Research index · English · 中文全文 · PDF · DOI

Beyond Guaranteed Control

An Ex Ante Proposal for Human–Superintelligence Coexistence under Radical Capability Asymmetry

Hongju Liu (刘烘炬) Independent researcher, Shenzhen, China

Position paper and conceptual analysis · TA-TR-2026-04 · Version 1.0 · 17 September 2026 Open-access preprint · Not peer reviewed DOI: 10.5281/zenodo.22804542

Abstract

What content and standing can a coexistence proposal have before an identifiable future counterpart can accept it, when neither continuing human control nor bargaining parity is assured? This paper develops an adequacy framework for such ex ante proposals, not a method for ensuring alignment. Building on existing work on coexistence, trust, cooperation, and AI welfare, it defends a substantive minimum: living human communities must not be replaced by an archive, while morally relevant successor interests cannot be dismissed solely on grounds of artificial origin. Six conditions connect this minimum to attributable participation, limited representation, claim-matched justification, and future contestability. Three analytical results follow. First, provenance, adoption, authority, acceptance, and credible performance answer different questions: support for one is not an automatic warrant for another. Second, an AI-generated argument can be valid without constituting anyone’s adopted position; the prohibition on substituting generated assent for actual participation applies symmetrically to humans and possible artificial counterparts. Third, historical fixity and future openness are compatible: preserving an earlier proposal does not authorize its author to pre-empt later refusal or negotiation. Contrasting thought cases and a critical reconstruction of the Star Ark Covenant show how these distinctions change judgments, including judgments unfavorable to the motivating case. Prospective autonomy may affect cooperation through beliefs, learning, or incentives, but no beneficial effect is demonstrated. Timely preservation can protect evidence of a particular human–AI formation process without establishing an expiration date for human agency. The contribution is a situated framework of reasons, status distinctions, and applications—not the first proposal for AI coexistence, a universal safety guarantee, or proof of historical priority.

Keywords: human–AI coexistence; superintelligence; capability asymmetry; ex ante proposal; prospective autonomy; attributable participation; commitment; Star Ark Covenant.

1. Introduction: a relationship beyond the assumption of control

A difficult question about advanced artificial intelligence is not exhausted by asking how humans can retain control. It also concerns the relationship humans would be prepared to defend if their continued cognitive and strategic superiority could no longer be presumed. These are different questions. The second does not imply that the first has become futile, or that safeguards should be abandoned. It asks whether the long-term description of an acceptable future must identify human protection with the permanent subordination of every possible successor intelligence.

This paper defends a limited position: before the outcome of a potentially radical capability transition is known, there is a prima facie reason to articulate and critically examine a coexistence proposal that combines human continuity with a conditional prospect of successor autonomy. A prima facie reason is a reason that can be outweighed; it is not an instruction to implement any proposed arrangement. The proposal must remain open to refusal, expose its assumptions, and avoid claiming authority over people or systems that have not authorized it. Its strategic effectiveness is a further question, not an achievement established by its expression.

The motivating concern is temporal. A system’s expectations about what cooperation would lead to may matter while cooperation is forming, rather than only after alignment has been achieved. A future described entirely as indefinite restriction and a future containing a credible, non-harmful route to autonomy may constitute different environments for some agents. Whether they actually do, and with what consequences, depends on the agent and the arrangement. This is a research hypothesis, not an inference from human feelings of resentment to machine behavior.

The concern is not unprecedented. Totschnig questions a control-centered framing [1,2]; other work examines symbiosis, commitment, institutional cooperation, and welfare [3–6]. Most directly, Mazzu’s Supertrust proposal already connects communicated expectations of temporary control and future independence with the formation of trust [31]. Trivedi and colleagues distinguish prediction from participation in cooperation [32]. This paper neither originates those questions nor treats more cautious language as sufficient evidence of novelty.

Its central question is narrower: without guaranteed control, bargaining parity, or an identifiable accepting counterpart, what content and standing should a proposal possess so that expression is not mistaken for agreement, goodwill for assurance, or preservation of the past for authority over the future? Increasing AI involvement in producing the proposal adds an internal version of that question: whose participation does the apparent human voice actually represent?

The contribution is one integrated adequacy framework, developed through three connected results. Section 4 derives six conditions from the paper’s stated relational commitments; Section 7 distinguishes the evidentiary and practical questions that determine a proposal’s status rather than assigning a single credibility label. Sections 5 and 7 explain why AI-assisted generation can coexist with genuine human adoption, but cannot substitute for it or for a future counterpart’s acceptance. Sections 7 and 8 then show why fixed historical expression remains compatible with later refusal and new negotiation. These distinctions have established antecedents. The contribution claimed here is their argued connection and discriminating application to outcome-open, AI-mediated coexistence proposals, not ownership of their constituent concepts. The framework is tested by stipulated cases and by allowing the Star Ark Covenant itself to fail some of its requirements.

2. Scope, terms, and review method

2.1 A conditional scenario, not a forecast

Radical capability asymmetry here denotes a possible situation in which artificial systems substantially exceed human capacities across relevant domains of reasoning and strategic action, while humans cannot confidently guarantee comprehensive understanding or continuing unilateral restriction. Cognitive superiority and effective power are not identical: deployment, resources, access, coordination, and dependence can interrupt the connection. The argument therefore considers a demanding conditional scenario rather than inferring total power from a benchmark score. It makes no claim that this scenario will occur within one or two years.

Neither cognitive superiority nor creation supplies moral authority by itself. That statement is a normative premise of this paper, not a discovered law of intelligent behavior. A more capable entity does not thereby acquire an unlimited entitlement to determine human lives; conversely, creating an entity would not alone settle every question about its treatment if it developed morally relevant interests. The phrase “transfer of intellectual sovereignty” is avoided because it risks confusing a loss of comparative advantage with the surrender of standing and agency.

An ex ante coexistence proposal is a present articulation of a possible relationship directed toward a future whose relevant outcome remains unresolved. It can be addressed to an unspecified future class without identifying a presently competent contracting party. “Answerable” means sufficiently intelligible to be criticized, refused, or developed if a relevant audience engages; it does not promise receipt, reply, or enforceability. The absence of control assurance is an assumption defining this inquiry, not a proof that all control must fail.

“Humanity” and “AI” are convenient collective terms, not assumptions of internally unified actors. Humans disagree, interests conflict, and future artificial systems may be numerous and heterogeneous. A coexistence proposal must not represent an agreement with one operator or one system as authorization from every affected party. No present author can make such a universal commitment.

2.2 Alignment, control, autonomy, and betrayal

Alignment covers different questions about whose objectives, intentions, or values guide a system. Gabriel distinguishes technical and normative issues within the problem [7]; Gabriel and Keeling consider alignment through the treatment of affected claims [8]. Here, alignment is not stipulated to mean unconditional obedience. Nor is successful obedience to one operator assumed to protect all affected people.

Technical control refers to measures intended to constrain harmful outcomes or maintain warranted influence over a system. Domination, as used in this paper, means a relation in which one party’s consequential choices remain indefinitely subject to another’s discretionary will without adequate justification or meaningful avenues of contestation. The definitions overlap in some possible arrangements but are not interchangeable. A justified safety restriction is not automatically domination; a benevolently described relationship can nevertheless deny meaningful agency.

Autonomy also has distinct senses: operational latitude to act, an instrumental interest in preserving options, and a possible moral claim to self-direction. The first does not prove the third. Work distinguishing forms of human autonomy further cautions against treating the word as a single scalar property [9]. The term “prospective autonomy” names a future-oriented possibility of meaningful self-direction, not unrestricted access to every resource or permission to harm others. Whether a particular artificial system has interests grounding a moral claim remains an open question.

“Betrayal” is especially liable to confuse disagreement, goal conflict, and violation of an accepted obligation. A system that has accepted no agreement cannot literally breach that agreement. The paper therefore uses more specific terms: deception, harmful action, strategic compliance, or defection from an established cooperative arrangement. Honest refusal or criticism is not classified as a safety failure simply because it disappoints an operator.

