(Revised with Atlas Level 6–7 Deep Search v1 + v3; provisional pending further corpus expansion)

Abstract

Most discourse about AI maturity tracks capability—benchmark performance, autonomy duration, tool use, and integration breadth. These indicators are real, but they are not measures of legitimacy, enforceable authority, or cross-system governability. This paper proposes a seven-level governance stack for AI systems (with an optional Level 8) and maps major organizations and research communities onto it.

A core finding, consistent across two independent Atlas deep-search bundles, is that no production AI ecosystem appears to implement Level 6 federal meta-governance or Level 7 public constitutional identity end-to-end for AI runtimes. What exists instead is a mature set of transferable primitives—remote attestation, secure update frameworks, transparency logs, software supply-chain provenance, and federated identity—implemented in other domains. The compositional novelty is the absence: these primitives have not yet been assembled into an AI-specific inter-constitutional negotiation and public legitimacy layer.

This paper is a taxonomy and a claim about gaps, not a victory lap. It is intended as a reference that can be tightened into standards drafts, research agendas, and reference implementations.

I. The Wrong Signal

AI ecosystems increasingly treat capability scaling as a proxy for maturity: stronger models appear to imply safer systems, easier governance, and greater readiness for institutional embedding. The Atlas finding does not dispute capability progress; it disputes the assumption that capability progress implies the existence of enforceable legitimacy machinery.

A system may become more competent while remaining constitutionally ambiguous. It may integrate into workflows while lacking refusal semantics that are mechanically enforced. It may “behave well” without possessing a stable, auditable identity that binds behavior to a canonical governance contract.

II. The Governance Stack (Levels 1–7, optional 8)

Level 1 — Model Scaling and Capability Expansion

Primary object: models and training regimes.
Primary metrics: benchmark performance, generalization, tool competence, autonomy duration.
Typical outputs: better models; improved inference stacks; better evals.
Representative orgs: OpenAI, Anthropic, Google DeepMind, Meta AI, xAI, Mistral, Cohere.
What is being optimized: capability and product usefulness, sometimes with safety overlays.

Level 2 — Interaction Paradigms and Human-Centered Framing

Primary object: user interaction and workflow embedding.
Primary metrics: task completion, trust, usability, collaboration, “human in the loop” posture.
Representative orgs: Humans&, Microsoft Copilot ecosystem, human-centered HCI labs and startups.
What is being optimized: the interface and role boundary as a product choice—usually without artifact-bound legitimacy.

Level 3 — Orchestration and Agentic Execution

Primary object: multi-step planning, toolchains, model coordination.
Primary metrics: throughput, latency, success rates on long-horizon tasks, tool coverage.
Representative ecosystems: agent frameworks, orchestration platforms, tool calling systems; frontier labs experimenting with multi-agent loops.
What is being optimized: execution power. Governance is typically policy-overlay, not constitution.

Level 4 — Constitutional Substrate (Sovereign Runtime)

Primary object: a runtime that can state and enforce its own governing identity.
Required properties: deterministic canonical reference, version binding, override schema enforcement, and refusal semantics—preferably with CI and runtime wiring.
Who is thinking here: formal methods, safety-critical engineering, “declare then verify” infrastructure builders, constitutional governance systems.
Status in AI: rare in production; often discussed, seldom enforced end-to-end.

Atlas contribution: Level 4 has strong non-AI precedents: software supply-chain provenance and policy enforcement (SLSA, in-toto), and secure update governance (TUF). These are not “AI governance,” but they are functional proofs that artifact identity can be bound to enforceable pipeline constraints.

Level 5 — Inter-Constitutional Negotiation (Protocol Layer)

Primary object: interaction between sovereign runtimes with distinct governance regimes.
Required properties: handshake, compatibility negotiation, authority boundary declaration, structured translation of constraints, and refusal interoperability.
Status in AI: not implemented as a coherent layer.

Atlas contribution: the closest non-AI analogue is federated trust negotiation, not generic “APIs.” In v3, Atlas surfaces OpenID Federation as a mechanism family: signed entity statements, trust chains, federation operators, distributed policy distribution, and dynamic federation. This is “inter-realm governance metadata exchange” in another domain. It is not AI-specific, but it demonstrates the architectural pattern.

Level 6 — Federal Meta-Governance

Primary object: a pluralist system in which multiple governance regimes coexist without collapsing into monoculture or chaos.
Required properties: upgrade rules across regimes, dispute resolution pathways, fork/exit conditions, compatibility declarations, and meta-constraints preventing authority laundering.
Status in AI: effectively absent in production. Present mostly as discourse, policy talk, or implicit vendor dominance.

