# Matched rerun ledger task

Each agent reads [spec.txt](spec.txt) and one 32-transaction ledger, writes `summary.json`, runs [verify.py](verify.py), and finishes with `BENCHMARK_DONE`. The exact [prompt](prompt.txt), specification and checker are byte-identical copies of this rerun's frozen fixtures.

The five measured rounds use [ledger-01.json](ledger-01.json), [ledger-02.json](ledger-02.json), [ledger-03.json](ledger-03.json), [ledger-04.json](ledger-04.json) and [ledger-05.json](ledger-05.json). Every client receives the same fixture within each round, in each separate timing and memory pass. [ledger-00.json](ledger-00.json) is the excluded warmup.

To try a round, save its ledger as `ledger.json` in an empty directory alongside `spec.txt` and `verify.py`, then give your agent the prompt. This is a small deterministic ledger task, not a general coding-quality benchmark.

[Methodology and provenance](../README.md) · [Full report](../REPORT.md) · [All main attempts](../data/attempts.csv) · [Snapshot checksums](../SHA256SUMS)