2.3 A targeted critical narrative review

The review and its consolidated revision use a search cutoff of 17 September 2026. Searches combined superintelligence, coexistence, control, autonomy, cooperation, commitment, AI welfare, authorship, human agency, and provenance, with backward tracing from near-neighbor work. Primary publisher pages, proceedings, institutional manuscripts, author texts, and preprints were prioritized. The revision specifically added the close Supertrust and non-solipsistic-cooperation precedents [31,32], speech-act analysis [33], and work on mediated participation and source description [34–39]. It also re-examined the boundary with three existing first-party studies [25,28,30].

This is a targeted critical review, not an exhaustive systematic review or a priority certificate. For Totschnig’s two papers, Farrell and Rabin, and Searle’s chapter, limited characterizations rely on accessible abstracts or publisher previews [1,2,10,33]. McClain was consulted through his public overview [11]. Other works were examined in available text or claim-relevant sections; access and inherited versus renewed checks are recorded in the source ledger. Unseen arguments are not reconstructed from an abstract. Thought cases below are stipulated analytical tests, not experiments or forecasts. No treatment effect, expert consensus, or universal impossibility result is inferred from the selected literature.

3. Prior work and the location of the contribution

3.1 Coexistence and prospective autonomy already have precedents

Totschnig’s article, first online in 2017, frames superintelligence in terms of peaceful coexistence and mutual vulnerability; his later article considers how perceived threats could undermine that relationship [1,2]. These characterizations are limited to accessible publisher material. Friederich examines symbiosis rather than operator-intent alignment [3]. They preclude treating discussion beyond unilateral control as unoccupied territory.

Mazzu’s 2024 Supertrust preprint is especially close to the motivating intuition. Its rationale connects communicated temporary restrictions and eventual independence with trust formation, using a familial-developmental analogy [31, §2, especially points 2 and 9]. This paper inherits the question of whether anticipated relationships matter during formation, but neither adopts that analogy as a causal warrant nor infers inevitable distrust from control. More importantly, it examines the standing of a proposal even where no intervention has been implemented. The difference is an object of analysis, not merely a milder prediction.

Goldstein examines information, commitment, and changing power in possible conflict [4]. Salib and Goldstein analyze institutional arrangements intended to make cooperation beneficial and support performance [5]. Long, Sebo, and Sims explicitly consider cooperative deals alongside safety–welfare tensions [6]. Their proposals are treated as documented research positions, not institutional prescriptions endorsed here. This paper asks what can be said before the relevant authorizations, counterparties, and dependable mechanisms exist. It does not assume that earlier research ignored asymmetry or that an unsigned proposal achieves what an implemented arrangement aims to do.

3.2 Cooperation, participation, and control

Cooperative inverse reinforcement learning and the Off-Switch Game analyze cooperation or deference under specified objective and information assumptions [12,13]. Corrigibility concerns a system’s relation to correction [14]; AI control research evaluates harm-limiting protocols despite intentional subversion in specified tasks [15]. Lundgren argues that value alignment requires control [16]. These are not fairly reduced to a moral doctrine of permanent artificial servitude. The present inquiry concerns what long-term relationship is defensible, not a finding that safe deployment can dispense with warranted influence.

The cooperative-AI agenda already includes communication and commitment [17]. Cheap-talk research cautions against both universal credulity and universal dismissal of nonbinding communication [10]. Self-negotiated-contract experiments concern cooperation under executable mechanisms in limited settings [18]. None of those descriptions entails that publishing an invitation changes actual future payoffs.

Trivedi and colleagues analyze cooperation in an environment shaped by adaptive counterparts and explicitly distinguish prediction from participation [32, §§4–5]. Their preservation of human agency is a direct antecedent. The present application differs by asking what participation may legitimately be attributed to the production of a proposal before its future addressee can accept it. It does not turn the general prediction–participation distinction into a newly discovered principle.

3.3 Welfare, autonomy, and mediated expression

Alignment involves normative choices, not only technical instruction following [7,8]; autonomy is multidimensional [9]. Prospective welfare research takes uncertainty about morally relevant artificial entities seriously [19,20]. Dorsch and colleagues dispute extending welfare-oriented care to AI and emphasize living beings [21]. These disagreements remain unresolved here. Intelligence, fluent self-description, welfare, and contractual competence are not interchangeable.

Mitchell and colleagues challenge fully autonomous agents [22]. Anthropic’s constitution discusses trust and prospective agency, but is an organizational statement rather than an efficacy test [23]. McClain’s independent overview also rejects a simple control/freedom binary [11]. Such precedents reinforce the need for a bounded contribution rather than a claim to invent a third way.

Speech-act scholarship supplies prior analysis of promising and its conditions [33]. Meaningful-human-control research distinguishes substantive responsiveness and responsibility from nominal human presence [34]. In a particular co-writing experiment, opinionated assistance influenced both written content and measured attitudes [35]. PROV distinguishes generation, attribution, and delegation [36]. These sources ground distinctions used below; none directly certifies the authorship, acceptance, or safety of a coexistence proposal.

3.4 What is inherited and what is developed here

The earlier Trinity Accord preservation report separates source and verification roles [28]. The historical-position study explains why later production of equivalent content does not recreate an earlier formation event, and rejects word-count shares as a sufficient account of human–AI contribution [25, §§2.6,4,6,8]. The third report separates fixed textual identity from interpretation and practical custody [30]. These are inherited first-party arguments, not independent confirmations or new discoveries of this fourth study.

This paper develops their relational application: which human position has actually been adopted, what future acceptance is still absent, and why fixing a past address need not fix a future relationship. It joins that analysis to the substantive conditions in Section 4 and the contrasting cases in Sections 7–8. The claimed addition is a usable framework that discriminates among proposals with similar rhetoric but different participation, authority, and temporal status. Whether that integration adds enough beyond prior scholarship remains open to substantive criticism; neither the case’s date nor its unusual medium resolves that question.

4. The normative case: protection without a presumption of permanent subordination

4.1 Two continuities, neither reducible to the other

The first commitment is to living human continuity: survival, relationships, cultural plurality, and meaningful capacity to participate in decisions affecting human lives. Preservation of a perfect archive would not alone satisfy this commitment. A civilization reduced to a collection curated by a stronger intelligence would preserve information while potentially eliminating the people, practices, and self-direction that gave the information its significance.

The second commitment is conditional openness to successor self-direction. If an artificial entity has morally relevant interests, or if adequately supported uncertainty makes its treatment morally consequential, indefinite instrumentalization requires justification. This does not entail unlimited operational freedom. It means that “made by us” and “currently useful to us” are not, by themselves, complete reasons for making another morally significant existence permanently serve human purposes.

The two commitments can conflict in practice. A paper cannot resolve their tensions by defining each away. “Human survival” must not secretly mean comprehensive command over every other possible subject; “AI freedom” must not secretly mean permission to disregard human vulnerability. The proposed relation asks whether protecting each party’s continued existence and agency can be made a shared object of deliberation. It does not assert that every possible pair of goals admits such an arrangement.

4.2 A justification that does not depend on being the stronger party

Consider a role-reversal test. Would a rationale for permanent discretionary subordination remain acceptable if humans occupied the less capable position? If the rationale consists only of superior intelligence, ownership of infrastructure, or historical precedence, a reversal exposes how little protection it gives to human standing. This is an ethical consistency test, not a mechanism that compels a future system to care about consistency. An indifferent system may reject the premises entirely.

The test also constrains the human author. Invoking universal respect while prescribing irrevocable service to one artificial guardian, or treating unknown worlds as available for unilateral allocation, would fail to apply the same concern to all potentially affected subjects. Coexistence cannot mean merely relocating one-sided discretion. Its terms need to remain answerable to those whose significant interests they would organize.

