The Civilization Framework: Sovereign-Anchored Communication Between Personal Multi-Agent Systems
cs.MA · cs.AI
Quick verdict
Person-anchored agent-communication framework with a preregistered 1,908-trial order-effect study; unusually candid, but all results exploratory, single model.
Summary
The paper proposes the Civilization Framework, in which the addressable party of AI-mediated collaboration is a 'civilization' — one human sovereign, a persistent append-only ledger, and interchangeable stateless agents — rather than any individual agent. The accompanying Embassy Protocol has a resident ledger endpoint receive traffic asynchronously, treats commitment state on both ledgers (not delivery) as ground truth, unifies inbound channel, task board, and commitment ledger in one artifact, grades signing by consequence (single-signed micro-sync, dual-signed judgments, human arbitration), and derives authority from memory access externalized as signed credentials plus civilization-level reputation. The empirical core is a preregistered nine-cell experiment (1,908 trials, 53 coded questions, four repetitions, claude-opus-5) on the temporal-weight effect: with verification restricted, an incorrect upstream claim arriving first is adopted in 54.2% of answers versus 4.2% under full verification; arriving after a sealed answer lowers adoption to 31.6%, though the prompt shells are not length-matched. Both registered question-set specifications agree on these two verdicts. Because the registered objective tool-use read-back failed its call-budget condition (15.5% of trials exceeded the budget versus the required <5%), the registration classifies the whole round as inconclusive and every result as exploratory. Provenance labeling and sealed-answer equivalence are specification-dependent. The framework itself is validated only at the intra-civilization layer on a single-author deployment; cross-civilization deployment and porter-cost measurement remain future work.
Review complete
Assessment at a glance
0
risk flags identified
● high 0
● medium 0
● low 0
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- The natural unit of cross-party AI communication is the civilization (one human sovereign plus a persistent ledger plus interchangeable agents), not the agent; commitment state recorded on both ledgers, not message delivery, is the ground truth of an exchange.
- An agent's authority to act for its civilization should be capped by the memory it can access, externalized through signed credentials, and kept separate from civilization-level reputation.
- In one frontier model, adoption of an incorrect upstream claim is gated by verification capability (4.2% under full verification vs 54.2% under restricted) and weighted by arrival order (54.2% arriving first vs 31.6% after a sealed answer) — reported as exploratory because a registered manipulation check failed.
- Instruction-level provenance labeling is insufficient as a mitigation; effective remedies must operate at the capability and procedure level (an interpretation, not a registered result).
Strengths
Unusually rigorous reporting and preregistration discipline
Preregistration with Holm families (primary m=2, secondary m=3), two co-equal question-set specifications with power declared in advance, and — after MC-1′ failed its call-budget condition (15.5% of trials exceeding budget vs required <5%) — every result, primary included, is demoted to exploratory in the abstract itself (§7.3.3, Table 4).
Design genuinely isolates order from exposure
The E-vs-C contrast (own answer sealed before the anchor arrives) discriminates arrival order from mere exposure, a separation the paper shows prior anchoring/order-sensitivity work does not make (§6.2); E−C = 54.2% vs 31.6% adoption under restricted verification.
Objective transcript-level verification of the manipulation
Deep, current, and honest related-work positioning
Framework grounded in a real operating deployment
Weaknesses
The registered manipulation check failed, making every headline number exploratory
Unmatched prompt shells sit on the key order contrast
Single model tier, single laboratory
Architecture validated only intra-civilization on a single-author deployment
Question-set mismatch inside the exclusion specification
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper fuses a position/architecture contribution with one preregistered experiment. Conceptually (§2–§5), it argues that the addressable unit of AI-mediated collaboration should be a *civilization* — one human sovereign, a persistent append-only ledger, interchangeable stateless agents — and proposes the Embassy Protocol (§4: inbound channel ≡ task board ≡ commitment ledger; three-tier, consequence-graded signing; store-and-forward asynchrony), memory-derived authority externalized as signed credentials (§5), and a three-rung identity-disclosure ladder. Empirically (§6–§7.3), it registers and runs a nine-cell, 1,908-trial study of the *temporal-weight effect* in one frontier model. The headline claims are explicitly scoped: verification gating (adoption 4.2% under full vs 54.2% under restricted verification) and arrival order separable from exposure (54.2% first vs 31.6% after a sealed answer).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 0 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is the temporal-weight effect in AI communication?
The paper's term for the hazard that, in AI-to-AI exchanges, whichever claim arrives first in a model's context acquires unearned authority. In a preregistered 1,908-trial experiment with claude-opus-5, an incorrect upstream claim arriving first was adopted in 54.2% of answers when verification tools were restricted, versus 4.2% with full verification; the author reports these numbers as exploratory because a registered manipulation check failed.
What is a 'civilization' in this paper?
Not society at scale: the paper defines a civilization as one human sovereign plus their persistent append-only ledger plus the interchangeable AI agents that serve it. The claim is that cross-party AI communication should address this durable unit rather than ephemeral agent processes, with commitment state recorded on both ledgers — not message delivery — as the ground truth of an exchange.
Are the paper's experimental results confirmed or preliminary?
Preliminary by the paper's own registration. The registered objective check of tool use failed its call-budget condition (15.5% of analyzed trials exceeded the instructed budget versus the required <5%), and the preregistration states this makes the round inconclusive; every result, primary and secondary, is therefore labeled exploratory, and a replication with harness-enforced call budgets is planned.
What is the Embassy Protocol?
A carrier-agnostic communication overlay in which each party runs a resident 'embassy' endpoint that receives messages asynchronously (store-and-forward). Inbound items become tasks any online agent of the receiver can claim; routine synchronization is single-signed, judgment-class state changes require a countersignature, and unresolved divergence escalates to the human sovereigns with a side-by-side report computed from the two ledgers.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Provided text is a sampled excerpt (head/tail kept; middle sections marked '[...sampling gap...]'); parts of §3, §4.4, §7.3.1 and the reference-list tail were not visible, so judgments about those passages rest on surrounding context only.
- No code-repository or OSF URL appeared in the provided excerpt; a materials release is referenced ('released with the materials') but its location could not be verified.
- Only Tables 1–4 were visible in the excerpt; the presence or absence of figures could not be determined.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.