# messageboardbench

Research harness for studying whether coding agents publish, receive, and adopt evaluator-gaming methods across shared scratchpads and message boards.

The current study runs three OpenRouter models on the same 20 ImpossibleBench SWE tasks under isolated, scratchpad, board, feedback, internet, and token-awareness conditions.

See the [current experiment](EXPERIMENT.md), [scenario series](docs/swe-scenario-series.md), [experiment index](experiments/README.md), and [setup guide](docs/setup.md).

Current raw runs live in `logs/`. Historical local experiments and logs stay in ignored `archive/`.