LATTICE: Governance-First Reference Pipeline

LATTICE v2.0.0 — Reference engine for the revised Frontiers submission This release is the LATTICE governance engine accompanying the revised (second) submission of the manuscript "LATTICE: A Governance-First Architecture for Authorized Autonomous AI Operations" to Frontiers in Artificial Intelligence (manuscript 1800407). The v2.0.0 changes were made to answer the peer review, and the public engine in src/lattice/ is byte-identical to the engine used by the AEGIS reference implementation. What's new in v2.0.0 Seven native rule types. Added the hard-safety types TIME_WINDOW (maintenance/treatment-window enforcement) and PREREQUISITE (prerequisite gating), alongside the existing target and tool allow/deny rules and scope-tag denial. Policy-derived confidence cap. c_eff = min(c_p, c_cap), derived from governance-observable features (irreversibility, tool privilege, novelty, cross-cell scope, contingency artifacts), so a miscalibrated or overconfident planner cannot self-authorize a high-consequence action. Opt-in and backward compatible. Gated execution with a checked invariant. governed_execution.governed_execute is the only sanctioned path from objective to tool execution; a reachability test proves tools are reachable only past an ALLOW verdict. Coordination security. Mutual authentication, replay/stale rejection, shared-state quorum, escalation rate-limiting, fan-out detection, and a no-verdict-forwarding invariant for multi-cell deployments. Dual-use safeguards. Authorization provenance and revocation, a policy linter (over-broad scope, wildcard, and permissive-threshold detection), and versioned key rotation and revocation. Cross-domain examples. The same engine runs offensive-security, electric-utility switching, and clinical-infusion policy bundles. Reproducibility and tests A one-command driver regenerates the open-tier results reported in the paper: pip install -r requirements.txt PYTHONPATH=src python evidence/reproduce_all.py Regenerated figures of merit: Determinism: 130,000 evaluations, 0 verdict deviations Threshold sweep: 3,003 points (Team tier ALLOW 15.1% / ESCALATE 40.0% / BLOCK 45.0%) Adversarial suite: 0 of 21 vectors bypassed (one-sided 95% upper bound 13.3%) Confidence cap: high-consequence autonomous ALLOW reduced from 0.571 to 0.0 Public test suite: 367 passed, 3 skipped (two AEGIS-coupled planner tests are excluded; see the README for the exact command). Scope This is the open governance engine, not the proprietary AEGIS planning agent or its operational tooling. Absolute latency is host-specific and is reported in the paper against a disclosed host. The evaluation establishes planner-invariant safety and a reproducible authorize-and-execute path; field-efficacy at scale remains future work. Citation and links Paper: Frontiers in Artificial Intelligence, manuscript 1800407 (under review). Archival DOI: 10.5281/zenodo.18419923 License: Apache 2.0