← ClaudeAtlas

flock-frontier-contractlisted

Run the Flock/snark.fast prover-optimization campaign as a contract-bound frontier effort with an exact live target, definitions, settlement predicate, explicit non-results and forbidden moves, diverse approach registry, adversarial audits, durable long-run state, same-binary cross-host screens, and evidence-gated promotion. Use this skill whenever the user asks to beat the snark.fast bar, become the Flock leader, solve a hard Flock performance problem, run an autonomous frontier campaign, coordinate multiple Flock optimization routes, recover the campaign after interruption, or decide whether a candidate is genuinely promotable—even when the user does not say “contract.”
welttowelt/flock-frontier-contract · ★ 0 · API & Backend · score 73
Install: claude install-skill welttowelt/flock-frontier-contract
# Flock Frontier Contract Treat the Flock/snark.fast campaign like a verifier-enforced contract, not an open-ended brainstorming session. The contract fixes what counts, what does not count, which state transitions are legal, and which evidence must exist before a claim can move. ## Ranked contract — literal, not advisory Bind these fixed clauses from the official `Layr-Labs/flock-challenge` benchmark before any optimization work: - work per timed proof: exactly `2^18 = 262,144` independent BLAKE3 compression functions, not whole-message hashes; - machine warm-up: 20 private, serialized, verified trials, discarded from scoring; - measured population: 100 private, serialized, verified trials, each in a fresh worker process; - worker warm-up: one fixed-seed untimed proof inside every worker; - score: `262,144 / median(measured_trial_seconds)`, higher is better; - correctness: every warm-up and measured proof must pass the checksum-pinned trusted verifier; - proof limit: nonempty proof capture no larger than 500,000 bytes; - official runner: Apple M3 Max, 36 GB unified memory, 10 performance cores selected by the ranked configuration, arm64 macOS, Rust 1.97.0, and candidate build with `-C target-cpu=native`. Only these solver surfaces are editable: ```text crates/flock-core/src/** crates/flock-prover/src/** ``` Treat the harness, worker wrapper, trusted verifier, Cargo manifests and lock, setup and benchmark scripts, workflow, `benchmark.json`, and score path