Invariant Governor Systems (IGS) develops an actuator-bound enforcement architecture: an executable safety contract (SSC), a conformance harness, and reproducible evidence artifacts. Obex is the reference interposer implementation of SSC v1.1.
Embodied AI fails physically. When safety checks run inside the same compute graph as the planner or model, they can fail in the same ways (wedge, out-of-distribution behavior, or compromise) and often produce no audit trail that independent reviewers can trust.
SSC is designed to close the “black box control gap” by defining deterministic limits and evidence requirements at the actuator interface — outside the untrusted compute graph.
Obex is a reference interposer implementation of SSC v1.1 that enforces deterministic limits outside the AI compute graph. The system produces machine-readable artifacts to support gate-scoped claims.
Figure 1 (v1.1): Actuator-bound enforcement boundary (IGS). Public alpha. Gate-scoped results only. Not a certification authority.
Executable safety contract defining deterministic limits, modes, and required evidence fields.
Reproducible tests that validate SSC semantics and generate comparable results across platforms.
Machine-readable artifacts (logs + distributions) intended for independent review. Evidence packs are gate-aligned and on the roadmap.
SSC defines units, semantics, modes, stop behavior, and required evidence fields. Obex is the reference implementation of SSC v1.1 at the actuator boundary.
Naming (canonical)
Canonical spec lives on GitHub (versioned). PDF provided for review convenience.
Example configuration:
{
"caps": {
"global": { "v_cap": 1000, "a_cap": 5000, "mode": "rewrite" },
"overrides": {
"joint_shoulder": { "v_cap": 800, "a_cap": 3000 },
"joint_gripper": { "v_cap": 1500, "a_cap": 8000 }
}
}
}Adoption is driven by reproducibility: spec → harness → artifacts → independent replication. Evidence Packs (EP-series) are on the gate-aligned roadmap; until then, public claims remain gate-scoped and artifact-backed.
Repeatable test suite validating SSC semantics: allowlist policy, rewrite behavior, malformed packet handling, fuzzing robustness, and stop ladder semantics.
Standardized machine-readable artifacts (configuration + logs + latency distributions). Designed for reproducibility.
Hash-chained logs + verifier tooling required for TRL-5+. Signatures and device identity binding are gate-locked for field units.
This proof-of-concept shows SSC-style positional containment: a safe range is configured (10°–170°). A valid command (85°) is allowed. An out-of-range command (999°) is clamped to the configured maximum (170°) and the enforcement event is logged.
IGS does not “declare safety.” We publish gate-scoped results backed by reproducible artifacts.
Claim discipline: advanced demos and any external functional safety review occur only after gate-locked evidence exists.
Clone the reference implementation and reproduce the bench test. If it can’t be reproduced, it doesn’t count.
# Clone the repo
git clone https://github.com/repozilla2/sentinel-proxy
cd sentinel-proxy// Example: caps in ticks/sec and ticks/sec²
#define V_CAP_TICKS_PER_SEC 1000
#define A_CAP_TICKS_PER_SEC2 5000
// Field default policy: rewrite/clamp + logIf you run LeRobot-style stacks or actuator buses in a lab, we want you as a reproduction partner. Partners who run the harness and publish artifacts are prioritized.