Independent

A successful outcome must move beyond "normative aspirations" (calling for AI to be "inclusive" or "safe") and establish falsifiable architectural constraints. Success would be defined by: Adoption of the Corrigibility Invariant: Recognition that AI systems are only legitimate if they are structurally reversible. The LWD-R Standard: Establishing that transparency requires the disclosure of not just code, but Logic, Weights, Data, and Representation (ontology). Compute Equity: A global commitment to address "Compute Capture," ensuring that the Global South has the physical capacity to FORK (retrain) models, not just run proprietary weights. The Action Boundary Protocol: A requirement for a deterministic "envelope" around stochastic AI to prevent Governance Denial of Service (GDoS) and preserve human oversight. Working papers below https://anivar.net/corrigibility/dpi/corrigibility-framework-dpi.pdf https://anivar.net/corrigibility/epi/corrigibility-framework-ai.pdf

Transparency, accountability, and human oversight;Open-source software, open data and open AI models;Interoperability of governance approaches;AI capacity-building;

These selections address the Structural Preconditions for public trust: Transparency (LWD-R): Essential to prevent "Ontological Capture," where a model's hidden categories become non-contestable administrative facts. Capacity-Building: Legal permission to modify AI is meaningless without the compute resources to retrain models. Addressing Compute Capture is the only way to satisfy the FORK test at a global scale. Interoperability: We must ensure "Functional Exit Equivalence," allowing citizens to move between different AI-mediated infrastructures without losing "Workflow Sovereignty." Human Oversight: This must be enforced via Action Boundaries-deterministic validation layers that ensure probabilistic AI outputs remain under binding human-defined constraints.

Yes, two critical issues are currently overlooked: Variety Drift: AI models are "frozen snapshots" of reality. As the world evolves, a "Temporal Variety Gap" grows. Governance must treat corrigibility as a persistence condition, requiring mandatory revalidation and retraining cycles as environmental variety expands. Epistemic Delegation (Workflow Capture): When states adopt proprietary "AI Blocks" or orchestration layers, they risk a high Workflow Capture Coefficient (WCC). Even if the underlying model is "open," if the workflow orchestration is proprietary, the state loses its ability to alter policy-mathematically surrendering its interpretive sovereignty to a vendor.

The current governance gaps are driving a shift from the Rule of Law to a Rule of Workflow, where administrative power is increasingly mediated by proprietary, incorrigible systems. Significant Challenges Ontological Capture: The primary gap is the lack of transparency in Representational (R) schemas. AI models in welfare and identity sectors generate classifications (e.g., "high-risk") that become non-contestable facts. Without the CODE test (LWD-R disclosure), these latent categories bypass democratic review, creating systemic exclusion that cannot be legally remedied. Compute Capture: Our sector faces a "training forkability" crisis. While "open weights" are marketed as transparent, the astronomical compute cost to retrain models remains a structural barrier. This renders the FORK test impossible for non-corporate actors, meaning we can observe AI errors but lack the physical infrastructure to fix root causes in the training data. Governance Denial of Service (GDoS): In the emerging Agentic Economy, the absence of Action Boundaries is a critical risk. Stochastic AI agents interacting with mandatory infrastructure lack the biological resilience to handle bureaucratic friction, potentially leading to automated queue saturation and systemic collapse. Significant Opportunities Edge Topologies: Shifting toward Mobile Edge and Small Language Models (SLMs) restores the EXIT guarantee. Local inference allows citizens to perform essential functions without central telemetry, bypassing "Workflow Capture." Standardised Agentic Skills: Adopting open protocols like SKILL.md provides a deterministic "envelope" for AI. This allows the state to swap underlying models without losing Workflow Sovereignty, driving the Workflow Capture Coefficient (WCC) toward zero and ensuring that administrative power remains structurally reversible and accountable to the governed.

The AI Dialogue can move international cooperation beyond "normative alignment" toward architectural interoperability. Its most vital role is establishing a Universal Standard for Corrigibility, ensuring that AI-mediated infrastructure remains structurally reversible and accountable to the governed. Specifically, the Dialogue should: Establish the LWD-R Transparency Standard: Standardize the disclosure of Logic, Weights, Data, and Representation (ontological schemas). This prevents "Ontological Capture" by ensuring that the categories through which states "see" their citizens are transparent and contestable across borders. Coordinate on Compute Equity: Address "Compute Capture" by facilitating global access to retraining infrastructure. International cooperation is hollow if the Global South can only run proprietary weights but lacks the physical capacity to FORK (retrain) models to suit local contexts. Harmonize Action Boundary Protocols: Promote a deterministic "envelope" for autonomous agents (the Action Boundary Protocol). This ensures that agentic cross-border transactions do not trigger a Governance Denial of Service (GDoS), maintaining systemic stability as we transition to an agentic economy. Monitor Variety Drift: Create a multilateral mechanism to track the "Temporal Variety Gap." Since AI models are frozen snapshots of reality, the Dialogue can coordinate mandatory revalidation cycles to ensure models do not drift into instability as global environmental variety expands.

