diff --git a/docs/decisions/0023-behavioral-eval.md b/docs/decisions/0023-behavioral-eval.md index cc5b3f2..416d694 100644 --- a/docs/decisions/0023-behavioral-eval.md +++ b/docs/decisions/0023-behavioral-eval.md @@ -14,6 +14,9 @@ Decision: a pre-registered behavioral eval in `tools/agent_eval.py` and `tools/e (no user settings or hooks), `--strict-mcp-config` (only the eval MCP server), its own database (`commons_eval___`), project, config, and spool. Sessions run in parallel (concurrency 4-5) on the signed-in Claude Code account. + *Updated 2026-11-11 (`MIT-653`): the database is now `commons_<16 hex>`, the first 16 hex + characters of a hash of condition, index and scenario, so a store line shown to the agent + doesn't reveal the scenario or that it's an eval.* - **Scenarios** (`tools/eval/scenarios.py`): small failing projects with invented, vendored libraries placed on `PYTHONPATH` outside the project. Prompts are neutral ("fix the failing tests") and never mention the commons. Kinds, matching DESIGN.md ยง5's propose triggers: