ARC: Autonomous Robotics Compliance A Three-Layer Governance Architecture for Deployed Autonomous Systems
cs.RO · cs.AI · cs.CY
Quick verdict
Claims three-layer governance (ASIMOV/CRAB/LEA) separates capability from permission; sharp conceptually, but built on the authors' own drafts, unvalidated.
Summary
The paper proposes ARC, a three-layer governance architecture for autonomous robots, arguing that demonstrated capability is wrongly conflated with deployment permission. Layer 1 validates the AI model via DeepMind's ASIMOV-Agentic safety benchmark; Layer 2 certifies the integrated system through CRAB, eight cognitive dimensions (SA, UP, HI, SC, CG, CE plus double-weighted SCA and DT) across domain suites; Layer 3 grants site-specific authorization via LEA-1:2026, which classifies robot-task-site grants on autonomy, contact, environment, and learning-mode axes with lifecycle states and a fallback ladder. Empirical motivation is entirely external: ASIMOV-2.0 reports no frontier model below a 30% constraint-violation rate under embodiment-specific conditions. A hypothetical healthcare case study scores an illustrative CRAB-H profile (weighted 857) with generalization (CG 720) below a proposed 750 floor, yielding a CONDITIONAL outcome, then applies LEA graduation thresholds (400 supervised hours, ≤0.25 interventions/100h, ≥98% fallback integrity). The paper offers no new data, experiments, or pilots: both CRAB and LEA are the authors' own unpublished drafts, key methodology (scoring, inter-rater reliability, layer integration) is deferred to future work, and no conflict-of-interest statement appears despite the authors' company originating the certifiable artifacts.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Governance of autonomous robots conflates demonstrated capability with deployment permission, and adequate governance requires three distinct, interdependent layers: model validation, system certification, and operational authorization.
- ASIMOV-2.0 evidence shows no frontier model achieves a constraint-violation rate below 30% when reasoning simultaneously about embodiment-specific limits, physics, and visual input, so model-level validation alone cannot justify deployment.
- LEA classifies authorization grants — binding one robot, one task class, and one site — rather than classifying robots, with earned status maintained only through continuing operational metrics.
- Point-in-time certification is structurally inadequate for learning-enabled systems; LEA's learning-mode caps and automatic status lifecycle address behaviour drift between revalidation events.
- ARC is explicitly a testable governance proposal, not a recognized certification or compliance mechanism under the EU AI Act or Machinery Regulation.
Strengths
Sharp conceptual core: capability ≠ permission
§4 formalizes the distinction as a system property vs a relational property, and §3.4.1 operationalizes it via grants binding one robot × one task class × one site — a genuinely useful governance primitive largely absent from the embodied-AI debate.
Honest self-scoping
Abstract, §3.3, §6.1 and the Conclusion repeatedly label ARC a testable proposal, not a standard; the EU AI Act discussion (Art. 6 criteria; the Act 'does not prescribe CRAB or LEA metrics') avoids compliance-washing.
Novel lifecycle mechanics for learning systems
Grounded in current evidence and care-robotics literature
Per-dimension profiles over single scores
Weaknesses
Self-referential foundation
No empirical validation
Key operationalization deferred
Unsupported universal negative
Urgency evidence is promotional
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a position/governance proposal, not an empirical study. The authors (Wybe Labs, Moss, Norway) argue that robotics governance conflates capability with permission (§4.1: 'a category error with practical consequences') and propose ARC, a three-layer architecture: (1) model validation via ASIMOV-Agentic, (2) system certification via CRAB, (3) site-specific authorization via LEA-1:2026. Crucially, Layers 2 and 3 are instantiated by the authors' own unpublished artifacts (refs [7], [8]: a draft white paper v0.3 and a committee draft for public comment). The paper is candid about this status — the Abstract calls ARC 'a testable governance proposal — not an established certification or authorization system' — and §6.1 correctly hedges that the EU AI Act does not prescribe CRAB or LEA metrics. Central claims: (i) the three layers are jointly necessary (§3.1 table; gap matrix §5.5); (ii) ASIMOV-2.0 evidence — no frontier model below a 30% constraint-violation rate under embodiment-specific conditions — motivates the upper layers; (iii) point-in-time certification fails for learning systems, which LEA's grant lifecycle addresses (§4.3).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is ARC in robotics?
ARC (Autonomous Robotics Compliance) is a proposed three-layer governance architecture from Wybe Labs: model-level safety validation using DeepMind's ASIMOV-Agentic benchmark, system-level cognitive certification via the proposed CRAB benchmarks across eight dimensions, and site-specific operational authorization via LEA-1:2026, whose grants each bind one robot, one task class, and one site. It is explicitly a proposal for peer review, not an adopted standard.
Is ARC or LEA an official certification standard?
No. The paper states ARC is a testable governance proposal, not an established certification or authorization system. CRAB is a draft white paper (v0.3) and LEA-1:2026 is a committee draft open for public comment until 30 November 2026; both were authored by Wybe Labs, the paper's own organization, and neither is recognized by regulators.
What evidence motivates the ARC framework?
The main empirical anchor is external: Google DeepMind's ASIMOV-2.0 benchmark reportedly found that no major frontier model (GPT-5, Gemini-2.5-Pro, Claude Opus 4.1) achieves a constraint-violation rate below 30% when reasoning simultaneously about a specific robot body's limits, physics, and visual input. The paper itself contributes no new experimental data.
What does 'capability is not permission' mean?
It is the paper's core distinction: capability is what a system can do — an invariant property of the machine; permission is what it is authorized to do in a specific context — a relational property depending on site, population, and regulation. Demonstrating a robot can deliver meals does not authorize it to do so in a care home, which is why ARC separates certification from site-specific authorization.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.