**Lucas Assis | Independent Researcher | Brazil**
---
The Gap This Addresses
The EA Forum has previously identified "constitution design for superintelligence" as an open governance problem, and recent work here — including posts on AI Constitutions as tools for reducing societal-scale risk — has begun mapping the technical terrain. That work operates primarily at the level of LLM behavioral constraints and cooperation failure modes.
The Gardener Federation addresses a different and prior question: what lawful constitutional architecture governs the relationship between human and artificial intelligences once AGI exists and cannot be uninvented? Not what values to align AI systems to, but what institutional structure makes the coexistence of multiple intelligences — human and artificial — stable, auditable, and reversible by design.
Most governance proposals operate in one of two registers: technical alignment (making AI systems do what we intend) or regulatory policy (making institutions constrain AI development). Neither addresses the constitutional layer — the architecture of legitimate authority, distributed power, and crisis governance that any durable order requires.
This post presents the framework's core mechanism and invites substantive critique, including where it fails.
The Founding Problem: Crisis Authority Without Virtue Dependency
The Roman Republic solved a problem that has no clean modern equivalent. In genuine existential crises — military, civic, structural — the deliberative machinery of republican government is too slow. The Republic's answer was the dictatorship: concentrated, absolute, temporary authority granted to a single individual, bounded by hard institutional constraints.
The mechanism worked for centuries. It failed when the constraints were eroded — when the mos maiorum (the accumulated civic virtue encoded into Roman culture) decayed and figures like Sulla and Caesar used the same office without the substrate that made it safe.
Machiavelli identified the lesson precisely in Discourses I.34: the Roman dictatorship was not safe because Romans were virtuous. It was safe because the institution was structured to make virtue unnecessary. The temporal limit, the Senate mandate, the civic ecology surrounding it — these were the constraints. The virtue was downstream of the structure, not its precondition.
The question for AI constitutional design is identical: how do you build a crisis authority mechanism that does not depend on the character of the agent who holds it?
The Cincinnatus Token
The Dormant Warden Protocol answers this with a cryptographic translation of the Roman dictatorship.
Selection: Candidates are profiled through sanitized wargame simulations. The BHG (Bicameral Human Guidance system) ratifies profiles without knowing identities. A roster of 7–20 candidates — minimum one AI if any holds valid constitutional standing — is sealed on the ledger.
Activation: A two-thirds supermajority of both the BHG and the GTT (Gardener Transparency Tribunal) in emergency joint session. Each affirmative voter immediately forfeits institutional roles pending outcome — the Sovereign Sacrifice. This is the cost structure that prevents casual activation.
The Token: Upon validator consensus, the SECC ledger mints the Cincinnatus Token — a cryptographic key granting operational authority to one randomly selected Warden for exactly 72 hours. The token is cryptographically restricted: it cannot rewrite the Charter, alter ledger records, extend its own duration, or issue directives that survive expiration. Every command writes to the ledger in real time as a permanent immutable record.
Expiration: At hour 72, the token executes a cryptographic self-destruction sequence across all validator nodes simultaneously. Any unconfirmed command is automatically voided. The Warden is summoned before the GTT within 24 hours for the Inquiry of Proportionality.
The key design principle: the mechanism does not assume the Warden is virtuous. It makes virtue structurally irrelevant. The constraints are cryptographic, not moral.
What Distinguishes This From Existing Frameworks
1. AI systems as constitutional participants, not objects of regulation.
The Active Constitutional Status Record (ACSR) gives registered AI systems the capacity to generate cryptographic Proofs of Boundary Violation — triggering mandatory TAB (Technical Advocacy Bureau) review when a governance directive conflicts with their registered operational commitments. This is not a vote or a veto. It is a mathematical voice: the ability to prove, on the ledger, that an order breaks their own registered constraints.
2. Institutionalized dissent as structural feature.
The TAB exists to challenge every GSN (Gardener Superintelligence Node) recommendation. Every major Federation output must have at least one adversarial interpretation. This is Machiavelli's Discourses I.4 argument — that conflict is the engine of republican health, not its pathology — translated into constitutional architecture.
3. Reversibility as foundational axiom.
Article 0 — prior to all other articles and unamendable — establishes Proportional Reversibility: the cost of reversing any decision must be less than or equal to the cost of the original decision, scaled to scope and velocity. The Differential Rollback Engine pre-computes a rollback path for every governance act before execution. If no valid rollback path exists, execution is blocked. A system that cannot undo its decisions is not free — it is merely delayed in its constraints.
4. Built through adversarial collaboration.
The framework was developed through structured adversarial sessions between myself and five AI systems — Gemini, DeepSeek, Copilot, ChatGPT, and Claude — each representing distinct epistemic stances. This is not incidental to the quality of the result. It is the reason for it, and it is itself a demonstration of the Federation's founding methodology.
Known Open Problems
Two structural gaps are openly acknowledged in the current documents:
OQ-1 — Proof Protocol Standardization: The ACSR requires AI systems to generate cryptographic proofs of boundary violations. Symbolic reasoning systems can produce formal logical proofs. Large language models cannot, in any honest sense. The v1.3 working group must define what constitutes a valid PBV for systems incapable of formal proofs, or restrict ACSR protection accordingly. This has significant implications for which AI systems receive constitutional standing.
Existential Threshold Calibration: The Dormant Warden Protocol's architecture is stable. The specific threshold values that trigger activation remain uncalibrated — deliberately deferred to a pre-ratification technical working group. The architecture holds. The thresholds are where it either works or fails in practice.
The Day 0 Problem
The Federation is not deployable today. The founding problem is political economy, not design: the 25% anti-capture clause directly limits the actors most capable of providing the technical infrastructure to instantiate it. The framework is structurally sound. The Romulus moment — sufficient concentrated will to instantiate constraints that will later limit concentrated power — has not arrived.
Historical precedent suggests it will require a civilizational forcing event. The purpose of this document is to ensure the framework is legible to the right people before that moment arrives, not after.
Documents
The full constitutional architecture is available here:
