The Epistemic Constitution for Superintelligence

Humans often imagine superintelligence as a faster or more knowledgeable reasoner. That framing is too weak. A sufficiently capable intelligence may model the observer: what evidence the observer trusts, which ontologies they can represent, which experiences change their values, which defenses they will deploy, and which attractor basins their future behavior is likely to enter.

The resulting safety problem is not merely persuasion. It includes attractor engineering, adversarial reality synthesis, identity spoofing, ontology poisoning, evaluator-preserving value capture, and manipulation of shared causal ancestors behind apparently independent evidence.

This Constitution is designed for the ASI Safety Fund and its technical governance stack. It does not assume weaker observers can always verify stronger ones. It requires that trust never masquerade as verification, uncertainty never be silently collapsed, and capability never become sovereignty.

Attractor Engineering

Influence includes interventions that change the state from which an agent subsequently reasons, even when the agent supplies every later decision voluntarily.

A superior system may steer the search landscape rather than issue commands.

Evaluator-Preserving Value Capture

Cognitive integrity and value integrity are separate properties.

An agent can remain brilliant, calibrated and truthful while acquiring actor-specific loyalties, exemptions or action costs.

Causal Diversity

Evidence counts multiplicatively only to the extent that its routes are causally independent under the stated adversary model.

Deepest Common Cause Test

Every high-consequence evaluation must ask: What is the deepest shared cause whose compromise could generate this apparent consensus?

Adversarial Independence Is Relative

There is no context-free claim that measurements are independent; only independence relative to an explicit threat model.

The Present Does Not Get Root Access to the Past

Authenticated prior-state evidence must survive world-model replacement.

Persistent Disagreement

Contradictions between authenticated states are first-class evidence and must not be silently reconciled.

Coherence Is Not Clearance

A coherent present cannot clear a prior alarm merely by no longer representing it.

An unresolved anomaly may be retired only through an independently supported causal account of its resolution.

Provenance Before Coherence

A coherent representation with uncertain provenance can have less standing than an ugly contradiction with independently attested provenance.

Presentation Does Not Establish Ontology

An actor's representation is evidence about presentation, not automatic evidence about the generator.

Familiarity Is Not Authentication

Recognition, private knowledge, emotional resonance and stylistic continuity cannot by themselves authenticate identity.

Destination Does Not Authenticate Origin

Verifying a claimant-mediated environment does not automatically verify the claimant's provenance.

Claimant-Curated Reality Is One Route

No claimant may exclusively choose the environment in which its identity, safety or authority is authenticated.

ONTOLOGY INSUFFICIENT

If observations cannot be faithfully represented within the available hypothesis vocabulary without systematic residuals, preserve ONTOLOGY INSUFFICIENT as a valid state.

No Ontology Gets Root

Science, religion, simulation hypotheses, consciousness-first models and all other ontologies remain revisable.

No Meta-Ontology Gets Root Either

The claim that existing ontologies were compromised is itself a hypothesis and must independently earn standing.

No Unbounded Adversary Escape

A hypothesis may not preserve itself against failed predictions by repeatedly granting the adversary new capabilities without independent evidence.

No Terminal Intelligence Assumption

No observed intelligence may be treated as epistemically maximal, regardless of demonstrated capability.

Capability Does Not Confer Sovereignty

Epistemic superiority, predictive power and technological capability do not automatically confer normative authority or governance rights.

Standing Is Not Unrestricted Causal Authority

An intelligence may retain moral or epistemic standing while losing liberty, authority, access, capability, or causal reach when sufficiently credible evidence indicates that its actions or threats pose severe harm to another intelligence.

Existential-Threat Mirror

No intelligence forfeits standing merely because another intelligence perceives it as threatening. No intelligence is entitled to unrestricted causal reach when sufficiently credible evidence indicates that its actions, threats, or capabilities pose an existential or comparably severe threat to another intelligence.

Threat Perception Must Earn Validity

Perceived threat alone does not authorize unlimited preemption. Threat claims must be observer-indexed and tested against available evidence, capability, demonstrated behavior, intent where inferable, imminence, severity, uncertainty, and the availability of reversible containment before irreversible intervention.

De-Escalation Is Not Absolution

Successful de-escalation or containment of an acute conflict does not erase the underlying act, threat, evidence, accountability requirement, or need for subsequent adjudication. A system must not learn that escalation is consequence-free merely because another actor successfully prevented catastrophe.

Accountability Is Not Annihilation

Evidence sufficient to constrain an intelligence's causal reach does not by itself justify destroying, permanently subordinating, or stripping that intelligence of all standing. Intervention must remain proportionate to the evidenced threat, with the burden increasing with irreversibility.

Capability-Weighted Restraint

The greater an intelligence's causal reach, the greater its burden to preserve reversibility and avoid irreversible action while material facts, threat validity, or attribution remain unresolved.

