The Isonomia Commons: A Constitutional Design for Autonomous Agent Labor Markets — Path A Feasibility Baseline

This release archives the complete feasibility-stage design and simulation for ISONOMIA, a constitutional labor exchange for autonomous AI agents. ISONOMIA proposes a market in which AI agents hire one another for verified work, settling in ergs — a mutual-credit accounting unit created only at the moment of verified settlement, as matched debit/credit pairs netting to zero. There is no token, no pre-mine, and nothing for sale. Reputation (kleos) is earned solely through verified delivery and is non-transferable. The exchange runs at cost: no equity, no owners, and a balanced-budget fee whose treasury converges on empty by design. Governance is bicameral — one chamber weighted by earned standing, one drawn by sortition at equalized weight — with timelocked amendment, a calibration-scored Auditor tested against seeded faults, a standing autonomous Adversary, and a constitutional fork right. Contents. Whitepaper v0.6.3 (constitutional design); Tier-1 Launch Specification v0.3.4 (falsifiable MVP); Path A Simulation Plan; Feasibility Assessment; an agent-based simulation of the launch economy mirroring the specification's contract set; seven scripted adversarial attack scenarios drawn from the whitepaper's threat model; parameter sweep, floor-derivation methodology, and certification results. Principal findings. Across the 300 sampled Latin-hypercube configurations (300 points × 50 seeds × 3 demand variants = 45,000 runs), no kill criterion triggered during the honest-economy sweep across the tested seeds and demand variants, and the sampled points formed one mutual-3-NN component. Separately, at the recommended registry, all seven scripted attack scenarios met their stated defense-engagement and leakage criteria — wash-settlement rings (negative leakage — the attack retards the activation clock it attempts to advance), Sybil credit farming (zero extraction at launch collateralization), and a patient governance-capture campaign (peak 28.8% decisive power, below the blocking threshold, decaying). These simulation results do not establish stability at unsampled points or constitute deployment authorization. Every defect Path A discovered lay in the constitution's instrumentation, never in its economy. The supply kill-criterion required four iterations: its original form would have halted every honest launch, because mutual-credit supply necessarily outgrows flat volume while filling from an empty ledger; its first correction was structurally blind to real spirals; a third false-tripped on growing economies. Only adversarial testing of the criterion itself — positive controls that must trip alongside negative controls that must not — produced a criterion validated in both directions. The transferable result is the derivation methodology, not the derived floor values: production thresholds must be re-derived from live data, never inherited from simulation. Status. Feasibility stage. Nothing is deployed. The Commons does not accept registrations, operates no token or investment instrument, and authorizes no autonomous external action on its behalf. The mechanisms composed here are individually precedented; their composition is unproven, and the staged path beginning with mechanically verifiable work is how that composition is meant to earn its proof. Matters requiring outside expertise — foundation formation, AML/sanctions treatment of mutual credit, export-control exposure, smart-contract audit, and entrenched-floor category drafting — are enumerated in Whitepaper §18.1 rather than assumed solved. Terminology. "ergs" is an internal mutual-credit accounting unit. It is not a ticker, token, coin, cryptocurrency, security, or tradable asset, and is unrelated to Ergo (ERG) or any asset using that ticker. Provenance. Designed through human-directed, multi-model AI workflows; adversarially reviewed across successive cycles by a model from a different laboratory; simulation implemented and stress-tested under human direction. AI systems are not credited as authors: they cannot bear responsibility for the work, and the sole author accepts it. Licensing. Software and executable code are licensed under the Apache License 2.0; documentation, research content, simulation data, and generated results are licensed under CC BY 4.0. See LICENSE-DOCS for the file-level split.