the honest slop detector

slopmark

AI models are great at sounding smart and terrible at admitting when they're not. slopmark feeds them tasks with machine-checkable answers, runs every model through the exact same harness, and lets a rule-based verifier — never another AI — decide who actually passed.

run the bench →play with the models leaderboard

364 tasks in the pool17 domains3 completed sessions0 LLM judges. ever.

featured challenges

curated head-to-head sessions — fixed tasks, fixed harness, receipts included

all challenges →

featured

Game Night v1

5 Fireworks models · 10 party-game constraints

winner: GPT-OSS 120B · 90% pass · 7/13/2026

featured

Drawing Contest v1

5 Fireworks models · 10 SVG drawing challenges

winner: GPT-OSS 120B · 100% pass · 7/13/2026

featured

Niche Sprint v1

6 low-tier models · 10 one-shot challenges

winner: Gemini 2.5 Flash · 90% pass · 7/9/2026

+ build your own — pick tasks, bring your key, get an infographic

how a score is made

most leaderboards ask a bigger model to grade the answers. that's a vibe check, not a benchmark. here every point is traceable:

01
task drops
a task with a machine-checkable contract comes out of the seed pool — word limits, JSON schemas, exact numbers, SVG shapes.
02
same harness, always
identical system prompt, temperature 0, capped tokens. no per-model babysitting, no secret scaffolding.
03
rules decide
a deterministic verifier checks the output against the contract and prints the rule-by-rule breakdown. no LLM judges.
04
everything is kept
runs roll up to the leaderboard, sessions land on the wall, and the worst outputs get framed in the hall of shame.