DOTA 2 · Lineups & drafts
Data, scores and uncertainty
How Bladeform uses professional matches, high-rank games and OpenDota records, and how lineup scores, probability estimates, sparse samples and update scopes differ.
How can you verify a Bladeform assessment?
Check match facts, model output and author interpretation separately. Match sources establish what was drafted; a retained run records what a model computed for specified inputs. Neither alone establishes prospective accuracy or a trading edge. Public cases expose inputs, sample context, assumptions and results so those claims can be checked.
Evidence needed for a citable assessment
| Evidence layer | What to retain | What it establishes |
|---|---|---|
| Match facts | Sources, match ID, map, sides, draft sequence and actual roles. | Identifies the map. Eventual roles may not have been known before play. |
| Run conditions | Analysis time, data snapshot, patch coverage, sample counts, role inputs and hero-pool restrictions. | Defines the scope. A frozen run is not necessarily the current production model. |
| Model output | Component evidence, composite score, candidate ranks and opposing replies. | Records the comparison under that configuration. A rank is not a selection probability. |
| Predictive validation | Consecutive predictions recorded before play, all outcomes, calibration and time-separated evaluation. | Evaluates forecasting ability. One retrospective example cannot substitute for it. |
What does 51.10 mean in the Yandex–Spirit case?
It is Spirit’s composite score in that retained offline run, not a 51.10% win probability. The case records five Yandex final-pick candidates and Spirit replies; the actual OD pick ranks 43 in its corresponding candidate list. Scores and ranks depend on the recorded roles, data and candidate-pool assumptions.
What can this case establish?
It demonstrates alternative-pick comparisons with fixed inputs and retains all 24 draft actions, roles, data context and outputs. It was run after the match using eventual roles. The case does not include a full replay fight review or historical market prices, so it cannot explain every cause of the loss or establish a pre-match trading edge.
What happens when the patch or data changes?
Keep the original case time, inputs, snapshot and outputs. Record new runs with their dates and changed conditions rather than silently replacing old evidence. Recompare when samples, hero pools or roles change. A documentation update date does not mean every model dataset was refreshed together.
Inspect the public case and JSON evidence: Yandex–Spirit map 2
Where does the data come from?
Analysis draws on professional and high-rank match history. Team and player information uses public records such as OpenDota. Scope and update times vary by metric; consult the corresponding snapshot and sample notes inside the tools.
What does lineup analysis compare?
Lineup overview compares roles, three lanes, counters and synergy. Actual lanes, player hero pools and team records affect the analysis context. Check automatically assigned roles against the target match.
How should draft recommendations be used?
Draft simulation compares candidates against current picks and bans using available role, lane, lineup and player-pool information. Recommendations help compare options and do not replace in-game judgement.
How do scores differ from probabilities?
Overall scores and player mastery scores describe or rank options; they are not win rates. Separately shown win probabilities are model estimates. Historical replay performance is not the accuracy of future predictions.
What does limited evidence mean?
Missing data does not mean a hero is weak, a lane is easy or a team has disbanded. Small samples and old-patch records have limited relevance. Professional and ranked samples are not the same population.
Who maintains Bladeform?
Bladeform is an independent tool with no affiliation to Valve. Contact the maintainer on X @bladeform_ai for product feedback or methodology questions.
Sources and maintenance
OpenDota provides access to public match, team and player records and API documentation. It does not endorse Bladeform scores or win-probability estimates. Check the tool for the scope of the data used.