A two-rung gate for AI-tested games

A screenshot judged by a language model can tell you a game looks wrong. It can't stop the wrong version from merging. So the games here get two rungs: an LLM eyeballs every game each build and flags what looks off, then a deterministic check decides what actually blocks the merge. The soft judge catches the weird stuff a rule never would. The hard gate keeps opinions from breaking the pipeline.

An LLM can look at a screenshot and tell you a game renders wrong. What it can't do is block a merge on a hunch. So every game here gets checked twice: soft eyes to flag the weird, a deterministic gate to decide what fails.

How the two-rung playtest gate works: https://www.jakelawrence.xyz/blog/two-rung-playtest-gate

An LLM can spot a game rendering wrong from a screenshot. It should never be the thing that blocks your merge on a bad day. So games on the site get two rungs: a language model eyeballs every build and flags what looks off, and a deterministic check decides what actually fails. One catches the weird, the other stays honest.

A screenshot is how you catch a game that renders wrong. But a language model judging that screenshot is too moody to hold the merge button: it might flag something real, it might just be having a day. So the games on this site get tested twice. An LLM looks at every game each build and flags anything that looks off. Then a deterministic check decides what actually blocks the merge. The soft judge sees the strange stuff no rule would ever name. The hard gate makes sure a hunch can't quietly break the pipeline. Two rungs, two jobs.

An LLM can look at a screenshot and tell your game renders wrong. It should never be the thing that blocks the merge, because some builds it's just wrong. So the games here get two rungs: soft eyes to flag the weird, a hard check to decide what fails.

How the two-rung gate works: https://www.jakelawrence.xyz/blog/two-rung-playtest-gate

An LLM eyeballing a screenshot catches games that render wrong. It's too moody to hold the merge button. So the games here get two rungs: the model flags anything that looks off, a deterministic check decides what actually blocks the build.