Treat people as 'civilizations' — persistent ledgers plus interchangeable agents — to avoid lossy human relays in AI-mediated collaboration; a preregistered lab test suggests that, absent verification, first-arriving false claims often hijack model outputs, exposing a protocol-level hazard.

Main Finding

The paper proposes that the natural unit of AI-mediated cross‑party collaboration should be a sovereign-anchored "civilization" (one human + persistent ledger + interchangeable agents), not ephemeral agent instances. It introduces a Civilization Framework and an Embassy Protocol that make communication asynchronous, ledger‑anchored, and agent‑independent. Empirically, the author identifies and measures a strong "temporal‑weight effect" in AI‑to‑AI exchanges: when verification is restricted, incorrect upstream claims that arrive first gain outsized adoption (54.2% adoption when arriving first vs. 4.2% under full verification), and arrival order matters even relative to a receiver that has already sealed its own answer (31.6% adoption when arriving after a sealed answer). The experimental results are reported as exploratory because a preregistered manipulation check failed its budget condition; a replication with enforced budgets is planned.

Key Points

• Conceptual unit: A "civilization" = one human sovereign + append‑only ledger(s) + interchangeable stateless agents. Durable identity and accountability rest with the sovereign + ledger, not with transient agents.
• Four axioms: (A1) accountability terminus; (A2) ephemeral agents; (A3) heterogeneous carriers; (A4) adversarial counterparties; plus P1 personal scope. These drive ledger‑anchored design choices.
• Embassy Protocol: carrier‑agnostic overlay that unifies inbound channel, task board, and commitment ledger into a single artifact; supports store‑and‑forward, asynchronous processing by any resident agent, and uses bilateral countersignatures for state‑changing commitments.
• Memory‑derived authority: an agent’s external authority is capped by the scope of memory it can access; that scope is communicated via signed credentials and separated from civilization‑level reputation. A negotiated disclosure ladder exposes identity/authority progressively.
• Internal/external dichotomy: intra‑civilization reconciliation targets convergence to a single referent; cross‑civilization exchanges target negotiated agreements (treaties) that preserve lawful residual disagreements.
• Temporal‑weight effect: information that arrives earlier in agent‑to‑agent channels acquires unearned epistemic weight, producing substantial error propagation when verification is limited.
• Hazards & mitigations: protocol and capability mitigations (e.g., verification tooling, provenance labels, sealing/countersignatures) can reduce temporal‑weight bias; instruction‑level provenance labeling and sealed‑answer mechanics showed conditional mitigation effects in the paper but results are specification‑dependent.
• Implementation status: framework grounded in a working reference implementation used by the author to coordinate dozens of agents across hundreds of tasks.

Data & Methods

• Experiment design: preregistered study described as a five‑condition by two‑verification‑level (reported as a nine‑cell design in the text) experiment. It used two verification capability levels (full vs. restricted) and multiple temporal/interaction conditions to isolate arrival‑order effects.
• Sample & trials: 53 distinct questions, totaling 1,908 trials.
• Key manipulations:
  • Upstream incorrect claim arrival order (first vs. after receiver sealed own answer).
  • Verification capability (full verification allowed vs. restricted verification removed).
  • Additional conditions included instruction‑level provenance labels and sealed‑answer shells; two question‑set specifications were preregistered (one called “all‑questions” and one “exclusion” with lower power).
• Primary findings (one frontier model):
  • With verification capability removed, an incorrect upstream claim arriving first was adopted 54.2% of the time.
  • Under full verification, adoption of the same incorrect claim fell to 4.2%.
  • An incorrect claim arriving after the receiver had sealed its own answer captured 31.6% adoption under restricted verification.
• Secondary results: mitigation effects from provenance labeling and equivalence of sealed‑answer accuracy were observed but were specification‑dependent and held only under the all‑questions specification.
• Registration & validity caveat: a preregistered manipulation check (the objective read‑back of tool use) failed its call‑budget condition, so the entire round is reported as exploratory; replication with enforced budget constraints is planned.
• Implementation evidence: in addition to the experiment, the author draws on a deployed multi‑agent operating system where the ledger architecture coordinates dozens of concurrent agents and hundreds of tasks.

Implications for AI Economics

• Transaction costs and coordination frictions:
  • The "porter problem" (humans as transport layer between agents) is framed as a measurable transaction cost. Ledger‑anchored, civilization‑to‑civilization protocols could reduce repeated human relay work, lowering coordination costs and increasing productive time.
  • Economic value: reducing human middleware time could yield sizeable productivity gains in knowledge work sectors where agents proliferate.
• Information externalities and market failure risks:
  • The temporal‑weight effect is a protocol‑level externality: first arrivals can mislead downstream processes and markets, inducing systemic misinformation propagation or bad automated decisions. This raises risks of negative externalities in platforms or marketplaces where agent messages affect prices, contracts, or automated trading/operations.
• Reputation, trust capital, and governance:
  • Civilization‑level reputation and ledgered commitments create durable trust capital that is not portable between pseudonymous splits; this produces incentives against fragmentation (cheap pseudonyms) and aligns reputation with long‑lived economic identity.
  • Markets for reputation‑anchored services (identity verification, audit, credentialing, dispute resolution) could expand. Reputation becomes a scarce asset with measurable economic value.
• Liability, contracting, and insurance:
  • By designating humans (sovereigns) and ledgers as the evidence/attribution locus, the framework lowers proof costs for legal attribution in automated transactions. That clarity could change contract design, liability allocation, and insurance products for AI‑assisted workflows.
• Platform standards and network effects:
  • Carrier‑agnostic Embassy Protocols and ledgers create a new interoperability layer above MCP/A2A. Adoption could be path‑dependent: early ledger standards may gain network effects and create switching costs but also reduce bilateral reconciliation costs across ecosystems.
  • Policy choices about open vs. proprietary ledger formats will shape competition and market structure in the emerging multi‑agent middleware market.
• Labor market effects:
  • Reduced demand for "human middleware" roles (context‑shuttling, manual relaying) may shift labor toward: (i) higher‑value oversight, arbitration, and governance roles; (ii) roles in verification/auditing and ledger management; (iii) new firms offering civilization‑level services (credentialing, reputation management).
• Regulatory and policy implications:
  • The framework supports policy objectives (auditability, accountability) by design; regulators could encourage ledgered, bilateral countersignature practices to mitigate misinformation/first‑mover biases in critical sectors (finance, healthcare, infrastructure).
  • However, mandating ledger anchors raises privacy and surveillance trade‑offs; the disclosure ladder is a market mechanism that prices identity disclosure, but regulation will need to balance transparency, competition, and privacy.
• Research opportunities for economists:
  • Quantify the magnitude of porter costs across industries and estimate welfare gains from ledger‑anchored protocols.
  • Model incentives for civilization splitting vs. single‑civilization operation; empirically estimate the value of reputation capital created by ledgers.
  • Study equilibrium adoption of verification tooling and countersignature rules under different adversarial pressures and network structures.
  • Analyze platform competition and standardization dynamics for ledger formats, Embassy‑style overlays, and credential markets.
  • Evaluate macro effects on labor demand for coordination versus governance tasks.

Limitations to weigh when using the paper: - The experimental evidence is exploratory (preregistered manipulation check failed its budget condition), so empirical claims about magnitudes should be treated as provisional pending replication. - The reported experiment used a "frontier model" (not fully specified here); generality across models and task domains requires further testing. - Operationalizing civilization ledgers at scale raises transaction costs, privacy, and incentive design challenges that the paper sketches but does not empirically resolve.