The README advertised "zero bugs shipped over two weeks" on a single user and
two projects. That was an anecdote that read like a result. It is now replaced
by what can be shown, including the parts that are unflattering.

Held-out benchmark, PARTIAL: an existing SWE-bench Multilingual harness resolved
7 of 12 instances actually run (58 of the 70-instance subset produced empty
patches and were never attempted), all on Sonnet in small-fix mode. No same-model
single-agent baseline exists, so the figure has nothing to compare against.

Process behaviour over 340 completed tasks in 6 real projects, 235 with a
parseable stage ledger. Every rate is conditional on the stage returning a
verdict, and the denominators are published per row: a reviewing stage objected
in 39.3% (70/178, or 29.8% of the full ledgered corpus), the fixer needed a
second round in 5.3% (9/169). The loop converges on the first pass almost always,
though the tail is real ...