Working paper presents nine safety invariants for autonomous agents, highlighting operational requirements and implications.
Series S4 (Stewardship & Governance) of the Viridis Canon — the operational alignment layer. Nine machine-verified safety invariants (33 Lean 4 theorems, zero sorry) that an autonomous conservation agent's runtime can be required to satisfy: (I1) authorized-view non-interference, (I2) secret non-observability, (I3) delegation containment (child scope ⊆ parent), (I4) authority preservation under trusted equivalence, (I5) spend safety, (I6) uncertainty-gated irreversible action, (I7) protected-state equivalence, (I8) blast-radius containment / loss cap, (I9) fail-closed human escalation. Each invariant carries both a positive theorem and a refutation of its unsafe case (non-vacuity audited). Complements the strategic alignment results P3 (Goodhart impossibility) and P9 (conservation operator). Scope. Lean certifies the safety properties as stated, not that deployment code refines them. Working paper; not peer-reviewed.
No takes yet. Share an insight, caveat, or question.
Hart et al. (2026) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: