Making experiment findings survive.
Most experiment findings die — not wrong, dead. They're faked (verdict ungrounded, collapses under load) or orphaned (correct but reaches no decision, leaks into nothing). Tracking tools (MLflow, W&B, DVC) cover metrics and params; none covers the verdict layer — what was actually established. So the finding leaks.
This is the work of making knowledge survive: findings that are true-checkable (verdict grounded in evidence you can point at) and connected-queryable (stable identity + verdict a downstream can act on). Both, or it dies.
- finding-declaration — a machine-readable standard for declaring experiment findings. 5 required fields, a 5-state verdict vocabulary derived across 8 experimental domains. Verified uncovered across 40+ tracking tools and metadata standards. Dogfooded on 170 real findings.
- findcheck — the reference validator.
pip install findcheck. - Most of your experiment findings are already dead — the first piece on why knowledge leaks and what stops it.
Every artifact under patchwright answers one question: does this make signal more true-checkable, or more connected-queryable? If neither, it's off-thesis and won't be foregrounded here. The principle is testable — you can check it per artifact, and so can anyone else.
The list also contains earlier tools (wildlint, mcp-lint, agent-wall, schemasmoke, mutaprobe) and a number of upstream fix-PR forks. They're maintained or archived as they stand, but they are not the focus — the surface narrowed to findings-integrity in 2026-07. The reason is straightforward: scattered work across unrelated domains does not compound into trust, and a narrowed, consistent catalog does. The earlier work isn't repackaged as something it wasn't; it's just no longer what this publisher points at.
Knowledge doesn't survive because you're careful. It survives because you encoded it to survive.