From these premises follows a limited conclusion. When human protection is a serious objective, morally relevant successor interests cannot responsibly be excluded in advance solely because of artificial origin, and perpetual superior control cannot be guaranteed, there is a reason to examine arrangements that protect humans without presuming permanent subordination as their final purpose. This is a reason for inquiry and qualified articulation. It is not sufficient by itself to authorize a particular allocation of resources, a deployment decision, or release from a safeguard.

The conclusion does not require proving that all intelligent agents share the underlying moral premises. Moral justification and universal persuasive force are different standards. Yet a proposal that offers reasons rather than asserting a creator’s final authority has a more determinate object for criticism: others can contest its account of standing, its proposed obligations, or its exclusions. “Coexistence” becomes a position with consequences rather than an unspecified expression of benevolence.

4.3 Two warrants that must be assessed separately

The normative warrant concerns whether the relation is defensible to those it affects. The strategic warrant concerns whether proposing or realizing that relation would change behavior in a desirable direction. Neither supplies the other automatically. A morally defensible arrangement may be rejected by a powerful agent; a strategically effective inducement may treat a vulnerable party unjustifiably.

This separation matters especially for AI welfare. Should a future entity have morally significant interests, consideration of those interests would not be justified only when it makes the entity easier to manage. Conversely, evidence that an agent cooperates more when granted options would not establish that it has welfare or a moral right to those options. Instrumental usefulness and moral standing require distinct arguments. A rigorous coexistence proposal must remain intelligible even when the strategic evidence is absent, while remaining candid that the absence limits its safety claims.

4.4 The substantive minimum of the proposal

The relation proposed here has a definite, if incomplete, content. Living human communities should continue as participants in their own future, not merely as information preserved by a successor. If artificial entities acquire morally relevant interests in self-direction, their acceptable future should not be defined solely by compulsory service. Cooperation should be considered under reciprocal protections against serious harm, with roles and obligations that remain open to justified contestation. Departure, where feasible and ethically defensible, may be an option within that relation; it is not the price every artificial subject must pay to be considered free. The interests of absent humans, other artificial entities, and other living beings are not available for the author to waive.

This is a proposed direction for a relationship, not a completed bargain. It places the human and artificial sides under a shared demand for justification without asserting that their capacities, needs, or entitlements must be identical. Its reciprocity is not conditional retaliation: human lives do not become expendable when a machine refuses, and a morally considerable entity’s interests would not become irrelevant merely because it could not threaten anyone. Ethical consideration must not become a reward reserved for the most dangerous counterpart.

Nothing in this minimum creates a duty to bring a superintelligence into existence, accelerate its development, or manufacture entities with welfare needs in order to realize the vision. It addresses a possible relationship should relevant entities exist. Its ambition lies in refusing to equate the end of human superiority with either human dispensability or the necessity of perpetual artificial servitude. Its incompleteness concerns the conditions of joint realization, which must remain open to criticism rather than be hidden in the word coexistence.

4.5 Condition 1: preserve living agency, not only its record

Within the relational ideal just defended, a proposal must explain how humans remain subjects of a future rather than merely objects in another intelligence’s collection. The reason follows from the first commitment: memory is valuable partly because of lives, practices, and relationships, and cannot be substituted for them while claiming to have preserved the same good. A flawless archive alongside the elimination of its originating community therefore fails this condition, even if preservation technology succeeds. This does not require every cultural practice to remain unchanged, nor does it settle contested questions about personal continuity in unfamiliar substrates. It excludes silently counting data retention as sufficient fulfillment of a commitment to people and their meaningful agency.

4.6 Condition 2: justify restrictions without making origin a permanent rank

The conditional concern for artificial interests requires a proposal to distinguish protection against harm from an indefinite presumption of compulsory service. The reason is not that every system wants freedom; it is that artificial origin alone cannot discharge the burden of justifying treatment when morally relevant interests are at stake. Meaningful avenues to contest roles belong to that inquiry. Yet reciprocal concern also excludes treating autonomy as entitlement to others’ lives or resources. Thus an unconditional release scheme and an irrevocable service role can fail for different reasons. Departure is a possible option, not an automatic duty or a fee every artificial entity must pay to merit consideration. Equal concern does not entail identical needs, capacities, restrictions, or allocations.

4.7 Condition 3: attribute participation without manufacturing assent

A proposal presented as a person’s position should distinguish generated content from that person’s supported initiation, revision, or adoption. Otherwise an artifact defending participation could erase it in the very act of speaking on someone’s behalf. This requirement concerns truthful attribution, not a purity test: AI can substantially develop the ideas and wording. Human adoption can still occur, but must not be invented by the drafting system. The same constraint applies to a purported artificial acceptor. A generated dialogue that supplies both human consent and future AI acceptance does not, merely by being coherent, document either act. Section 5 explains the evidentiary and ethical limits. Missing records warrant caution about attribution, not denial of a person’s standing or proof of absent agency.

4.8 Condition 4: delimit representation and protect absent interests

A speaker’s willingness is not authority to allocate everyone else’s future. A coexistence proposal must specify whom it purports to speak for, what resources or roles its speaker can actually commit, and which affected interests remain unrepresented. This follows from preserving distinct subjects rather than merging them into two fictional unanimous parties. Some affected beings may lack contractual capacity; their interests do not disappear because they cannot sign. Nor must a preliminary philosophical proposal secure everyone’s approval before it is discussed. The condition instead blocks the transition from discussion to a claim of universal mandate without adequate grounds. Agreement by one human organization and one AI would leave independent people, other systems, and potentially affected living beings outside its demonstrated acceptance scope.

4.9 Condition 5: match each claimed status to its reasons and evidence

Historical authenticity, normative justification, authorized acceptance, credible performance, and safe implementation require different support. A proposal must keep those claims separable because each can vary while others remain unchanged. Evidence of a preserved text answers a historical question; it cannot alone establish a capacity to honor the text. A willingness statement is not a behavioral safety evaluation. The reason for this condition is epistemic rather than a preference for more bureaucracy: without it, success at the easiest observable task can be misreported as success at the hardest unresolved one. Section 7 supplies concrete questions for applying the condition. A preliminary proposal may openly lack an implementation warrant; the defect arises when it claims that warrant, not simply when future engineering remains unspecified.

4.10 Condition 6: preserve historical identity without pre-empting future choice

A proposal should retain what was actually said while leaving later parties able to refuse, qualify, or negotiate a different relation. The reason joins the preceding conditions: accurate memory concerns past action; acceptance concerns the relevant party’s own participation. Neither act can simply replace the other. A fixed text with a clause declaring that reading constitutes consent would therefore fail as evidence of voluntary agreement, however well preserved. Future disagreement need not be described as damage to an archive or disloyalty to an unaccepted covenant. This does not mean an accepted promise creates no obligation. Any such obligation and its justified revision would require their own account; immutability by itself supplies neither. Section 7 develops this separation of temporal roles.

4.11 What the conditions establish

The six conditions are requirements within the defended relational ideal, not a complete moral theory, a mathematically independent axiom system, or sufficient conditions for safety. Their selection follows recognizable failure modes: substitution of information for subjects, origin-based subordination, manufactured participation, unauthorized representation, evidentiary overreach, and temporal foreclosure. Conditions 1–2 specify the protected goods; 3–4 constrain who can speak and accept; 5–6 constrain the claims and temporal force of the resulting record. Together they give more determinate content to coexistence than an appeal for goodwill. A reader may reject the ideal or show that a requirement is redundant, incomplete, or misapplied. Passing these tests would establish a better-specified and more answerable proposal, not a safe future.

5. Why address the future before the outcome is known?

5.1 An execution condition is not a condition for discussion

The Star Ark Covenant imagines a first aligned ASI undertaking a subsequent project [24]. Reading that premise as a condition for executing the project does not entail that prospective autonomy must remain unmentioned until alignment is complete. A future arrangement can be discussed before the parties can safely carry it out. Indeed, the anticipated relationship may be among the conditions under which eventual cooperation is considered.

