Correctness
Independent SQLite ground truth judges memory results instead of asking an LLM to grade itself.
MESA QA
An external, detachable test-engineer system for long-running behavioral checks against MESA through its canonical MCP surface.
EVALUATION SCOPE
Independent SQLite ground truth judges memory results instead of asking an LLM to grade itself.
Profiles cover sustained remember/recall cycles, temporal updates, corrections, forgetting, and session rotation.
Scenarios restart candidate processes and verify that expected memory survives.
BOUNDED REPAIR
When a reproducible defect is found, QA can write a failing regression, apply a minimal repair, restart the candidate, and verify the result.
The pinned QA report records a live first-write INVALID_ARGUMENT failure. Earlier Lite harness evidence includes a candidate session verdict of FAIL; it does not validate Core 0.9.3.