Atlas contribution: the closest analogues are standards bodies and federation ecosystems (identity and internet governance), plus selective lessons from blockchain governance. Atlas v3 strengthens the point: the primitives for cross-realm trust and negotiated compatibility exist; what does not exist is an AI constitutional federal layer that binds those negotiations to enforceable runtime identities and refusal semantics.

Level 7 — External Legitimacy Signaling and Public Verifiability

Primary object: public, third-party verifiable claims about runtime state, provenance, and governance events.
Required properties: auditability and verification without requiring trust in operator narrative.
Status in AI: largely absent end-to-end, despite strong primitives elsewhere.

Atlas contribution: v1 and v3 show mature building blocks:

  • Remote attestation architectures (RATS) and standardized claim tokens (EAT) — prove “what state am I in?”
  • Transparency log ecosystems (Certificate Transparency; Sigstore/Rekor; SCITT architecture) — prove “what was published, when, and has it been altered?”
  • Secure update governance (TUF) — prove “this update path is controlled, rollback-safe, thresholded, and key-rotatable.”
    These systems demonstrate that public verifiability can scale operationally. They do not, today, log “constitutional identity of AI runtimes” or “machine-enforced refusal semantics” as first-class, standardized objects.

Level 8 (Optional) — Civilizational Embedding Under Institutional Destabilization

Primary object: how federal AI governance interfaces with weakening institutions, contested authority, geopolitical fragmentation, and legitimacy crises.
Status: not primarily technical; requires institutional design, standards legitimacy, and adoption strategy.
Note: Level 8 is where “why society needs it” becomes operational rather than rhetorical.

III. Where Organizations Actually Sit (Provisional Mapping)

Primarily Levels 1–3 (capability, interface, orchestration)

OpenAI, Anthropic, Google DeepMind, Meta AI, xAI, most agent/orchestration ecosystems, most “human-centered” startups (including Humans& as described publicly).

Level 3–4 hybrid (enterprise governance overlays; not federal protocol)

Palantir and similar enterprise systems often provide audit, policy, and workflow control. That is governance inside a deployment perimeter. It is not inter-constitutional negotiation, federal compatibility, or public legitimacy signaling.

Levels 4–7 primitives (non-AI, production-grade)

IETF and ecosystem implementations around RATS/EAT, Certificate Transparency, Sigstore/Rekor, secure update frameworks like TUF, provenance frameworks like SLSA/in-toto, and emerging transparency services like SCITT. These are the most important evidence that Level 7 is viable as infrastructure—but they are not currently assembled as AI constitutional federalism.

Levels 6–7 in AI (end-to-end)

Atlas finds none in production.

IV. The Atlas Finding, Restated as a Design Constraint

The relevant conclusion is not “nobody is thinking about governance.” Many are. The conclusion is that:

  1. The ingredients for Level 6–7 exist in other domains (attestation, provenance, transparency, federation).
  2. The composition into AI constitutional interoperability has not been achieved—especially the semantic layer that would make attestations and logs meaningfully express “authorized/unauthorized,” “refusal invoked,” “canonical identity bound,” and “inter-regime compatibility negotiated.”

This shifts the novelty claim from invention to composition and semantic binding.

V. Implications: What a Level 6–7 AI Ecosystem Would Need

A credible Level 6–7 system would have to treat the following as first-class, machine-verifiable objects:

  • Constitutional identity (a canonical reference that is hash-bound and versioned).
  • Refusal semantics (conditions under which execution must fail, expressed as enforceable prerequisites).
  • Inter-regime negotiation artifacts (compatibility declarations, boundary contracts, override translations).
  • Public legitimacy artifacts (attested runtime state; transparency-logged governance events; provenance for governance changes).
  • Federal dispute and upgrade pathways (explicit arbitration routes and fork/exit conditions).

Without these, “governance” remains policy overlay and institutional narrative.

VI. Conclusion (Provisional)

The AI ecosystem is currently scaling capability and orchestration faster than it is building constitutional identity, inter-regime negotiation, and public legitimacy proofs. Atlas research indicates that Level 6–7 AI governance is not blocked by missing cryptography or missing distributed systems theory; it is blocked by the absence of an AI-specific semantic and protocol composition that binds legitimacy to runtime state and makes that legitimacy portable across sovereign regimes.

If institutions require pluralist AI governance under destabilizing conditions, the missing work is not merely “ethics” or “policy.” It is the construction of enforceable federal substrate—using mature precedent primitives, but assembling them into a new layer of constitutional interoperability.

(Standby document; to be amended with additional Atlas results, citations, and any discovered counterexamples.)