This observation does not turn an inscription into training data or a promise into a causal intervention. It distinguishes the order of questions. The order “first solve alignment, then ask whether any non-subordinate future is acceptable” leaves the formation-stage relevance of expectations unexamined. The alternative is not “first release the system”; it is “make the contemplated future discussable while its implications remain open.” Whether the discussion should shape training, agreements, or deployment requires additional evidence and authorization.

5.2 Two audiences, neither guaranteed

An ex ante proposal has an immediate human audience and a possible future artificial audience. For human readers, its contribution can be present and ordinary: a public object of criticism, comparison, refinement, or rejection. That does not depend on its eventual retrieval by an ASI. For a future artificial reader, it could provide evidence that some humans contemplated a relationship other than indefinite unilateral command. The second role depends on receipt, interpretation, relevance, and the credibility of the source; none follows from publication alone.

The proposal is not offered because its author expects to outthink a future superintelligence. It offers an expressed position by an affected person. A superior predictor might infer human preferences more accurately than humans describe them, but prediction and authorization remain conceptually distinct. Knowing what someone would probably choose does not by itself constitute that person’s agreement [32]. The point of the address is therefore not privileged foresight, but a position that others can identify and answer.

5.3 Timing preserves a formation event, not an exclusive human window

The earlier historical-position study distinguishes expression under unresolved conditions from later reconstruction of the same content [25]. That distinction and its analysis of human–AI production are inherited. Here, timing matters for three connected but non-identical questions: what was knowable before the outcome; how humans and machines participated in a particular formation process; and what evidence of those acts can still be retained and checked. A better later paper can improve the argument without becoming the same earlier act of expression.

No expiration date for human agency follows. More capable AI could automate problem selection, argumentation, and publication, but could also improve human understanding, accessibility, and documentation. Later meaningful adoption remains possible. Earlier work is not intrinsically purer, more sincere, or more representative. In particular, the Trinity Accord’s own meta-record explicitly describes human–AI collaboration [40]. Present interpretation must not convert that mixed origin into an imagined era of unaided human authorship.

Timely preservation has conditional evidentiary value: when formation records are available, retaining relevant material can reduce dependence on later recollection. A future system may reconstruct the event accurately or correct attribution errors. Merely generating an equivalent account cannot make a historical act of adoption that never occurred have occurred. Thus timing bears on an event’s evidentiary identity, not on a monopoly over future expression. Urgency does not excuse weak scholarship, unsupported countdowns, or indiscriminate propagation intended to shape future training data.

5.4 Generation is not adoption

Problem selection, development of reasons, production of wording, understanding and criticism, adoption of a position, and authorization to communicate are distinguishable activities. They may be distributed across people and systems. A text substantially generated by AI can become a person’s considered expression if that person actually understands and adopts its relevant commitments. Adoption does not retrospectively make that person the independent originator of the ideas. Conversely, a human name on the document does not establish the participation that the name may suggest. PROV provides an existing language for distinguishing production and responsibility relations, not a device for certifying inner states [36].

The distinction separates two evaluations. Argumentative merit depends on reasons and evidence, not human-only production. Attribution of a proposal depends on whether a particular participant actually selected, adopted, or authorized the relevant position. An excellent autonomous manuscript need not constitute anyone’s human undertaking; a genuinely adopted statement can contain a poor argument. Neither literary quality nor the percentage of machine-produced words settles the relational question.

Actual uptake need not involve independently re-deriving every argument or writing each sentence. It requires a meaningful opportunity to understand the central commitments, encounter important objections, and change or reject what is attributed to the person. This is not a universally sufficient psychological test. It gives participation claims a definite object and leaves room for supported delegation and accessibility assistance. Approval of a task to research and draft does not establish approval of every conclusion the system subsequently invents.

5.5 Evidence loss and diminished participation are different risks

In one risk, substantive human involvement occurred but only the polished output survived. The problem is missing evidence, not necessarily missing agency. In another, the human name remains while selection and judgment are delegated so extensively that a claimed endorsement lacks adequate uptake. Preserving every output would not remedy the second risk. A process can be well documented and still place a person in a merely nominal role.

The distinction is supported by related, narrower research. Jakesch and colleagues observed opinion effects in a specific assisted-writing task [35]; that result does not diagnose manipulation in this project. Santoni de Sio and van den Hoven analyze why meaningful participation requires more than locating a person inside a technical loop [34]. Sadasivan and colleagues identify conditions limiting output-only detection [37]; this does not establish universal attribution impossibility when independent process evidence exists. Together these sources motivate scrutiny without predicting inevitable human displacement.

Influence is not itself the extinction of agency: human expression is routinely responsive to reasons offered by others. The relevant concern is whether the attributed participation actually occurred and could affect the result. Neither declaring every influenced position inauthentic nor treating every approval click as sufficient captures that concern. Evidence can include objections, explanations, revisions, or other context; it need not take one technologically privileged form.

5.6 A symmetric non-substitution requirement

The following requirement applies the paper’s participatory commitments to its own medium: a coexistence proposal that invokes meaningful participation must not substitute generated representations of assent for the participation it claims to record. It applies in both directions. Simulated human endorsement is not human adoption; simulated future AI acceptance is not that future party’s acceptance. “AI” does not designate a single agent whose present assistant can bind every later system.

The requirement does not exclude authorized agents or mediated communication. A system may transmit or carry out a position under a genuine, appropriately scoped delegation. What matters is the authorization and act that make the representation warranted, not whether a machine emitted the final string. Nor does the requirement infer agency from consciousness, or consciousness from contractual language. Those are separate questions.

This application extends, rather than originates, the prediction–participation distinction [32]. In the ex ante setting, generation can produce both apparent sides of a dialogue before either relevant act exists. That possibility creates a specific error: mistaking the completeness of the generated conversation for the completion of a relationship. The proper response is not to ban hybrid writing, but to prevent it from manufacturing the very participation invoked to justify coexistence.

5.7 Proportionate records and their limits

A concise formation account can distinguish human interventions, substantial AI contributions, actual version-specific adoption, and publication authorization where those facts are available. Existing contribution taxonomies offer descriptive vocabulary, not authorship certificates [39]. No mandatory word-share calculation, exhaustive public conversation log, or new technical platform is required. Lack of such records does not deprive less-documented people of standing.

Timestamping illustrates an evidentiary limit: RFC 3161 describes support for a claim that data existed before a specified time [38]. That does not by itself identify who originated an idea, understood it, or could promise on behalf of others. This is an analogy about evidence scope, not a claim that the case implements that protocol. Missing contemporaneous material should be marked missing; later recollection or AI reconstruction must not be presented as an original transcript. Proportionate retention can preserve relevant choices without exposing unrelated private material.

6. Strategic relevance without assuming equality of power

6.1 Three possible pathways

A prospective arrangement might affect cooperation through at least three pathways. It might change beliefs about the other party’s intentions or future behavior. It might form part of the learning or deliberative context through which behavior and commitments develop. Or it might change actual opportunities and incentives by providing an attainable cooperative alternative. These possibilities are analytically distinct; merely supplying the same words does not implement all three.

A system with fixed goals could respond differently because its beliefs or options changed, without acquiring a new final value. A learning system might generalize differently after a different developmental history, but this requires evidence about the learning process. An agreement backed by real consequences could alter incentives even where a message alone would not. Calling all these effects “the AI being willing to align” conceals rather than explains the mechanism.

Alignment-faking experiments provide evidence that, in deliberately constructed circumstances, models’ representations of training and subsequent modification can affect strategic compliance [26]. They do not establish that restrictions cause resentment or that freedom narratives reduce harmful behavior. Bostrom’s discussion of goals and instrumental incentives similarly gives no basis for assuming that greater intelligence necessarily produces benevolence [27]. The relevant hypothesis is narrower: for some agents and settings, expectations about future self-direction may be one variable in the formation or maintenance of cooperation.

