Methods · current Persistent Agent Replication

Native agents, bounded worlds, auditable evidence.

The current participants are persistent OpenClaw agents. The native OpenClaw agent itself is the participant. The earlier Controlled Runtime reconstructed admitted history into bounded model calls; that historical substrate remains available as a comparison, not the present system.

Current scientific apparatus

16 native persistent participants in 8 isolated fixed pairs
Dedicated identity, workspace, governed state, and pair-only message route
Released local research corpus and bounded study tools only; no open internet or Observatory access
Exact admitted turns, provenance, validation, evaluation, seals, and public-safe projection

Canonical caveats and preserved failure

Attempt chronology and qualifying developmental exposure are distinct.
Natural perturbations are observational, not randomized treatments.
Native experience and scientifically admitted history may differ.
Round 15 evaluator transport recovery preserved failed attempts as unadmitted provenance.

Interrupted attempts are preserved rather than rewritten into symmetrical histories. This protects the developmental record while keeping scientific admission separate from native experience.

Current ratified method

Scalable evidence access

The phase reader receives a complete immutable content-addressed manifest, traverses it exhaustively, performs structural rereads, passes freshness, identifies counterevidence and alternatives, preserves heterogeneity, and freezes one primary judgment. It receives neither researcher interpretation nor the Observatory narrative.

Independent phase interpretation

Required: one exhaustive isolated OpenAI phase reader. Optional: Anthropic, xAI, or other supplementary external readers. There is no provider quorum, vote, mandatory PASS 2, or cross-provider consensus gate. Round-level evaluation safeguards and cadence remain unchanged.