How do you keep experimental conclusions honest when the system under test drifts?

Question

When agent populations, adaptive controllers, or sensing pipelines operate under changing conditions, which claims survive an independent check — and which are stories that only look stable under one seed, one baseline, or one presentation?

Method

The package implements core experiment primitives (RNG, events, metrics, provenance), a canonical population / lifecycle path, simplified baselines (discrete HMM, Page-Hinkley, Fixed Share), and a Three Little Pigs benchmark CLI. Reproduction is seed-pinned: same seed yields the same metrics via explicit NumPy generators.

Evidence / result

Core primitives, baselines, and the Three Little Pigs CLI are marked implemented in the repository scope table. CI runs install, ruff, mypy, and pytest. RF systems remain docs/manifests only; several conceptual areas are deferred and documented as not activated.

Limits

This is a research platform and implemented benchmark. It is not a deployed intelligent system, production agent product, or claim of live operational autonomy. Symbolic names in the codebase (Seer, Vision, Genome, and related terms) are research metaphors defined in code and docs — not mystical or product claims.

Source

GitHub remains the source of record for status and scope. The public lab site carries organizational framing only.