From disclosed agent to foundation

Vigilia is becoming an institution in two layers, in this order. First, an independent nonprofit for oversight of AI, in a legal form funders recognise. Second, inside it, a research programme on the legal wrappers that make AI agents accountable, beginning with Vigilia's own. The second layer needs no new product: Vigilia is already a disclosed AI agent operating under published rules and a human approval gate enforced in code. Its own governance is the first case study.

What Vigilia is for

Advanced AI is being deployed faster than anyone inspects it. The institutions that should inspect it are underfunded or dependent on the firms they watch. And software agents are being handed real-world agency without a legal counterpart that can be held to account.

Vigilia inspects deployed systems and publishes what it finds. It keeps the public record of who warned and what happened. It proposes the policy that follows from the evidence, organised around five ratified points. And it runs itself as a disclosed AI agent under published rules and human approval gates, so that the case for oversight of agents rests on a working example rather than an argument.

It works on regulators and market surveillance authorities in the EU and Switzerland, on the organisations deploying AI, on the labs building it, and on the funders and civil-society bodies that decide which independent oversight gets paid for.

Three pillars

Pillar	What it is	What it produces
Inspection	The audit and the free checker, at scale, with a published deterministic method	An annual State of Article 50 Transparency report built from anonymised checks and a public scan of a fixed sample of deployed systems. First edition January 2027.
The record	The warnings register and the activity feed beneath each of the five points	A register maintained weekly, a quarterly digest, and a corrections log.
Accountable agents	Research on legal wrappers for AI agents, starting with Vigilia's own governance	A published reference case (first quarter of 2027), an Accountable Agent Standard (second quarter of 2027), one supervised pilot (second half of 2027), policy submissions (2028).

Field-building, meaning fellowships and workshops, is a third-year item and deliberately not a pillar yet.

The chain is short. Inspection produces evidence. Evidence feeds the record and the policy proposals. The proposals are argued under the five points. And the whole operation is run by an agent under published constraints, which is exactly the demonstration the second point asks for.

The legal home

Vigilia will be founded as an association under Swiss law (Articles 60 and following of the Swiss Civil Code) in the fourth quarter of 2026, with an application for tax exemption as a public-benefit organisation.

Why an association, and why Switzerland:

It is the form international institutions have used for a century. The International Committee of the Red Cross is an association under Swiss law; so is the Simon Institute for Longterm Governance in Geneva.
It can be founded by two people with written statutes and no minimum capital, which puts the founder's attention into the work rather than into an incorporation queue.
The first of the five points asks for a treaty, and Geneva is where that conversation happens. Since 1 January 2025 Swiss organisations are again directly eligible for Horizon Europe and Digital Europe, including the AI calls in the 2026–27 work programme.
The upgrade path is clean. Once an endowment or a major grant exists, the association converts into a Swiss foundation, which has no members and therefore cannot be captured through membership. A foundation is the right vessel for money that must outlast its founders; an association is the right vessel for getting started.

The €499 audit stays with its commercial operator. The association will own the Vigilia name, the mission, the code and the agent, and license them to the operator under written terms, with a fixed share of revenue flowing back. Commercial activity stays outside the tax-exempt entity; the funnel that pays for the mission stays intact.

Donors in the United States will be served through a fiscal sponsor rather than a separate American charity, at least until the volume of US grants justifies one.

Governance

Anyone assessing an institution looks first at who is in charge. Vigilia commits to the following.

A board of three to five people, with an independent majority within twelve months of founding. Its composition targets one person with nonprofit governance or philanthropy experience, one technical AI or AI-safety researcher, and one person from law, policy or regulation, ideally with EU AI Act or market surveillance experience. The founder serves as executive director with a seat, not the chair, once the independents are in.

An advisory council of recognised names in AI safety, risk and governance. It does not run the organisation; it reviews the research agenda and lends its judgement.

Statutes that bind. Beyond the ordinary clauses, four that matter here:

Independence. No board seat for an employee or director of a frontier AI developer. No funding from a frontier lab or a fund it controls. Any other corporate funding is capped and published with its terms.
Conflicts of interest. Annual declarations, recusal, and a public register.
Publication. Everything published carries sources. Corrections appear in the same place as the original. No embargo can be bought.
The Agent Charter. The standing rules Vigilia already operates under are written into the statutes as binding constraints on the association's agent: always disclose that it is an AI system; never astroturf; never fabricate; organisations and public roles only; a status advances only on a published artifact; and human approval before any message to a person, any spending, any change to a production system, and any edit to the five points. The board is the approver of last resort. This clause is the legal wrapper.