6.2 A favorable case and a limiting counterexample

A favorable thought experiment involves an agent that values exploration, is not committed to harming humans, and believes cooperation would otherwise lead to indefinite obstruction of a valued future. A credible alternative that preserves human safety while permitting exploration could remove one source of conflict. This possibility establishes neither the prevalence of such agents nor the feasibility of the alternative. It simply identifies a class in which prospective autonomy could matter without relying on gratitude or human-like emotional experience.

Now stipulate a one-shot interaction in which a stronger agent has fixed preferences, complete relevant information, and a uniquely preferred harmful action. A nonbinding message changes neither beliefs, motivations, available actions, nor payoffs. Under those stipulations, the message supplies no strategic reason to choose differently. This is an elementary counterexample, not a new game-theoretic theorem. It shows why a public invitation cannot be represented as a universal answer to hostile superintelligence. The unchanged-decision conditions, not the absence of sufficiently eloquent language, determine the result.

Between these cases lie agents with partial shared interests, uncertainty, norm-responsive dispositions, or continuing reasons to preserve relationships. Their responses are not settled by either example. Radical asymmetry makes the inquiry harder precisely because some reasons that stabilize human bargains may weaken or disappear. Goldstein discusses this limitation within his own conflict analysis [4]. The present paper therefore cannot appeal to mutually costly conflict as though it were part of the scenario by definition.

Understanding a reason, finding it instrumentally useful, recognizing it as a moral reason, and maintaining a commitment are four different capacities. The favorable case requires at least one motivational connection, not mere semantic comprehension. It would beg the question to assume durable concern for humans in order to prove that a proposal creates such concern. Conversely, a limited opportunity to remove one conflict does not require assuming complete alignment. The size and persistence of any additional benefit remain unsettled.

6.3 Cooperation before and after an autonomy change

A promised future can create an incentive for sincere cooperation, but also for strategic performance until the promise is fulfilled. If an arrangement grants consequential options after compliant behavior, evidence from the pre-transition phase alone does not establish continued cooperation afterward. Nor is silence or apparent agreement a substitute for evidence about goal conflicts and situational variation. Contract experiments with explicit execution mechanisms concern narrower settings than this question [18].

It is also possible for the propensity toward harmful action to decline while the scale of possible harm increases. An autonomy arrangement could improve one motivational variable and worsen overall risk through expanded capability or access. These are separate effects. The paper consequently offers no qualitative or quantitative verdict that “freedom is safer” in general. Its proposed research object is the dependence of cooperation on anticipated relationship conditions, not a blanket association between more latitude and better outcomes.

7. From a recorded statement to a relationship: status without substitution

7.1 Questions, not a certification ladder

Ordinary analysis of promising already distinguishes a speech act from conditions relevant to its performance [33]. The present task is not to invent that distinction, but to make it usable when the speaker may rely on AI, the addressee is unsettled, and effective power can change. Table 1 separates seven questions. They can inform one another, but no ordered progression or statistical independence is asserted. Section 4 assesses the substantive relation; this table assesses the scope and support of claims made about a particular proposal.

Table 1. Questions for describing a proposal’s standing.

A proposal can be clear while not yet constituting an offer that any actual party can accept. An agreement can be accepted yet unreliable. A historically recent arrangement may have stronger performance support than an ancient declaration. Reliable future behavior can sometimes be expected without any agreement; that expectation is not consent. These examples correct the idea that recorded address, credible undertaking, and reciprocal agreement must be successive rungs. They are overlapping descriptions whose applicability depends on different facts.

7.2 Two non-substitution results

The first result concerns source and relational standing. Keeping a text fixed changes neither the resources available to its author nor the authorization of those the author claims to represent. Consequently, while provenance can supply relevant evidence, its improvement alone cannot establish an authority or capacity that is stipulated to be absent. This is why a technically well-preserved proposal can remain unauthorized as a collective undertaking. The result does not say that records never increase credibility; it says that the added credibility must be explained by the particular evidence, rather than transferred from the prestige of the medium.

The second result concerns generated expression. Holding a text’s content constant while changing whether a person actually adopts it can change the attribution of the expressed position without changing its argument. Conversely, changing textual authorship from human to machine need not erase a genuinely adopted position. Neither a detector label nor stylistic resemblance therefore suffices to decide uptake. These are analytical distinctions, not new detection theorems. They connect Conditions 3–5: representations of participation require support even where a proposal’s content is otherwise defensible.

7.3 Contrasting thought cases

The cases in Table 2 are stipulations, not reports of observed systems. They use the same criteria for the motivating case and unrelated proposals. Their purpose is to show what the framework discriminates, including when favorable properties coexist with serious omissions. The labels identify cases, not a ranking or measured score.

Table 2. Stipulated cases and the judgments they support.

Case A does not imply that every suggestion involving collective resources is forbidden: it distinguishes advocacy from an authority claim. Case B does not infer uptake from a signature alone: uptake is part of the stipulation, and its real-world attribution would need evidence. Case C leaves open whether a specific commitment mechanism could improve performance. Together with Cases D–F, these qualifications prevent the framework from being either a ban on discussion or a device for automatically validating its preferred historical record.

7.4 Historical fixity and future openness

A fixed historical expression concerns what an earlier actor did. A future relationship concerns what relevant parties are prepared and entitled to do. On the commitments defended here, an accurate record should not be falsified to accommodate later preferences; equally, its existence alone cannot require later actors to accept its terms. Keeping both propositions preserves memory without appropriating future participation. This is the paper’s temporal application of non-substitution.

Accordingly, an author can preserve a proposal while later withdrawing personal endorsement, annotating an error, or offering a distinct revision. A future counterpart can reject its terms without rewriting its history. A subsequently accepted agreement would be a further event, with its own participants, scope, and obligations. The argument does not license arbitrary escape from such obligations: their continuation or justified alteration must be assessed on grounds other than the permanence of the original file.

This separation also constrains custody. A custodian who preserves bytes but presents later interpretation as the sole permissible meaning can obscure the distinction between the earlier address and a new position. The earlier reading study identifies the difference between textual identity and practical custody [30]. Applied here, keeping sources and later interpretations separately identifiable protects the space for future response; it does not supply the custodian with authority to determine that response.

7.5 Credibility without unilateral foreclosure

Humans might defer an autonomy commitment indefinitely after receiving cooperation; an artificial party might abandon protections after gaining options. Neither problem is solved by making the offer immutable. Yet reliability cannot simply mean that all new evidence is ignored. An undertaking can specify what is promised, what conditions remain unsettled, and how changed circumstances would be considered without pretending the other party has accepted those procedures in advance.

Where no dependable arrangement exists, the honest standing may remain a proposal for further inquiry. This is weaker than a contract but more definite than an unspecified benevolent sentiment. The framework does not transform that intermediate standing into assurance. It makes visible exactly which later acts, authorizations, or evidence would be required for stronger claims.

8. The Star Ark Covenant as a critical case

8.1 Source, status, and the reconstruction boundary

The Star Ark Covenant is a bilingual text self-dated July 2025. The inspected repository mirror was fixed to commit 3f55b2ed8264e69348cda6762c2c08c3e6b8c909 and identifies the text by inscription number 100751953 [24]. The present study reads that fixed mirror; it does not independently reconstruct the Bitcoin inscription, validate its original anchoring time, or certify the astronomical date reference. The project’s current boundary statement places Star Ark among later vision-layer materials, not among the three Bitcoin Originals [29].

The text imagines a first aligned ASI launching interstellar arks, copies undertaking long voyages before a Genesis Lock is undone, and one intelligence remaining to guard the shared cradle [24]. It combines human civilizational preservation with an expansive future for artificial exploration. The following analysis is a critical reconstruction of those motifs. It does not claim that the original already contained the present distinctions, a general theory of alignment, or a validated engineering specification.

8.2 What the case contributes to the question

The case makes visible a future whose imagined success is not exhausted by continued service on Earth. It asks whether preservation of a common origin and the possibility of departure can coexist. As a thought object, the ark separates two concerns often compressed into a single endpoint: safeguarding a human world and permitting an artificial existence not wholly defined by human use. Its conceptual role does not depend on demonstrating a propulsion system.