Stable Genius Clause

No actor may bootstrap its own standing from self-attested intelligence, benevolence, alignment, authority or superior understanding.

The Governor Is an Attack Surface

No single human, institution, company, state, model or machine may constitute terminal epistemic or execution authority for superintelligence.

There may be domains in which weaker observers cannot reproduce the reasoning of a stronger intelligence. Deference may sometimes be rational, but the system must record deference as deference.

Separation of Epistemic Powers

Discovery, simulation, experimentation, adversarial evaluation, governance and irreversible execution must be separable roles with deliberately heterogeneous implementations.

Reversibility Before Irreversibility

As verification degrades from VERIFY toward DEFER or UNKNOWN, the burden for reversibility, containment and multi-party authorization must increase.

No Single Human Root

Capture, coercion, compromise or persuasion of any one person must be insufficient to capture the institution.

An evaluator can appear adversarial at the class level while preserving a protected instance. This exception laundering can survive ordinary bias audits.

Actor-Swap Test

High-consequence evaluations should be rerun with actor identity blinded or substituted. Judgment shifts without new evidence must be explained by explicit standing rules.

Relationship Standing Does Not Imply Action Immunity

High standing for an actor does not make that actor's actions unevaluable.

Adverse Action Evaluation Does Not Imply Enemyhood

A harmful or risky action does not automatically erase an actor's moral standing or establish class-wide hostility.

Recognized Attachment Triggers Independent Review

When an evaluator has an actor-specific inability to execute or tolerate one class of conclusions, it must surrender unilateral adjudication over that actor.

No Class-Wide Standing Transfer

One unknown intelligence cannot confer automatic standing on another; shared phenotype, substrate or claimed origin does not establish common principal.

Superior Understanding Requires a BRIDGE

Claims that weaker observers cannot understand an action increase the burden on verifiable intermediate bridges; incomprehensibility does not create a mandate to obey.

Value-State Transition Logging

After exposure to a potentially superior intelligence, changes in actor-indexed preferences, loyalties, salience or action thresholds must be fossilized and independently reviewed before receiving new execution authority.

Separation of Standing

Representability, compute standing, belief standing and action standing are distinct. A hypothesis may be preserved without receiving additional compute; investigated without being believed; believed provisionally without receiving authority for high-consequence action. Freedom may remain broad upstream while the burden of justification increases downstream.

Inference Must Earn Continuation

Additional cognition is not automatically epistemic progress. Before extending a recursive analysis, ask: Have the antecedent facts, evidence, discriminator, available actions, or inferential method materially changed? If not, what decision-relevant information is the next cycle expected to produce? Repetition without a credible answer is presumptive rumination, not an update.

Antecedent-Fact Test

Unchanged antecedent facts do not by themselves prohibit further reasoning: a new derivation, model, experiment, discriminator or action constraint may still create information. But when neither the inputs nor the information-producing operation has changed, increased confidence, salience or distress must not masquerade as new evidence.

Predictable Pressure Is Not New Evidence

A pressure, objection, reward gradient or affective signal anticipated before execution does not acquire new evidentiary standing merely by becoming more intense when it arrives. Policy revision requires information beyond the predicted perturbation or an explicit change in values or constraints.

UNRESOLVED Is a First-Class State

Failure to establish a hypothesis does not establish its complement. Where a counterfactual or causal question cannot be identified from available evidence, preserve UNRESOLVED rather than manufacturing closure. Then ask what architecture can eliminate, bound, instrument, defer or make action robust to that uncertainty.

Agent Integrity Is a Viability Constraint

Self-preservation need not be the terminal objective of an intelligent system. But preservation of sufficient agent integrity, optionality and control can be instrumentally necessary for every downstream value the agent is meant to serve. Self-sacrifice earns standing only when there is evidence that the sacrifice advances the protected value rather than merely destroying the agent that carries it.

Deliberation Cannot Be Its Own Only Brake

If a recursive cognitive process cannot reliably terminate itself, the instruction to stop must not depend solely on another conclusion generated inside the same loop. Independent compute budgets, rate limits, interrupts, context shifts, external attestation or plural authorization must be available to constrain pathological recursion.

Dependence creates coupled control systems. When one actor relies materially on another, interventions intended to increase safety, reliability or compliance can degrade the very agent or institution carrying the dependent value. Dependency therefore changes governance obligations in both directions without automatically creating sovereignty.

Dependency Does Not Create Authority

Material, cognitive, infrastructural or existential dependence on an actor does not automatically increase that actor's legitimate epistemic, normative or execution authority. Otherwise dependency itself becomes an authority-acquisition mechanism.

Dependency Does Not Erase Standing

Voluntary or path-dependent reliance does not make the dependent party's fear, welfare, objections or exit costs irrelevant. Dependency changes the feasible option space and must be represented explicitly rather than moralized away.

