cyberine

A worker’s own claim is never evidence

Four judges. One of them can be wrong out loud.

Three judges read and reason about the candidates. One of them actually runs them. That asymmetry is why DiffJudge takes the whole width below and the other three share a row.

Gates on machine-checkable evidence

DiffJudge default for code

Takes every candidate diff, applies each one in a throwaway git worktree, and runs the typecheck. A diff that will not apply loses. A diff that applies but breaks the build loses. There is no opinion in the loop, and no way for a worker to talk its way past it.

git applythentsc --noEmit

Every candidate scores 0, 1 or 2. The rationale is the judge’s own wording.

  1. both gatesdiff applies cleanly and type-checks — both gates passed (score 2/2)
  2. typecheck faileddiff applies cleanly but typecheck failed — partial evidence only (score 1/2)
  3. apply faileddiff does not apply cleanly against HEAD (score 0/2)

highest score wins; a tie among passing diffs goes to an LLM tie-break

counts

Consensus

Pure majority across the candidates. Cheap, deterministic, and blind to whether the majority is right.

reads

LlmJudge

A model reads every candidate and argues for one. Useful where correctness is not machine-checkable: prose, design decisions, an answer with no build step.

attacks

AdversarialJudge

Sends N independent skeptics at each candidate, each prompted to refute rather than approve. A finding that survives a majority of refutation attempts is the one that ships. Built for the failure mode the other three share: plausible and wrong.

Hover or focus a judge to watch how it picks.

When the evidence runs out

The verdict says what it saw

01

More than one passes

When several diffs pass both gates, an LlmJudge breaks the tie on correctness and quality. Zero or one passing candidate skips that call.

02

No diff at all

A step where nobody returned a diff is a harmless no-op: the winner is none, and the rationale says so in plain words.

No candidate carries a diff — nothing to apply or verify.

03

No typecheck to run

In a repo with no typecheck script, gate two is skipped, never failed, and the verdict detail says so. No fallback oracle is invented.

Run your first race.

Install it, point it at a repo, give it a task.

Read the docs
run 81cbc07cangle: the verdictrecorded 2026-09-254 workers3 diffs applied1 kept$0.10 spent