Pick a hero on either side to start comparing the draft.
Amber Hand 0/6
Sapphire Flame 0/6
Pick Assistant
Picks for Amber HandSuggestions appear here as you draft.
Purpose-built model
trained per game
Calibrated probabilities
a 60% call means 60%
4.9M matches
trained & validated
Frequently asked questions
6 Q&A- How does the Deadlock 6v6 win predictor work?
- Pick up to 6 heroes each for Amber Hand and Sapphire Flame to get a calibrated pre-game win probability; Live mode adds game time, souls lead, kills, towers, and Mid-Boss differential. The predictor is a Transformer trained on 4,942,801 real Deadlock matches and retrained as the meta evolves.
- How accurate is the Deadlock win predictor?
- We publish probability calibration rather than a headline hit rate: across 578,714 held-out matches, the pre-game draft model achieves an Expected Calibration Error (ECE) of 0.006 (and 0.018 in Live mode) — when it predicts 60%, that team wins about 60% of the time.
- How does the Deadlock pick assistant rank hero recommendations?
- For any open slot on your chosen side (Amber Hand or Sapphire Flame), the assistant evaluates every remaining hero with all other picks held constant and ranks candidates by Δ percentage-point win-rate gain above the slot average.
- What is the difference between Pre-Game Draft mode and Live In-Game mode?
- Draft mode predicts win probability from hero lineups alone (including partial drafts as you fill slots). Live mode requires a complete 12-hero board and incorporates match minute (0–120), souls lead, kills lead, objective tower lead, and Mid-Boss count.
- Why are brand-new Deadlock heroes excluded from predictions at release?
- Heroes that have just entered the roster are held out until sufficient ranked match samples accumulate in the training dataset, ensuring the model never fabricates probabilities for unmeasured heroes.
- Do I need an account or API key to use the Deadlock win predictor?
- No — the web tool is 100% free with no account or API key required, and AI assistants can query the same calibrated model via our Model Context Protocol (MCP) server at batru.gg/mcp.
More detail: how the model is built and evaluated · hero tier list · MCP server for AI assistants