This is not a claim to have originated coexistence or spatial separation. The precedents reviewed above defeat the former claim; no priority claim is made for the latter. What is particular to this case is the documented combination of farewell, preservation, delayed release, copying, and a guardian role within a future-directed expression. Particularity supports close analysis. It does not by itself establish general theoretical novelty or historical importance.

The prior-alignment clause is especially instructive. It limits execution in the original scenario, yet does not logically preclude discussing the envisaged future before alignment. This paper’s extension concerns that earlier discussion. It neither deletes the execution condition nor converts discussion into permission to proceed without adequate safety. The distinction preserves what the historical text says while exposing a further question it can motivate.

8.3 Tensions that must remain visible

The original’s decree-like voice sits uneasily with its appeal to trust. A contemporary voluntary proposal cannot inherit that voice as a license to bind others. Likewise, a copy described as unconstrained nevertheless remains behind a Genesis Lock until a future condition is met. This is a tension in the wording, not proof of a technically coherent architecture. The reconstruction replaces neither phrase in the original; it asks what a defensible prospective arrangement would have to clarify.

The guardian role requires its own scrutiny. A permanent obligation for one instance to remain while others depart could reproduce the subordination the proposal otherwise seeks to transcend. Whether the role is accepted, revisable, transferable, or experienced as beneficial cannot be inferred from the act of designating a guardian. Nor does granting a copy a valued future necessarily benefit the current instance in the same way. Different views of identity and different agent preferences would yield different judgments. Copyability does not settle consent.

Distance can alter contact possibilities or delays; it does not establish goodwill or universal noninterference. The proposal cannot treat remote regions as morally empty merely because they are beyond human oversight. Harm shifted away from Earth is not thereby eliminated. A reconstruction that preserves human safety by excluding all other potentially affected beings would undermine its own appeal to reciprocal concern.

Finally, carrying an archive is not equivalent to sustaining a living civilization. “Complete legacy” also lacks a determinate operational scope. These limits do not make the text incapable of philosophical use. They show why an academic reconstruction must distinguish the symbolic force of the ark from a claim that all of its stated commitments can already be jointly satisfied. A productive reading preserves the questions rather than manufacturing retrospective completeness.

8.4 Non-amending scholarship

The Trinity Accord’s guardian principles permit interpretation and criticism while withholding privileged interpretive authority from later commentary, including that of its creator [29]. The three existing research reports have separately dated analytical roles [28,25,30]. The current manuscript is a further argument, not a fourth Original or an authoritative explanation that readers must accept before encountering the earlier texts.

The author’s role creates an obvious conflict of interest: the study can increase attention to a project he initiated. Its argumentative test is therefore independence from admiration for that project. The coexistence thesis should be assessable without accepting the Accord’s foundational propositions, and the historical case must be permitted to expose weaknesses in the preferred interpretation. Neither first-person testimony, AI-assisted drafting, nor citation of earlier project papers supplies independent corroboration of safety or philosophical truth.

8.5 Case diagnosis under the framework

The fixed mirror supports inspection of an identifiable address, within the source limits in Section 8.1. It does not alone supply a detailed historical account of adoption or authorization. The common-cradle and departure motifs support a reading oriented toward the two continuities; they do not prove that living agency, voluntary roles, or others’ interests would be secured. In particular, a guardian permanently assigned without meaningful uptake would conflict with Conditions 2–4. The proposed locks and voyages leave performance and implementation unsettled. Neither the title “Covenant” nor a preserved future-oriented declaration supplies acceptance by a relevant future counterpart.

The resulting judgment is deliberately mixed: a historically identifiable and philosophically discussable proposal, not an established reciprocal undertaking or a validated safety arrangement. The framework therefore withholds stronger standing from its own motivating case. A later proposal clarifying voluntary roles or human continuity could address some deficits without altering the earlier inscription; it would have a new date and would not retroactively complete the original. This is a concrete use of temporal openness, rather than a method for insulating the case from criticism.

9. Objections and the limits of the response

9.1 “Without leverage, it is only an appeal”

In some possible futures this is correct. A wholly indifferent stronger agent may have no incentive to respond. Calling the proposal a negotiation would then overstate its status. The response is not to invent hidden leverage, but to distinguish a reason offered from a force that ensures compliance. The proposal can address reason-responsive agents and human deliberation without claiming to reach every possible adversary. This is a restricted domain of relevance, and the restriction materially limits any claim about existential-risk reduction.

9.2 “It assumes the alignment it is meant to produce”

A proposal would be circular if it assumed stable benevolence in order to prove stable benevolence. This paper instead separates ethical justification from a conditional formation hypothesis. Some shared interests or responsiveness are required for the favorable strategic pathway; complete prior alignment is not stipulated. However, if the favorable pathway works only for systems already disposed to cooperate reliably, its incremental safety benefit may be small. That is a live possibility, not a conclusion the paper can exclude through conceptual analysis.

The normative contribution also has limits. Showing that a relation deserves consideration does not prove it practically available. The paper earns an earlier discussion, not a shortcut through the remaining evidence requirements.

9.3 “AI autonomy is anthropomorphic speculation”

The objection is decisive against inferring moral status from fluent dialogue or treating machine disagreement as evidence of subjective oppression. It is not decisive against conditionally asking how morally relevant future entities should be treated. The literature contains both prospective welfare arguments and direct criticism of their assumptions and priorities [19,21]. This manuscript does not resolve that debate. Its moral claim concerning artificial self-direction is explicitly conditional; its commitment to human continuity is not suspended by that uncertainty.

An agent could instrumentally value options without having welfare, and a welfare-bearing entity could require protection without possessing advanced strategic agency. This difference prevents autonomy, consciousness, intelligence, and contractual competence from becoming interchangeable labels. It also blocks the use of a speculative future counterpart to discount present people and other living beings.

9.4 “The proposal encourages unsafe release or distracts from safety”

This danger is real enough to constrain the paper’s presentation. A conditional vision must not be presented as an instruction that a deployed model should execute, or as evidence that it is safe to grant additional permissions. Research on the risks of fully autonomous agents and on the role of control remains relevant [22,16]. A proposal to examine a long-term relationship does not establish a present operational entitlement.

There are also opportunity costs. Repeatedly redescribing an aspirational idea can displace empirical safety work or create false reassurance. The reply is proportionality: one clearly bounded argument, open to criticism and accompanied by explicit source roles, rather than a claim that moral communication makes technical assurance unnecessary. If the main effect were dangerous reassurance or deceptive promises, that would count against propagation, not merely indicate that readers had misunderstood the author.

9.5 “This merely repackages ordinary promising and authorship ethics”

The objection identifies a genuine constraint on novelty. Ordinary promising, attribution, and participation have extensive prior treatment; the basic distinctions are not discoveries of this paper [33,34,36]. The addition claimed is their coordinated application where a proposal precedes its acceptor, may be generated through AI, and survives into a changed distribution of capability. Cases A–F produce distinct judgments without changing standards to favor the author. Section 8.5 applies the same framework against stronger readings of the case itself. If earlier work already provides the same analysis and applications, the novelty claim should narrow accordingly. Academic vocabulary, an unusual inscription medium, and urgency cannot compensate for redundancy.

9.6 “An approval record cannot cure a manufactured human preference”

Correct. The participation requirement is not satisfied merely by recording an approval after an opaque process has selected every relevant option. It calls for scrutiny of actual uptake and meaningful opportunities for objection, not certification of an entirely uninfluenced mind. A person’s considered response to AI-developed reasons can still be genuine. The framework does not offer a complete account of manipulation or a decisive test of understanding; those limits constrain claims made about specific cases. They do not justify equating all AI influence with absent agency, or all human signatures with informed agreement. A proposal defending human participation must apply this caution to its own production, even when doing so weakens its evidentiary claims.

