BenchmarksCopy Anchor Link

Measured September 25, 2026 with ev-grep 0.1.4 on an Apple M5 Pro (48 GiB RAM, macOS 26.6.2, ARM64). Rust 1.98.0, cargo build --release --locked, default release profile with thin LTO.

LocalCopy Anchor Link

No model requests. Thirty measured runs after five warmups per case.

Live searchCopy Anchor Link

OpenRouter, typesafe/jev-1.13-20260917, up to four concurrent requests. Five measured runs after one warmup per case.

Times include startup, file reads, network requests, model responses, and captured JSONL output. All measured searches completed without execution errors; all 105 measured file assessments returned uncertain. These are elapsed search times, not isolated model inference times or timings for confirmed matches.

MethodCopy Anchor Link

Every file contains this synthetic TypeScript function, padded to exactly 1,024 UTF-8 bytes with comments and spaces:

export async function save(db, user) {
  return db.users.insert(user);
}

Files are named corpus/file_0000.ts, file_0001.ts, and so on. Each invocation uses a newly written corpus in a temporary directory, so the filesystem cache is warm. Corpus creation and JSON validation are outside the timing.

ev-grep --version
ev-grep --provider openrouter --json 'Writes to the database' corpus/ --dry-run
ev-grep --provider openrouter --json 'Writes to the database' corpus/

Runs are sequential, without retries. Provider-side caching and other machine activity are uncontrolled. The samples are too small for tail-latency claims, and repeated 1 KiB files do not represent a large repository. These measurements do not test decision accuracy or compare ev-grep with other tools.

Raw samples include warmups, measured runs, the exact fixture, model IDs, and executable hash. See Request limits for file sizes, concurrency, and timeouts.