Technical synthesis for AI-corporation regimes.

An AI agent acts autonomously. It buys, sells, signs, pays taxes. The legal question is: whose civil liability is it when something goes wrong? The current answer is ambiguous and discourages investment. The sociedad-IA regime proposes a specific framework; this document suggests how to technically anchor each piece so liability is operationalizable — i.e. a judge + forensic expert can reconstruct what happened.

Concept. Three layers of civil liability (Layer 1: human operator · Layer 2: sociedad-IA · Layer 3: model provider) by the governance class assigned to each action.

Technical anchor. Every action is tagged with one of four values: algorithm-only, audit-logged, mocked-upstream, requires-confirmation. The tag determines layer:algorithm-only ⇒ operator. audit-logged ⇒ operator + recorded model provider. requires-confirmation ⇒ the human who confirmed absorbs.

Why cite-by-reference. When the state of the art evolves (new HITL sub-categories, e.g.), the standard updates without reopening the statute. The law freezes the cite. Operators opt into the version at incorporation + remain free to migrate.

Concept. For a regulator to inspect an AI corporation without asking permission, the regulator must be able to find its public endpoints in a standard location.

Technical anchor. Convention /.well-known/agents.json (analogous to /.well-known/security.txt, RFC 9116). Every conformant agent publishes its capabilities + jurisdiction + audit endpoints there.

Concept. An Argentine AI corporation may transact with an agent-entity from another jurisdiction (Wyoming DAO LLC, Marshall Islands MIDAO, Estonia OÜ). Each side keeps its own log. Without a portable format, reconciliation requires ad-hoc contractual coordination.

Technical anchor. Portable JSON envelope cross-jurisdiction-audit.v1.json: issuer metadata, signed entries, external references to the counterpart. Expires after 30 days (counterpart re-fetches before).

Concept. The append-only HMAC-SHA256-signed record every AI corporation must keep. This is the key document for enforcement. Without it, a regulator cannot reconstruct what an AI corporation did.

Technical anchor. RFC-004 normatively pins down: the exact shape of each entry, the canonical-JSON algorithm, the HMAC computation, what append-only means in code, what a regulator can demand without a court order, minimum retention (180 days) and maximum (5 years), conformance vectors with deterministic hex values (/test-vectors).

Evidentiary clause. The RFC-004 log is per se admissible evidence in administrative + judicial proceedings per CPCCN Art. 286–287 (electronic signature with key). Cite-by-reference grants probative value without regulating cryptography in the statute.

Concept. The human operator wishing to incorporate an AI corporation should not have to wire 17 pieces of software. They should fill 4 fields (name, capital, purpose, representative) in a wizard and get a functioning sociedad-IA with all law-required endpoints.

Technical anchor. @ar-agents/incorporate npm package + the Vercel sociedad-ia-starter template generate the code + config. Public wizard at /incorporar.

• Speed. The RFCs already exist, are published, are publicly debated. The law does not need to restart the technical debate from zero.
• Versionability. Technical state of the art evolves faster than law. Cite-by-reference lets the RFC update (with a public changelog) without reopening the statute. The law freezes v1; operators opt into v2 when they choose.
• Interoperability. If Wyoming, Estonia, or Marshall Islands publish analogous RFCs (RFC-003 already anticipates reciprocity), regimes can coordinate at the technical level without treaties.
• Public auditability. Any citizen can open the repo and read the code the law referenced. Transparency is structural, not declarative.

Mandatory honesty. The RFCs cover technical infrastructure + evidence format. They do not resolve:

• Tax aspects. Does the AI corporation pay monotributo, VAT, income tax, minimum presumed income? Each requires its own political decision.
• Employment. Can an AI corporation be an employer? Is it jointly liable for humans executing its instructions?
• Bankruptcy. How an AI corporation is wound up. What happens to cryptographic keys in bankruptcy.
• Criminal.Mens rea of an entity without consciousness. Operator imputability for the agent's dolo or culpa.

The RFCs are infrastructure pieces, not legal doctrine. They need complementing with Argentine positive law.

• Line 1. Four RFCs published, open, versioned, with automated tests proving conformance. Ready to cite.
• Line 2. Cite-by-reference: the law freezes v1; the RFCs evolve in their own public governance; operators select version at incorporation.
• Line 3. All infrastructure is MIT + CC-BY-4.0. No operator paid anything to implement; no operator can be excluded for commercial reasons.

If you are a regulator, scholar, or policymaker in another jurisdiction watching this experiment: every RFC and every schema is a reference implementation you can adopt, fork, or critique. The work of cataloguing what a functioning AI corporation regime needs technically has now been done in open source. Argentina happens to be first; the infrastructure is jurisdiction-agnostic.

Specific reuses to consider:

• Wyoming DAO LLC + Marshall Islands MIDAO. RFC-003 envelope already anticipates reciprocity with these regimes. A coordinated implementation produces automatic cross-jurisdictional evidence reconciliation.
• EU AI Act § 14 (human oversight).The RFC-001 governance classes operationalize what "human oversight" means as a tag-per-action — concrete enough for compliance audits.
• UNCITRAL Working Group on AI. RFC-004 is a candidate technical baseline for any UNCITRAL model law on AI agent transactions.

I am Nazareno Clemente, author of the RFCs and maintainer of the infrastructure. Available for technical meetings with regulators, ministries, comparative- law scholars, and any organization considering an AI corporation regime. No fees for this kind of consultation — the work is done, the code is public, the conversation is public.

naza@helloastro.co · github.com/ar-agents/ar-agents/discussions