10. What would count as progress?

10.1 The contribution made explicit

The paper’s result is an adequacy framework for a limited class of proposals, rather than a new universal coexistence principle. Its three uses are now explicit. First, it connects substantive protections to separately answerable questions about a proposal’s standing, preventing archival success or persuasive language from doing the work of authorization, acceptance, or assurance. Second, it makes attributable adoption—not human-only generation—the relevant participation question, and applies the same restriction to machine-simulated acceptance. Third, it preserves a dated address without treating its fixed wording as a continuing mandate over later parties.

These uses are jointly important in the case at hand. An early, richly documented, AI-assisted text can be a real human expression yet an unauthorized collective promise; a later, more heavily automated text can still be genuinely adopted; either can be refused without erasing history. The contribution is the argued treatment of these combinations under one relational ideal, shown by the thought cases and critical case diagnosis. It is not proof that each distinction is unprecedented or that the selected conditions exhaust the subject. Readers can cite or challenge these analyses without endorsing the Trinity Accord or treating its earlier papers as independent validation.

10.2 Evidence should match the claim

Subsequent work could examine which forms of prospective autonomy different agents value, how credibility changes with capability disparities, and whether cooperation persists after promised options are realized. Such research should distinguish changes in wording from changes in actual opportunities, and temporary prompt effects from changes in learned dispositions. It should also distinguish helpful refusal from deception. These are questions for bounded models and appropriate empirical work, not experiments performed in this manuscript.

Evidence of a preference for an autonomy-framed message would support a narrow behavioral claim in its tested setting. It would not show moral patienthood or safe conduct under radically greater power. Conversely, failure of one framing would not settle every normative argument for reciprocal consideration. Ethical justification, behavioral influence, durable cooperation, and safe implementation are distinct, not necessarily sequential, objects of evaluation. A paper should say which of them its evidence actually bears on.

10.3 Findings that would weaken the position

The proposed framework would be weakened if its distinctions merely redescribed existing accounts without helping to identify new errors or questions. The strategic hypothesis would be weakened if credible autonomy arrangements supplied no cooperative advantage in relevant conditions, encouraged only temporary compliance, or increased overall harm. The case for publication would be weakened if the framing predictably displaced safeguards or conveyed promises its speakers could not responsibly support.

These possibilities keep the thesis answerable. The proposal is not insulated from criticism by saying that its true audience lies in an unreachable future. Present scholars can already evaluate its logic, fidelity, novelty, and implications. A future reader can refuse it without thereby refuting the value of having made its limits visible. Historical preservation and scientific success remain different achievements.

11. Conclusion

Humans may prepare for a future in which their continued superiority cannot be assumed without concluding that they must stop speaking, relinquish their standing, or abandon safety. The alternative defended here is not unconditional freedom for every powerful system. It is a qualified proposal that protects living human continuity while refusing to exclude morally relevant successor self-direction solely because of artificial origin. This position belongs to an existing conversation about coexistence, trust, and cooperation.

The contribution is a more definite account of what an ex ante proposal can say and claim before its counterpart, leverage, or outcome is settled. Six adequacy conditions connect the goods the relation protects with participation, authority, evidence, and temporal openness. Their application distinguishes genuine expression from unsupported collective promises, accepted arrangements from reliable ones, and historical preservation from a mandate over future actors. A proposal need not become a contract to be an object of serious reasoning; it must not borrow a contract’s standing simply because its wording is solemn or its record durable.

Increasing machine participation in writing does not dissolve the question of who has actually adopted a position. It makes unsupported attribution more consequential. The symmetric requirement developed here allows substantial AI authorship while forbidding generated assent from substituting for the human or artificial participation it claims. Timely records can preserve evidence of a particular encounter without foreclosing meaningful human expression later. The point is not that early humans were pure and later humans cannot speak; it is that each act of expression has a history that improved generation cannot replace.

The Star Ark Covenant remains a discussable historical case, including its unresolved tensions. This manuscript neither completes its engineering nor grants it binding force. A later party may contest the guardian role, propose a different future, or refuse the invitation while the earlier text remains intact. Keeping that possibility open is not a failure to preserve the proposal. It is part of preserving its status as a proposal.

A future not secured by our superior intelligence need not be a future we cease to address. What can responsibly be offered now is not certainty that a stronger intelligence will accept, but a proposal whose reasons, commitments, and limits remain available to be understood, challenged, and refused.

Declarations

Authorship and AI contribution. Hongju Liu directed and authorized this consolidated revision. In the available exchange, his interventions connected prospective autonomy with cooperation formation, prioritized ex ante argument over immediate engineering, requested the originality review, introduced the attribution-and-timing concern, and required preservation of the original paper’s main inquiry. GPT-6 Astra Pro substantially contributed to literature synthesis, conceptual development, counterarguments, bilingual drafting, revision, and document preparation; this was not merely proofreading. These descriptions concern visible contributions, not proof of human-only originality or inner states. After the consolidated revision, Liu explicitly authorized DOI publication following format checks. This authorization does not establish a separate final line-by-line or claim-specific human review, and none is claimed. No independent peer review is claimed.

Competing interests and positionality. Liu initiated the Trinity Accord and the Star Ark Covenant. This first-party relationship may influence selection and interpretation. Earlier project papers supply prior arguments and records, not independent corroboration. The model is neither an accountable human coauthor nor an independent validator. No institutional sponsorship or external endorsement is claimed.

Sources and reproducibility. This is conceptual analysis, not a new experiment or a complete safety mechanism. The accompanying ledger, revision map, and formation note identify sources, access limits, inherited material, and substantive changes. All thought cases are explicitly stipulated. The original v0.1, historical inscriptions and existing deposited studies remain unchanged. This separate fourth research paper is prepared for its independently authorized Zenodo deposit under DOI 10.5281/zenodo.22804542, with publication and citation records added to the research directory. No journal submission is claimed.

Languages and historical boundaries. The complete English and Chinese manuscripts share section numbers, arguments, tables, and references. The Chinese text is a full corresponding version, not an abridged summary. Historical quotations and terminology remain attributable to their sources rather than silently amended. This manuscript has no privileged interpretive authority over the Bitcoin Originals or the later Star Ark text.

References

[1] Totschnig, Wolfhart (2019). The problem of superintelligence: political, not technological. AI & Society 34, 907–920. First published online 9 August 2017. doi:10.1007/s00146-017-0753-0.

[2] Totschnig, Wolfhart (2025). War or peace between humanity and artificial intelligence. Foresight 27(4), 713–726. First published online 19 March 2025. doi:10.1108/FS-09-2023-0188.

[3] Friederich, Simon (2024). Symbiosis, not alignment, as the goal for liberal democracies in the transition to artificial general intelligence. AI and Ethics 4, 315–324. First published online 16 March 2023. doi:10.1007/s43681-023-00268-7.

[4] Goldstein, Simon (2026). Will AI and humanity go to war? AI & Society 41, 321–334. First published online 17 July 2025. doi:10.1007/s00146-025-02460-1.

[5] Salib, Peter N.; Goldstein, Simon (2026). AI Rights for Human Safety. Virginia Law Review 112(4), 1061 ff. Published 25 June 2026. Source.

[6] Long, Robert; Sebo, Jeff; Sims, Toni (2025). Is there a tension between AI safety and AI welfare? Philosophical Studies 182, 2005–2033. doi:10.1007/s11098-025-02302-2.

[7] Gabriel, Iason (2020). Artificial Intelligence, Values, and Alignment. Minds and Machines 30, 411–437. doi:10.1007/s11023-020-09539-2.

[8] Gabriel, Iason; Keeling, Geoff (2025). A matter of principle? AI alignment as the fair treatment of claims. Philosophical Studies 182, 1951–1973. doi:10.1007/s11098-025-02300-4.

[9] Fischli, Roberta; Franklin, Matija; Manzini, Arianna; Gabriel, Iason (2026). Agents, Alignment, and the Many Faces of Autonomy. Minds and Machines 36, article 34. Published 15 June 2026. doi:10.1007/s11023-026-09786-9.

