that hero yet.

Refresh and scale version

The footer shows the source data date and parser version used for the current build. The launch refresh path only publishes a new Heroplan pin when every variant parses and no new unvalued clauses are introduced.

Parser and valuation improvements can move scores because the model has learned to price more mechanics. A safe source refresh also updates the roster reference, so a newly added hero can move existing relative scores.