FAQ
Fair questions.
The ones a skeptical engineer should ask before letting anything optimize their code overnight.
How is this different from an LLM in a loop?
A loop generates code and hopes its checks are meaningful. Sleepy runs an evolutionary search, sends each candidate through the exact evaluator contract in a bounded cell, and only measures survivors under a preregistered paired policy. The result is a reviewable, scoped claim rather than code that merely looks plausible.
What stops a candidate from cheating its own tests?
Eight fail-closed controls, detailed here. In the admitted guarded profiles, candidates run with no external network and zero Linux capabilities against an immutable authority snapshot, and every pass is checked with a deliberately poisoned harness that must fail.
Does my code ever leave my machine?
Not in the canonical executor/job path. Source, tests, prompts, provider payloads, candidates, and raw measurements stay in the client-owned executor. The code-blind control plane receives source-free scheduling state, exact receipts, and authenticated decision projections.
Whose API keys make the model calls?
Yours, direct to the providers you choose. The hosted control plane does not hold your model-provider API keys or proxy model inference; your client-owned worker invokes the models you configure under the budgets you set.
Do you touch my tests?
The hosted service never receives them. Your local worker reads the evaluator authority to build an exact contract, then candidates run against an immutable snapshot they cannot rewrite.
What do I wake up to?
A re-verified champion with local lineage and evidence — or an honest “no winner.” Candidate applications are restored during search; the current CLI finalizes only the verified winner according to the output setting you chose. Merging remains yours.
What languages work today?
Go benchmarks and tests, plus a deliberately conservative Rust profile with Criterion benches. Narrow on purpose: a language ships when its gate is as hard as Go's.
What do I still have to trust?
The primary execution authorities include your host kernel, container runtime, pinned image and toolchain, evaluator authority snapshot, and the Sleepy binary. The audit guide makes those assumptions explicit so a reviewer can verify the admitted boundary rather than relying on a blanket promise.
How do you make money from this?
Any pricing is disclosed before use. Two boundaries are structural today: your model relationship stays direct with your provider, and the canonical hosted control plane is code-blind rather than a resale path for model tokens.
Something this page didn't answer? Ask support or go back home.