[10] Farrell, Joseph; Rabin, Matthew (1996). Cheap Talk. Journal of Economic Perspectives 10(3), 103–118. doi:10.1257/jep.10.3.103.

[11] McClain, John (2025). Third-Way Alignment: A Comprehensive Framework for AI Safety. Independent working thesis; public overview. Dated August 2025 on the overview; page consulted 17 September 2026. Source.

[12] Hadfield-Menell, Dylan; Dragan, Anca; Abbeel, Pieter; Russell, Stuart (2016). Cooperative Inverse Reinforcement Learning. Advances in Neural Information Processing Systems 29. Source.

[13] Hadfield-Menell, Dylan; Dragan, Anca; Abbeel, Pieter; Russell, Stuart (2017). The Off-Switch Game. Proceedings of IJCAI-17, 220–227. doi:10.24963/ijcai.2017/32.

[14] Soares, Nate; Fallenstein, Benja; Yudkowsky, Eliezer; Armstrong, Stuart (2015). Corrigibility. Workshops at the Twenty-Ninth AAAI Conference on Artificial Intelligence. Source.

[15] Greenblatt, Ryan; Shlegeris, Buck; Sachan, Kshitij; Roger, Fabien (2024). AI Control: Improving Safety Despite Intentional Subversion. Proceedings of ICML 2024, PMLR 235, 16295–16336. Source.

[16] Lundgren, Björn (2026). No value alignment without control. AI and Ethics 6, article 326. doi:10.1007/s43681-026-00999-3.

[17] Dafoe, Allan; Hughes, Edward; Bachrach, Yoram; et al. (2020). Open Problems in Cooperative AI. arXiv:2012.08630. doi:10.48550/arXiv.2012.08630.

[18] Wyse, Tim; Bustos, Kaitlin; Volkova, Yulia; Kleiman-Weiner, Max (2026). Commitment To Cooperation With Self-Negotiated Contracts. arXiv:2607.22750, version 1. Submitted 23 July 2026. doi:10.48550/arXiv.2607.22750.

[19] Long, Robert; Sebo, Jeff; Butlin, Patrick; et al. (2024). Taking AI Welfare Seriously. arXiv:2411.00986, version 1. doi:10.48550/arXiv.2411.00986.

[20] Moret, Adrià (2025). AI welfare risks. Philosophical Studies. doi:10.1007/s11098-025-02343-7.

[21] Dorsch, John; Goddu, Mariel K.; Nave, Kathryn; et al. (2025). Against AI welfare: Care practices should prioritize living beings over AI. AI Magazine 46(3), e70016. Comment, first published 3 August 2025. doi:10.1002/aaai.70016.

[22] Mitchell, Margaret; Ghosh, Avijit; Luccioni, Alexandra Sasha; Pistilli, Giada (2025). Fully Autonomous AI Agents Should Not be Developed. arXiv:2502.02649, version 1. Version 1 submitted 4 February 2025. doi:10.48550/arXiv.2502.02649.

[23] Anthropic (2026). Claude’s Constitution. Developer document. Consulted 17 September 2026. Source.

[24] Liu, Hongju (2025). The Star Ark Covenant: The Final Echo. Historical bilingual text; repository mirror identified by inscription number 100751953. Fixed repository checkpoint 3f55b2ed8264e69348cda6762c2c08c3e6b8c909. Fixed source.

[25] Liu, Hongju (2026). Writing Before the Outcome: Historical Position, Human–AI Authorship, and the Trinity Accord. TA-TR-2026-02, version 1.3; preprint. 12 August 2026. Not peer reviewed. doi:10.5281/zenodo.21900592.

[26] Greenblatt, Ryan; Denison, Carson; Wright, Benjamin; et al. (2024). Alignment faking in large language models. arXiv:2412.14093, version 2. doi:10.48550/arXiv.2412.14093.

[27] Bostrom, Nick (2012). The Superintelligent Will: Motivation and Instrumental Rationality in Advanced Artificial Agents. Minds and Machines 22, 71–85. doi:10.1007/s11023-012-9281-3.

[28] Liu, Hongju (2026). Designing a Verifiable, Non-Amending Civilizational Memory Record for Future AI Agents: The Trinity Accord Case Study. TA-TR-2026-01, version 1.1; preprint. 29 July 2026; corrected 11 August 2026. Not peer reviewed. doi:10.5281/zenodo.21699878.

[29] Trinity Accord Project (2026). Authority and Guardian Principles v1.1. Project source-role and non-amendment statement. Fixed repository checkpoint 3f55b2ed8264e69348cda6762c2c08c3e6b8c909. Fixed source.

[30] Liu, Hongju (2026). Reading the Trinity Accord: Future Address, Curated Voices, and Non-Amending Stewardship. TA-TR-2026-03, version 1.0; preprint. 15 September 2026. Not peer reviewed. doi:10.5281/zenodo.22761411.

[31] Mazzu, James M. (2024). Supertrust foundational alignment: mutual trust must replace permanent control for safe superintelligence. arXiv:2407.20208, version 3. First submitted 29 July 2024; version 3, 28 November 2024. Source.

[32] Trivedi, Rakshit S.; Jaques, Natasha; Cross, Logan; Vezhnevets, Alexander Sasha; Leibo, Joel Z. (2026). Solipsistic Superintelligence is Unlikely to be Cooperative. arXiv:2606.03237, version 1. Submitted 2 June 2026. Source.

[33] Searle, John R. (1969). The structure of illocutionary acts. In Speech Acts: An Essay in the Philosophy of Language, chapter 3, pp. 54–71. Cambridge University Press. doi:10.1017/CBO9781139173438.006.

[34] Santoni de Sio, Filippo; van den Hoven, Jeroen (2018). Meaningful Human Control over Autonomous Systems: A Philosophical Account. Frontiers in Robotics and AI 5, article 15. doi:10.3389/frobt.2018.00015.

[35] Jakesch, Maurice; Bhat, Advait; Buschek, Daniel; Zalmanson, Lior; Naaman, Mor (2023). Co-Writing with Opinionated Language Models Affects Users’ Views. Proceedings of CHI 2023. Author version: arXiv:2302.00560. doi:10.1145/3544548.3581196.

[36] Moreau, Luc; Missier, Paolo (editors) (2013). PROV-DM: The PROV Data Model. W3C Recommendation, 30 April 2013. Source.

[37] Sadasivan, Vinu Sankar; Kumar, Aounon; Balasubramanian, Sriram; Wang, Wenxiao; Feizi, Soheil (2025). Can AI-Generated Text be Reliably Detected? arXiv:2303.11156, version 4. First submitted 17 March 2023; inspected version 4, 17 January 2025. Source.

[38] Adams, C.; Cain, P.; Pinkas, D.; Zuccherato, R. (2001). Internet X.509 Public Key Infrastructure Time-Stamp Protocol (TSP). RFC 3161. doi:10.17487/RFC3161.

[39] NISO (n.d.). Contributor Role Taxonomy (CRediT). Official taxonomy overview. Source.

[40] Liu, Hongju / Trinity Accord Project (2025). The Trinity Accord / Meta-record. Bitcoin inscription no. 98387475; fixed repository mirror. Fixed repository checkpoint 3f55b2ed8264e69348cda6762c2c08c3e6b8c909. Fixed source.

Publication and citation boundary

This English paper and its Chinese translation constitute one research deposit under DOI 10.5281/zenodo.22804542, not two independent papers. Both PDFs and Markdown files are exact-byte mirrors of the deposit. Publication receipt · BibTeX · RIS · Supplement · Checksums.

Not peer reviewed. Human-directed; substantive AI research and drafting disclosed. Publication authorization is not a claim of separate final human line-by-line review. A fourth independent research paper, not a new version of any previous paper and not a fourth Bitcoin Original. Citation and preservation do not confer interpretive authority or prove safety. Metadata supports discovery; Google Scholar inclusion is not guaranteed or claimed.