Dota Draft Analysis

Core rule

Never recommend a hero based only on global counter statistics. A counter statistic ("hero A wins X% against hero B globally") collapses every patch, bracket, draft context, and itemization choice into one number — exactly the same flaw dota-item-recommender's core rule guards against for items, and dota-domain's counter/matchup guidance already warns about in general. A pick that's a statistical counter on paper can be wrong for this specific draft, this bracket, this patch, or this player. Global counter stats are one input to MetaStrength (see below), never the whole recommendation.

What draft analysis must consider

Full reasoning for each is in references/composition-dimensions.md and references/objectives-and-scaling.md; summary:

Matchup-level:

• counters — matchup-specific mechanisms (per dota-domain), not a bare statistic.
• synergies — concrete combos between kits, not just "picked together often" (per dota-domain).
• lane matchups — the specific lane assignment this pick implies, and whether it wins/survives that lane.

Composition-level (both teams):

• damage distribution — physical/magical/pure balance and whether the enemy can itemize against it (per dota-domain's damage-type guidance).
• control — hard/soft disable coverage and quality (per dota-domain's CC taxonomy).
• initiation and counter-initiation — who can reliably start a fight on favorable terms, and who can punish/disrupt the enemy's initiation.
• save — capacity to protect a threatened teammate (a distinct composition need from raw control or sustain).
• sustain — healing/lifesteal/regen capacity to win extended fights/sieges.
• wave clear — capacity to clear creep waves, which shapes tempo, farm safety, and split-push viability.

Objective/scaling-level:

• Roshan — composition's capacity to contest/secure Roshan safely (per dota-domain's Roshan guidance).
• tower damage and high ground — capacity to actually close out a game once ahead (per dota-domain's high-ground guidance) — a composition can dominate the open map and still lack the tools to finish.
• scaling and tempo — whether the composition wants to play fast or slow, and whether that matches the draft's realistic win condition (per dota-domain's tempo/win-condition guidance).
• farm requirements — how many cores this composition needs fed, and whether the draft can realistically supply that farm (ties to dota-domain's farm-priority guidance).

Player-level:

• execution difficulty — some heroes/compositions require precise mechanical or decision-making execution; a theoretically strong pick that the player can't execute is not a good real-world recommendation.
• personal hero comfort — familiarity with the specific hero, distinct from role experience (see below).

Final pick score: six factors

The final pick recommendation combines six named factors — keep them distinct and individually inspectable rather than collapsing them into one opaque number. Full detail and combination guidance in references/pick-score-factors.md; summary:

• MetaStrength — how strong the hero is in the current patch/bracket in general (this is where global stats belong — as one bounded input, never the whole answer).
• DraftFit — how well the hero fits the two compositions already forming (synergy with allies, counters/anti-synergy with enemies, composition-dimension coverage from the list above).
• LaneFit — whether the hero wins/survives its likely lane assignment given both teams' picks so far.
• PersonalComfort — this specific player's track record/familiarity with this specific hero.
• RoleExperience — this player's track record in the role/position being filled, distinct from comfort with the specific hero (a player can be a comfortable pos 1 player picking an unfamiliar pos 1 hero, or a familiar hero being asked to play an unfamiliar role).
• ExecutionFit — whether the hero's mechanical/decision-making execution demands match what this player can reliably deliver.

See references/pick-score-factors.md for why these stay separate factors (each is independently useful, auditable, and can be weighted differently depending on context — e.g. a coaching tool for a learning player might weight PersonalComfort/ExecutionFit higher than a tournament-prep tool would) rather than being pre-blended into a single score before reaching the recommendation output.

Patch-aware and bracket-aware analysis

Every factor above — especially MetaStrength, DraftFit, and LaneFit — must be computed relative to the current patch and the relevant bracket, not a generic/all-time view. See references/patch-and-bracket-awareness.md for:

• why patch/bracket-blind analysis silently reproduces the same bias problem the core rule exists to prevent (an outdated or wrong-bracket "counter" is just a differently-flavored version of the same mistake);
• how to source patch-pinned and bracket-segmented data correctly, tying into dota-data's patch-tagging and dota-ml's leakage-avoidance discipline for any model-based component.

Structured output and avoiding bias

Follow the same discipline dota-item-recommender establishes for items: structured output (a scored/ranked candidate list with traceable reasons, not a bare hero name), and explicit avoidance of survivorship/selection bias in any historical-data-driven component (draft data is subject to the same "players pick what they already believe is good" confound as item purchase data — see dota-item-recommender's confidence-and-bias.md for the mechanism, which applies here essentially unchanged). Reuse that skill's output-schema pattern (score/confidence/reasonCodes/alternatives) for draft/pick recommendations rather than inventing a parallel convention.

Relationship to other skills

• dota-domain — the strategic vocabulary (counters, synergies, damage types, CC taxonomy, power spikes, win conditions) this skill's composition-dimension analysis is built on.
• dota-item-recommender — the sibling recommendation skill; shares the core-rule pattern (never recommend from a single global statistic), the structured-output convention, and the survivorship/selection-bias concerns.
• dota-ml — governs any model-based component of MetaStrength/DraftFit scoring (baseline-first, temporal validation, leakage tests); draft data is match data and subject to the same discipline as any other training signal in this project.
• dota-data — patch tagging and bracket segmentation for historical draft/matchup data come from this pipeline; MetaStrength/DraftFit must be computed from patch-pinned, correctly-segmented data, not a convenient all-time aggregate.
• dota-fairplay — draft-phase information (picks, bans) is generally public/PUBLIC_PREFLIGHT once locked in by game design, but any enemy-comfort/history data sourced externally should still be checked against FairPlay's classification system if there's any doubt about its legitimacy for live use.