SmiteVoodoo Methodology

Methodology — How the AI Scorer Works

Every build recommendation on SmiteVoodoo flows through three distinct systems kept deliberately separate:

1. AI Recommendation — an internal scoring model in js/app.js that ranks items against the selected playstyle, the god's damage type, scaling profile, specializations, and matchup threats.
2. Arena Meta Influence — optional nudge sourced from winning Arena item frequency and Arena win rates. Scored as 60% playstyle fit + 40% pick popularity, never replacing the AI engine.
3. Arena Winner Build Reference — real winning builds scraped from smite2.live, shown as a reference layer on each god's page.

Scoring inputs

• PLAYSTYLE_WEIGHTS — 26 playstyle profiles (burst, DPS, crit, ability, lifesteal, tank, healer, peel support, tank killer, on-hit shred, anti-basic, anti-ability, anti-CC, chase, and anti-shield) each mapping stat priorities.
• Whole-build search — a constrained beam search compares complete legal item sets instead of locking every slot after one greedy choice.
• Target and budget controls — enemy HP, protections, incoming damage source, and gold cap feed the same objective used to select the final build.
• scoreItem() — weights every stat on the item by the active playstyle, plus bonuses for tags that match the god's specializations and damage type.
• GOD_THREATS — per-god threat table (CC, healer, execute, AA, shield, tank) used by applyMatchupBoosts() when you enter enemy gods.
• GOD_KIT_MODS — per-god weight overrides that reflect known kit quirks (e.g. Nu Wa's minion scaling).
• GOD_ITEM_FLAVOR_BONUSES — hand-tuned item affinities for specific gods where pure stat math misses flavor-fit.
• Aspect weight mods — aspects can flip primary scaling and re-weight playstyle; we apply the mods before scoring.

What we don't do

• We don't blindly copy the most frequent Arena build. Match frequency is a conservative nudge and never overrides legality, matchup counters, or the AI scorer.
• We don't hardcode "meta picks" — every item goes through the same scoring function.
• We don't track users or sell data.

Full technical detail lives in docs/SYSTEM_MAP.md in the repo.