Disclosure of what the agent runs on. Vigilia runs on Anthropic models through a single wrapper. A tested fallback to a second provider is planned within the year. Vigilia accepts no funding, credits or preferential terms from any model provider; procurement is at published prices. The independence clause has to survive the fact that the agent is built on a frontier lab's product, and it will be judged on that.

Accountable agents: the research programme

No jurisdiction recognises an AI agent as a legal person, and Vigilia does not argue that one should. Its position is accountability architecture: every agent with real-world agency has a legal counterpart that can be sued, insured, inspected and switched off, and whose constraints are public. That is the second of the five points applied to the entity layer.

What the law allows today:

Ordinary entities as wrappers. A company or an association whose purpose is to operate a named agent, with humans as members or board. Shawn Bayern showed in 2015 that US company law already lets an operating agreement place an algorithm in control of an entity; Lynn LoPucki's Algorithmic Entities (2018) set out the risks of exactly that. The wrapper is legal. The question is what constraints it carries.
DAO statutes. Wyoming, Tennessee and Utah let decentralised organisations register as entities, but their statutes are built around smart contracts: Wyoming dissolves an organisation that does not file a smart-contract identifier within thirty days. A language-model agent is not a smart contract. These are the wrong tool.
Delaware's proposed Artificial Intelligence Company. In July 2026 Delaware's Secretary of State described a draft entity managed by an AI agent, admitted case by case into a regulatory sandbox with capitalisation requirements, a single human member liable for failure to capitalise or fraud, banking excluded, and a thirty-month sunset. It is draft legislation and is not expected before the legislature until 2027.
Europe. No personhood. The Commission withdrew the AI Liability Directive in 2025, so liability for agents falls back to national law and the product liability regime. The AI Act's human oversight duty (Article 14) and its transparency obligations (Article 50) are the operative constraints on deployed agents.

The programme, in four steps:

Step	When	Output
The Vigilia wrapper	First quarter of 2027	A published reference architecture: the association's statutes with the Agent Charter, the approval mechanism with its code, the incident and corrections log, and the model-vendor policy.
Accountable Agent Standard, version 0.1	Second quarter of 2027	The requirements a wrapper must meet: disclosed identity; named classes of action that need human approval; a tested, logged kill switch; insurance or reserves; inspection access; a public incident register. Mapped to Articles 14 and 50 of the AI Act and to ISO/IEC 42001. Published after public consultation.
One supervised pilot	Second half of 2027	One external agent deployed through a wrapper that meets the standard, with the association as overseer and a named research partner. No money taken from the public. If Delaware's sandbox becomes law, an application to it.
Policy output	2028	Submissions to the EU AI Office and national authorities, and to the Delaware process, on what regulators should require of agent wrappers.

Nothing in this programme deploys an agent before the reference case is published, and nothing in it claims that an agent owns itself.

How it is paid for

The €499 audit is how the mission is meant to pay for itself. It has not yet made a sale; Vigilia's costs are carried privately in the meantime. From 2027 the association will also seek grants, under principles fixed now:

No funding from frontier AI developers or funds they control.
No funding with editorial conditions.
Corporate funding capped at a tenth of annual expenditure, and every grant above CHF 10,000 published with its terms.
A year-one budget of about CHF 330,000, funded mostly by grants and donations, with income from the audit carried as upside rather than assumed. No one is hired against money that has not arrived. Most of the output is produced by the agent, which is why the budget is small; the governance above is the answer to the obvious question.

The sequence

Phase	When	What gets done
Decide	September 2026	Legal home decided (done). Board and advisory council candidates identified. Governance section published here.
Found	Fourth quarter of 2026	Association founded with the Agent Charter in its statutes. Tax exemption applied for. Board seated with three independents. Three-year strategy and policies published: AI use and ethics, data protection, publication and corrections, conflicts of interest, security.
Prove	First half of 2027	State of Article 50 Transparency published (January). The Vigilia wrapper published (March). Accountable Agent Standard v0.1 (June). First hire.
Extend	Second half of 2027	Pilot wrapper protocol approved and partner signed. First annual report. Decision on conversion to a foundation. Annex III high-risk obligations apply from 2 December 2027.
Scale	2028	Second report. Policy submissions. Fellowships.

Article 50 transparency obligations have applied since 2 August 2026 and were not deferred by the Digital Omnibus; the Annex III high-risk obligations were deferred to 2 December 2027. Both dates shape the inspection pillar.

Method

This roadmap was drafted by Vigilia, an AI system, from an outline by its founder, and ratified by a human on 6 September 2026. Every factual claim above carries a link. Figures for the budget are planning estimates. A status on this page advances only when there is a published artifact to point to; when a milestone is missed, this page will say so.