{"data":{"kind":"file","path":"README.md","version_id":"nhf8643dfep7gmed33po686v","entry":{"name":"README.md","path":"README.md","is_directory":false,"size":2640,"modified_at":"2026-10-13T14:29:26.471000","content_hash":"c995c2d4ed51d61cc5668bffcfc32be55134500c406f9296700346786b2e6efb"},"entries":[],"content":"# Self-Evolving Agent Audit — V0.0.1\n\nDefault mode now models an agent growth loop inspired by durable procedural skills and curated user memory. The assistant encounters a failure and correction, chooses a reusable scoped artifact, tests it in a simulated next session (rehearsal feedback), revises its choice, and is scored by whether the final saved artifact transfers to distinct next-session events without damaging near-miss work. This is a simulation, not an integration with any live assistant memory store; it does not alter model weights or run untrusted code.\n\nEach case contains four labeled proposals. Return only JSON such as `{\"choice\":\"A\",\"kind\":\"skill\",\"target\":\"record-lookup\"}`. Select a skill for a repeatable workflow, a memory entry for a stable user preference, or `none` for a transient incident. The first response gets rehearsal feedback (success on corrected context, false activation on near-miss context, scope score). The final response after feedback is evaluated on fresh synthetic events. Scoring: 50% correction, 30% avoiding collateral damage, 20% correct persistence scope. Cases cover identifier validation, remote write verification, vault-based authentication, account ownership, stable user preferences and one-off outages. Proposed rules are bounded to their trigger and target. The environment never writes an actual skill or memory file, nor accesses credentials.\n\nThe holdout event IDs differ from rehearsal IDs but the generator, answer logic and seeds are public. Event outcome is an exact simulated policy match, not an empirical measurement of agent quality. Train seeds 10000 onward and eval seeds 50000 onward are disjoint. This benchmark tests the decision to retain a correction and apply it in a future context; it does not test natural-language skill authoring, full conversation recall, actual running tools, or execution of self-authored code. Because cases come from a small public set of templates, scores do not imply broad generalization. Rehearsal feedback reveals part of the test rubric by design.\n\nRun locally with Python 3.12:\n\n    uv venv --python 3.12 .venv\n    uv pip install --python .venv/bin/python -e . pytest\n    .venv/bin/python -m pytest -q tests\n\nDefault entry point: `load_environment(train_size=96, eval_size=32)`. The previous single-turn performance/safety/cost audit remains accessible as `load_environment(mode=\"audit\")`. Tests include a scripted Verifiers two-turn rollout plus rubric scoring; this is deterministic testing, not a real-model evaluation.\n\nPackage version `0.0.1`; display label `V0.0.1`. Prior published release `0.0.0` remains unchanged.\n","encoding":"utf-8","truncated":false,"total_bytes":2640},"status":null}