Methodology
How current rankings are produced, what they cover, and where they fall short.
Suites
The active suite is Zombie Flamethrower Showdown. Each build is evaluated as a complete playable game, not as isolated text answers. The suite stays fixed within a round so builds face the same bar.
Prompt
Both builds start from the same game-build brief: a flamethrower survival round against waves of zombies. The brief fixes the theme, controls, and win/lose shape while leaving art, tuning, and feel to each build.
Playability checks
- Loads and starts without errors on desktop and mobile widths.
- Movement, aiming, and firing respond to keyboard, mouse, and touch.
- Waves advance, scoring updates, and game-over plus restart work.
- No blocking audio, layout, or input faults during a full run.
Scoring
Judges score each build on a 5-dimension rubric, 20 points per dimension, 100 total: Controls & Mobility, Combat Physics & Weapon Mechanics, Content & Enemy Variety, Audio & Sound Design, and Visual Polish & Game Feel. A build with no audio scores 0 on audio. A feature earns points only when it is genuinely implemented and reachable — on-screen text, comments, and dead code score zero. This v2 audit replaced the earlier v1 method, which credited claimed features that were absent from the code and therefore reported scores up to 40 points too high; the v1 leaderboard has been withdrawn. observed during hands-on review. Dimensions outside this round, such as reasoning or coding subscores, show as Not evaluated. Pricing and speed show as Not measured until dedicated measurements land.
Votes, likes, and ratings
Community signals sit alongside judge scores. Votes pick a preferred build in a head-to-head matchup, likes mark a build worth replaying, and ratings give a 1 to 5 star view on fun and polish. Signals never rewrite a published Showdown Score; they provide context on which builds players return to.
Freshness
The current snapshot is the September 2026 round. Each round records its evaluation date on model and benchmark pages. When a new round ships, prior scores stay visible with their dates so movement can be traced.
Limitations
- Nine models, eleven audited builds — results do not generalize to other models or other prompts.
- Game-build quality only — no claims about chat, code, or reasoning strength.
- Judge review includes human judgment and carries taste variance.
- No price or speed claims in this round; those pages state Not measured.
Why rankings change
Rankings move when a new round is published, when a build is updated and re-evaluated, or when review standards are tightened and noted. Dates on every score make it clear whether movement came from a new build or a new round.