Skip to content
bladeformBladeform Legacy
Lineup overviewDraft simulator
Sign inSign up
‹ My account

Usage and data guide

2026-11-19
Reading a result

Bladeform.ai uses public Dota 2 matches for draft research and review. An estimated win probability is conditional on the available data. A patch, substitution or lane swap can change the outlook.

Team Elo, player mastery, lane income and match probability measure different things. Missing observations, failed queries and pending updates have separate states. An empty history says nothing about a player’s ability to play a hero.

Sources, patches and dates

The professional snapshot at 18:14 Beijing time on 10 November 2026 covers 3,644 games from 7.41a–f for the combination rankings and draft model. Three designated exhibition categories are excluded. Subpatches are inferred from release dates. The lane snapshot uses 3,624 parsed pro matches with identifiable roles and early observations on the lane in question.

The hero matchup and synergy snapshot covers 100,000 public ranked games on 7.41f through 4 October. At least five ranks are known and the mean rank tier is at least 75, which includes Divine V and Immortal. The supplemental lane sample through 10 October has 1,915 parsed high-rank games, including 1,658 on the current subpatch. The 160 Divine I–IV and 201 Ancient games are separate reference groups.

Team and player queries read server-side copies of public OpenDota records. Regular rosters refresh roughly every 12 hours, hero pools and account histories every 24 hours. Rate limits, privacy settings and upstream gaps may delay updates. Refresh reads the available records. Hero snapshots update separately and do not track live matches.

Entering a draft and lineup

Paste the event and game number, then a team name, side and five heroes for each side. Alternatively list one player and hero per line below the team name. Use 待确认 for an unannounced pick. Radiant maps to A and Dire to B. Pick order is separate.

Use ten distinct heroes for a complete analysis. Link both teams to view player ratings earlier. The usual starters are prefilled and can be changed for substitutions or role swaps. Roles are inferred from pro and player records, independently of input order. A manual lock takes precedence.

Team Elo, pro pools and account history

Elo represents relative strength from professional opponents and results in the snapshot. The estimate identifies whether actual teams contributed. Unidentified teams use the available model baseline.

The pro pool includes games for previous teams within the selected window of 90, 180 or 365 days and either all patches or the 7.41 series. Wins and raw win rate are shown. Pool fit also weighs sample size, role frequency and activity over the last 30 days.

Public account history spans patches and non-professional play. Games with unknown roles can count towards account totals. Lists can be sorted by games, score or win rate. Heroes with few recorded games can top a win-rate sort.

Hero mastery and lineup scores

Mastery uses account games, win rate and personal pick rank. S starts at 85, A at 70, B at 55 and C at 40. Lower scores are D. Scores can exceed 100. Fewer than 10 games are marked as a small sample.

The first 200 games earn up to 45 points linearly, then one point per 100 games. Win rates from 30% to 70% map linearly to 0–45 points, with one point per percentage point outside that interval. Personal pick ranks within the top 3, 5, 8 or 12 add 10, 7, 5 or 2. At 6 or more games with a win rate of 33% or lower, the score is halved and graded after rounding to one decimal.

Positions 1–5 carry weights of 30%, 25%, 20%, 15% and 10%. S grades add 5, 8, 6, 4 and 4 points. Observed D grades subtract 8, 6, 5, 4 and 4. Incomplete data yields a partial score. The score difference has no direct conversion to win probability.

The performance term weights GPM and smoothed KDA equally, compares the same hero and role, and adjusts for the player’s other heroes. It uses 180 days of 7.41 pro data, a 45-day half-life and minimums of 3 player and 20 reference games. It carries zero weight after a 443-game replay found no overall improvement in probability error.

Lanes, income and confidence

Standard lanes are 1+5 versus 3+4, 2 versus 2, and 3+4 versus 1+5. Exact pro role matchups take priority, with parsed high-rank or series data as fallback. Sources are counted separately. The series includes the current subpatch. Core-only observations carry less weight. Conflicting sources are flagged.

Early income difference uses 35% of the cumulative 5-minute difference and 65% of the 8-minute difference from Radiant or the labelled A perspective. The 10-minute figure is separate. Income differs from net worth. A 95% reference interval describes uncertainty in the historical mean.

The early index weights economy at 55%, 8-minute experience at 30% and last hits at 15%. A directional signal needs 12 games and an interval that excludes zero. Strength bands use 0.2, 0.5 and 1 source-specific standard deviation.

With sparse exact four-hero observations, side lanes can use four opposing pairs. The four-pair estimate has low confidence and carries no BP score weight. Pair samples may overlap and cannot be added as exact full-lineup games. Inferred roles, core-only data, small samples and disagreement affect confidence.

Matchups, historical signals and match estimates

Matchup and synergy data describe associations with outcomes. High-rank estimates fit base hero strengths and shrink observed differences towards expectation. Players, roles, items and repeated participants influence results. Association alone does not establish a causal counter or lane outcome.

Top 200 and Top 500 are cutoffs of one ranking. The six families are 1+1, 1+2, 2+1, 2+2, 2+0 and 3+0, with at least six games. Sort by raw p, sample size and structure ID, keeping the lower-p mirror. Opposing groups use a two-sided Fisher test. Ally-only groups use a two-sided binomial test against members’ mean win rates in the same sample. Neither p nor adjusted q represents win probability. The current minimum q is about 0.747.

A historical signal describes direction relative to its own baseline. Patterns can cover the same games. No hit means no coverage. The match estimate distinguishes team strength, patch and side effects, heroes and combination adjustments according to available inputs. Contributions are expressed from side A.

The snapshot has 3,379 pro matches with both team IDs for model updates. In 819 matches after the old cutoff, rolling Elo called 487 correctly and M2 with hero information called 476. This check describes an older model on later games and does not establish the accuracy of the refit or any individual prediction.

BP suggestions

Captain’s Mode runs in three rounds of 7 bans and 2 picks, 3 bans and 6 picks, then 4 bans and 2 picks. Suggestions use revealed picks and bans, hero strength, counters, synergy, roles, pools and likely replies. Candidate scores rank the options at the current step. The top twelve candidates are evaluated through the next two actions. A ban subtracts 35% of our pick opportunity cost.

Recent pro usage is the main player reference. S and A heroes with at least 10 account games may receive a bonus shrunk by games / (games + 60), capped at 25% of personal fit and scaled by the pool weight. A missing record does not restrict legal picks.

Browser storage and conditional calculations

The current browser stores drafts and BP progress. Changing devices, clearing site data or using a private window can make these records unavailable.

The web conditional calculator applies to a single game with no draw. Signal mode shrinks one lowest-p combination with a neutral Beta prior. Manual mode uses the entered probability. The conditional 95% interval is unadjusted for selection or series dependence. Full Kelly is (decimal odds × probability − 1) / (decimal odds − 1). The calculator uses one quarter with an optional cap. It executes no transactions and supplies no betting advice.

Limitations, rights and corrections

Last-minute substitutions and lane changes are possible. Public records can be delayed, missing or revised. Read estimates alongside the actual lineup, patch and sample count. Models cannot guarantee an outcome.

Bladeform.ai is independent of Valve, Dota 2 and OpenDota. Game names, hero artwork and marks belong to their respective owners. For data corrections or rights questions, use the support contact and identify the team, player, hero and match time.

OpenDota ↗ · OpenDota API ↗

bladeformBladeform Legacy X @bladeform_ai ↗ Support the creator 隐私政策帮助与反馈 Guides and case studies Lineup text importAbout Bladeform Match case study
中文EnglishРусский