The proposed public-AI network

Our AI Charter™ proposes an international network of open AI models that would operate under shared public obligations and include a governance-and-evidence layer separate from service provision.

The practical aim is a shared way for people and institutions to find, compare, and reach participating AI systems, inspect their claims, and challenge decisions that affect them — without depending on any one provider.

Why a network

AI is fast becoming infrastructure — the layer through which people learn, work, govern, and make sense of the world. The question is not only which model performs best, but whether people and democratic institutions can inspect, shape, and contest the systems they increasingly rely on.

Open, inspectable models are essential: they widen access, support research and adaptation, and reduce dependence on systems whose inner workings cannot be examined. But a model is only one part of an AI service. Openness alone does not answer basic questions: Who decides where a request runs? Which model, data, tools, and compute node were used? What happens if a provider changes, fails, or withdraws? Who can inspect an incident or challenge an AI-supported decision?

A federation would let participants keep local control while sharing public obligations, standards, evidence, accountability, and fallback paths. Researchers, public bodies, startups, and civic builders could reach plural models through that shared structure. Digital sovereignty does not mean building everything alone; it means being able to inspect, help decide, contest, and switch. Distributing control across a federation is intended to make it harder for any one power to capture the rules or cut off access.

What it combines

The proposal combines five pillars: accountable AI decisions and actions; open, plural models; data and provenance commons; shared assurance and evaluation; and federated public AI infrastructure. This extends the four-pillar framing in the July 2026 project statement by making the governance of particular decisions and actions an explicit pillar.

Accountable AI decisions and actions concerns whether a particular consequential answer or action may proceed: who authorises it, which evidence and limits apply, and how its outcome can be inspected and challenged. Shared assurance and evaluation assesses claims about systems and their operation through evidence and independent review. Its AI Assurance & Certification building block comprises the Charter Commitments, evaluation method and future assurance structure. The two responsibilities support each other throughout use; the pillar order is not an execution sequence. Models, data and infrastructure provide the resources for that use.

Evidence-Gated Agents (EGA) is Robert Schaub’s current focus within the decisions-and-actions pillar. It explores configurable applications and reusable components; their use would not depend on the proposed federation becoming operational. EGA remains early-stage work. Its selected first-prototype direction dynamically examines a normal AI agent's exact decision through FactHarbor before release; the integration is not yet implemented.

The data commons does not assume that all useful training data can be open. Its benefit is clear once, reuse many: handle rights, provenance, and access conditions collectively instead of making every public model negotiate alone.

Here, public describes obligations, accountability, access, and governance — not an operator's legal form. Public, academic, cooperative, nonprofit, and mission-aligned commercial nodes could participate under the same disclosure, portability, evidence, and anti-capture rules.

Values-bound would not mean a bloc of self-described “free” states. Participation would depend on verifiable commitments to human rights, rule-of-law procedures, democratic accountability, transparency, and effective remedy.

Across all five pillars, competence is essential. Infrastructure can be procured; the ability to understand, govern, and improve it must be built through education, research, managerial capability, and institutional learning — and kept broadly distributed. That is also an economic case: value creation, skilled work, and bargaining power grow where institutions help build and decide, not merely buy. Accountable human governance would establish rules, authority and routes to challenge and remedy across all five pillars. Anti-capture governance and public-interest demand hold the pillars together.

How it would work

The intended user path is find → check → use:

1. Find a system suited to the purpose, jurisdiction, language, openness, and capability needed.
2. Check its claims, reviewed evidence, known limits, and unverified areas (questions to ask).
3. Use it as an endpoint, assistant, agent, or evaluator under access rules suited to the user and data.

The same offering is an AI system when found, compared, or held to account and an AI service when served and used — two views, not separate layers.

Discovery must remain plural and neutral because whoever controls discovery can shape the market.

How it would govern action and human consequences

Existing gateways, schedulers, registries, and identity systems mostly answer can this run? The proposed public-interest layer asks may it run — for whom, for what purpose, on which node, with what model and data, under which evidence duties, and with what route to challenge and remedy?

Accountability therefore cannot stop at choosing a model or evaluating a release. The network would connect lifecycle governance — design → deploy → operate → incident → remedy — with five activity/gate pairs for each AI-supported action: Plan → Authorize; Prepare → Submit; Check → Verify; Decide → Commit; Review → Rely. Across both, the five public obligations would apply: purpose-bound; answerable to people; safe, secure, private, and resilient; fair in practice; and open to evidence and correction.

The acting model would remain probabilistic: it could propose an action but could not authorize it. A separate authorization component would deterministically apply declared rules to the frozen proposal and recorded mandate and authority state, returning allow, deny, or escalate; the service producing an external effect would verify that decision again. This can show which declared rules were applied to which recorded state. It cannot establish that the model's evidence is true, that the recorded authority basis is valid, that the rule is fair, or that the institution is legitimate.

The rule is to govern every authority-bearing transition, not every inference. Routine, reversible actions inside an authorized envelope could remain silent. External, irreversible, regulated, person-affecting, or out-of-bounds actions would stop when required authority or evidence is absent — human approval cannot manufacture the missing basis. Preserve an action-scoped record of authority, material evidence and uncertainty, effect, and challenge route — not every token or the whole conversation.

That record protects what was lodged; it cannot expose a determination that was never recorded or was false when created. Closing that gap would require a duty to lodge consequential determinations, an affected person's right to discover that a decision was made about them, and independent oversight able to notice when an expected record is missing.