The AI Dialogue should build upon the G20 New Delhi Declaration, the UN Global Digital Compact, and the EU AI Act, but it must bridge the gap between "Presence" (policies on paper) and "Proof" (machine-verifiable legitimacy). Key Initiatives to Connect With: Open Source Initiative (OSI): Align with the Open Source AI Definition 1.0 to distinguish true training forkability from performative "open-washing." Agentic Skills (agentskills.io): Adopt this open-standard orchestration protocol to enforce deterministic boundaries around stochastic inference, ensuring Workflow Sovereignty. NIST and International Standards: Build upon safety-testing benchmarks but extend them to include Variety Drift monitoring. Added Value of the AI Dialogue: From Outcomes to Constraints: While existing mechanisms focus on "behavioral" outcomes (bias, accuracy), the AI Dialogue can focus on structural preconditions. It asks not "is the AI good?" but "is the AI correctable?" Mitigating Workflow Capture: It provides the mathematical tools, such as the Workflow Capture Coefficient (WCC), to help nations evaluate whether procuring specific AI systems constitutes a surrender of interpretive sovereignty. Institutionalizing FEE: The Dialogue can formalize Functional Exit Equivalence (FEE) for essential AI services, ensuring that even when a system is mandatory for survival, architectural guarantees preserve a "credible error signal" and prevent population-scale capture.

Stakeholders must move from consultation to adversarial verification. The Dialogue should be structured around the "Adversarial Manifest" model: States & Operators: Must provide an "Operator's Affidavit" (infrastructure.json), making machine-readable claims about their system's LWD-R transparency and Action Boundaries. Civil Society & Academia: Act as Adversarial Auditors. They should be funded to produce "Auditor's Findings" (audit.json) that red-team operator claims, specifically testing for Ontological Capture and Workflow Capture. Technical Communities: Should maintain the open Agentic Skills protocols and compute-accessibility indices to measure "Training Forkability." Recommendations for Structure: Layer-Decomposed Tracks: Separate tracks for the Identifier, Credential, and Wallet layers, applying the Weakest-Link Principle to ensure transparency at one layer isn't neutralized by opacity at another. Continuous Monitoring: Replace "one-off" summits with a permanent Variety Drift Observatory to track the divergence between frozen models and evolving environmental variety.

The most excluded perspective is the "Marginal User"—those for whom AI failure is existential rather than inconvenient. Who is missing: The Global South, not just as consumers but as "Sovereign Retrainers"; the digitally illiterate trapped in "Roach Motel" grievance systems; and independent researchers blocked by "Compute Capture." How to include them: FEE Mandates: Require all "high-stakes" AI to implement Functional Exit Equivalence (FEE), ensuring protected non-digital fallbacks for those excluded by biometric or algorithmic failure. Compute Grants: Establish a Global Forking Fund to provide the GPU-hours necessary for marginalized communities to retrain models that reflect their local Representational (R) ontologies. Legal Safe Harbours: Protect independent auditors from "Security-through-Obscurity" laws that criminalize measuring systemic error rates.

To move beyond "diplomatic theater," the Dialogue should utilize: Algorithmic Red-Teaming Marathons: Live, public "Capture the Flag" events where stakeholders attempt to bypass the Action Boundaries of proposed governance frameworks. Digital Sovereignty Simulations: Using the Workflow Capture Coefficient (WCC) to simulate how specific procurement choices affect a nation's long-term ability to alter its own policy. "Variety Drift" Stress Tests: Interactive workshops that subject existing models to "out-of-distribution" scenarios (e.g., new fraud types or linguistic shifts) to visualize the Temporal Variety Gap. Machine-Readable Deliberation: All policy proposals should be submitted as SKILL.md manifests, allowing autonomous agents to simulate the impact of the regulations on the agentic economy in real-time.

The most effective governance models move beyond "open-washing" toward structural corrigibility and Workflow Sovereignty. Concrete examples include: 1. Architectural Standards: Agentic Skills (agentskills.io) Adopted by over 18 major platforms, the SKILL.md specification provides a "deterministic envelope" for AI. It separates stochastic inference from execution by hard-coding tool permissions outside the model's context window. This creates a machine-readable Action Boundary, preventing prompt-injections and ensuring that autonomous agents remain under binding policy constraints. 2. Transparency Models: LWD-R (Logic, Weights, Data, Representation) While most "open" models only share weights, the Allen Institute (OLMo) and EleutherAI (Pythia) satisfy the CODE test by publishing complete training code and full data manifests. This enables Training Forkability, allowing independent researchers to diagnose and fix root causes in training data rather than merely observing symptoms in the outputs. 3. Federated Sovereignty: Estonia's X-Road X-Road demonstrates a Protocol Model of the state. By using a federated data exchange layer with open-source MIT-licensed code, it prevents single-point capture. It achieves Functional Exit Equivalence (FEE); if one node is captured or fails, system variety is preserved through other independent nodes, ensuring the infrastructure remains a "common" rather than an enclosure. 4. Machine-Verifiable Legitimacy: Adversarial Manifests The use of JSON Schemas (infrastructure.json and audit.json) converts policy rhetoric into falsifiable assertions. Operators declare their governance claims (e.g., "binding human-in-the-loop"), and auditors red-team these claims. Kontrast between the two generates a machine-readable Workflow Capture Coefficient (WCC), quantifying the risk of surrendering interpretive sovereignty to a vendor. 5. Regulatory Innovation: The EU AI Act's Tiered Approach By distinguishing between "GPAI models" and "high-risk" systems, it mirrors the Variety Drift principle, requiring more rigorous revalidation for systems where the gap between frozen model patterns and evolving reality poses existential risks.