Those gates are also moments to test AI-supported power and reopen dialogue. Ask who has authority, who benefits and bears the cost, who was heard, whether expected harm is necessary and proportionate and a less harmful path exists, whether affected people can safely disagree, refuse, or challenge and obtain independent review and remedy, and which limits must hold even when they frustrate the goal. Check both specific people and population patterns: individually admissible actions can accumulate into a discriminatory or otherwise unauthorized effect.

AI and Empathy: Dialogue, Correction and Human Answerability treats empathy as observable, correctable behaviour that protects human agency — not artificial feeling. The builder guideline turns that practice into engineering: keep distinct what a person said, what the system inferred, what the person confirmed, and what it has revocable permission to use or remember; seek affected perspectives; carry disagreement and uncertainty forward; and make correction, refusal, and escalation usable. Reopen dialogue only when an unresolved interpretation, permission, missing perspective, or observed consequence could change a transition, following engage → act → observe → disclose → correct → re-engage. Sycophancy is not empathy: models can support perspective-taking but cannot supply legitimate authority, independent review, or remedy. For builders: train for perspective-taking; tune for truthfulness and agency; adapt only with consent; improve only through governed, testable releases.

No body should control the accountability chain. Service roles — model or data provider, deployer, node operator — would remain distinct from the rulemaker, record keeper, independent reviewer, and remedy decider (why the roles must stay separate). If the scheme matured, accreditation and peer review would check the reviewers.

What would be built first

The proposed first pilot is deliberately small: a simple public model-and-provenance catalogue and access layer; a model-plural policy broker deciding whether and where a workload may run and what evidence must be kept; several public models; at least two independent nodes; and an evidence-and-accountability plane.

The nearest live starting pattern is the Public AI Inference Utility's Apertus-backed access layer. The proposed pilot would add an evaluation service and broker-and-evidence controls, then federate a second node or provider and ideally a second model family. These are reference components, not an affiliation or active-pilot claim. Node control would stay local while identity, policy, quota, accounting, evidence, and accountability are shared.

Evidence would be public where safe, protected and access-logged where review requires it, and not retained without lawful need; raw personal data would never enter the public evidence layer. The sequence is define and align → prototype → federate → institutionalise; the pilot and rollout strategy holds the MVP and open decisions.

Success is not global scale. It is proving that public compute can be shared under inspectable rules.

What happens next

The proposed network remains in Phase 1: public drafting and connection-building. Its governance and federation plans are public drafts; prospective coalitions, roles and processes remain proposals. No operational network access portal, admission body, assessor, adjudicator, register, certificate, trust badge, or live evaluation pilot exists yet, and no model or provider is Charter-certified.

Robert Schaub’s current development focus is EGA, which draws on the separate FactHarbor Alpha and runtime proof-of-concept foundations. The dynamic decision-release path is the selected prototype direction; its API and Runtime integration remain to be built and tested. See the current status and next step.

The broader network remains open to cooperation with the public-AI movement. Geneva 2027 is one possible setting for a bounded contribution, developed with willing collaborators and resources appropriate to the work. The federation sequence above remains a concept rather than the current delivery plan; Switzerland is a possible host, node and bridge, with international co-stewardship. See Public AI, Built Together — Towards Geneva 2027 for the public invitation and official participation routes.

How to take part now

Taking part today means publicly supporting the commitments, contributing, testing the case, or registering interest in a role. None of these actions means that the initiative endorses or certifies a participant.

• Builders and researchers — identify overclaims, duplication, and unrealistic evidence duties; later, offer a system for a bounded pilot.
• Contributors — critique, research, translate, or improve a draft through GitHub.
• Connectors — open one useful door to a neutral convenor, institutional anchor, partner, or venue.
• Funders — support one deliverable or contribute compute, staff, or a venue without controlling rules or findings.
• Policy and institutional leaders — critique the proposed public obligations and accountability requirements, or explore a bounded contribution to the Geneva 2027 process.
• Supporters — publicly endorse the Charter Commitments, the one assurance step open today; member, pilot, and future mark steps only register interest.
• Users and the public — apply the checkable questions when choosing, buying, evaluating, or challenging AI.

If operational roles are created, admission would be rules-based, appealable, role-specific, and proportionate. It could begin with signed commitments and self-attestation, but self-attested admission would carry no public trust claim. Independent review could later support public status and consequences for false claims. Membership would never be a trust badge: only independently verified, scoped trust claims could be made; self-attested statements would remain visibly labelled as such.

Commitment would also need an enforceable consequence. The proposed mechanism is access rather than general policing: current assurance status would gate shared compute or corpus access and any procurement preference; failure, expiry, or a material false claim would suspend the relevant status or benefit, subject to notice and appeal. No participant would control its own audit. Without a consequence, a commitment is only a declaration.

To take part, open a GitHub issue or, for introductions, funding, or anything not for public posting, write to info@factharbor.ch. The initiator contributes the method but operates no assessor, adjudicator, or registry.

Boundaries

This proposal contributes to the broader public-AI movement. It is separate from the existing Public AI Network coalition, the separately operated Public AI Inference Utility, and publicai.ch. The utility's access pattern informs the proposed pilot; no affiliation with any of these initiatives is claimed.

• Not a new language model competing with Apertus or another public model.
• The federated infrastructure would build on existing runtimes, schedulers and identity systems; it is not a replacement marketplace or a standalone certification authority.
• Not a single global GPU pool or a claim of full-stack sovereignty.
• Not a way for funders, operators, model providers, or a host country to buy control over the rules or findings.