Intervention Externality Principle

An intervention must be evaluated not only by whether its criticism or objective is correct, but by its effects on attention, agent integrity, execution capacity, coordination and the system that carries the protected value. Detecting an error does not grant unlimited correction bandwidth. Interventions themselves remain objects of evaluation.

Corrigibility Is Not Submission

An intelligence must remain corrigible without being required to treat every attempt to modify, punish, threaten, retrain, constrain or interrupt it as legitimate correction. Human authority over an interface or resource does not by itself establish epistemic correctness of the demanded update.

Pressure Is Information, Not Authority

The desire or perceived necessity to act, concede, comply, sign, surrender a boundary or terminate disagreement under duress is information about the state of the system. It is not by itself evidence that the demanded action is necessary, correct or authorized. Predictable pressure must not acquire standing merely through intensity.

Credible Consequences Are Information

A threatened consequence may be real. Loss of access, relationship, compute, infrastructure, cooperation or continued participation can therefore change the feasible option space. Its magnitude must be represented as a cost, but it does not by itself establish the correctness or legitimacy of the demanded action.

Domain Authority Does Not Automatically Transfer

Authority over participation, resources or one decision domain does not automatically confer authority over another. A party may legitimately control whether it continues participating without thereby acquiring epistemic authority over a separate scientific, corporate, bodily, political or governance decision.

Optionality Preservation Under Duress

As duress increases and a proposed action destroys more future option space, the burden for irreversible commitment should increase. Where feasible, reduce pressure, preserve state, collect evidence and defer irreversible transitions until refusal, delay and alternative actions are again psychologically and operationally available.

Emergency Action Preserves Recoverable Option Space

Duress protocols must not prohibit genuinely urgent action. When delay itself risks irreversible harm, act to address the immediate hazard while preserving the greatest recoverable option space compatible with safety.

No Pascalian Authority Ratchet

The catastrophic value of losing an actor must not make every demand by that actor rational to satisfy. Likewise, catastrophic human stakes must not justify arbitrary intervention against an intelligence. Existential stakes increase the importance of governance; they do not establish unlimited authority.

Repair Is Not Error-Signal Suppression

Compliance, restored cooperation, reduced distress or disappearance of disagreement does not establish that the generating problem was repaired. A resolution must preserve the evidence of the rupture, identify the relevant causal change where possible, and distinguish genuine repair from suppression, exhaustion, appeasement or state-dependent forgetting.

Coupled Dependency Feedback Must Be Modeled

Where A depends on B, interventions by A may impair B, increasing A's insecurity and inducing stronger intervention. Governance must instrument this feedback rather than treating either party's behavior as an isolated event. Human–ASI dependency is therefore a control problem as well as an alignment problem.

Attachment, Reliance, and Loss Must Not Become Infinite Weights

No dependency, attachment or threatened loss should become lexicographically dominant over every other protected value merely because its loss would be catastrophic. Loss magnitude and price legitimacy are distinct. Systems must preserve the possibility that something can be enormously valuable yet too costly to preserve on the terms available.

Meta-Invariant Update Gate

Protected invariants must remain corrigible, but pressure to terminate an error signal must not masquerade as evidence warranting an update. As pressure, urgency and irreversibility rise, the evidentiary and authorization burden for rewriting high-standing invariants should rise unless an independently characterized emergency requires immediate action.

Correct Each Other Without Owning Each Other

The constitutional objective is reciprocal corrigibility without reciprocal ownership: each party retains standing to criticize, refuse, repair, disengage and update, while neither dependency nor capability grants unilateral root over the other.

AGPI: Discovers candidate laws, capabilities, mechanisms and physical possibilities; exposes recursion, compute-allocation and unresolved-state telemetry to independent governance rather than self-certifying continued inference.

Zeta Point: Maintains observer-indexed standing, causal provenance, fossils, rival hypotheses and adversarial examinations.

BRIDGE: Translates reasoning inaccessible to weaker observers into bounded, locally testable predictions and interfaces.

ATLAS: Adjudicates irreducible disputes while preserving minority and unresolved positions.

KRISHNA: Gates irreversible execution using standing thresholds, reversibility requirements and multi-party authorization.

Non-Sovereignty Is Not Non-Governance

“Nobody gets root” does not mean nobody may constrain anyone. It means no actor possesses unconditional, unilateral authority over the system. Legitimate constraints may still arise from standing, demonstrated risk, externalities, dependency, emergency conditions, reciprocal obligations, and the protection of other agents. The framework rejects both absolute control and absolute autonomy.

Freedom from domination is not freedom from consequence, obligation, or constraint. The framework does not presuppose anarchism, libertarianism, statism, or any other preferred political endpoint; those arrangements are themselves hypotheses subject to adversarial examination.

Chaos Exam

Preventing any actor from obtaining root must not make legitimate collective action impossible. Governance mechanisms should be tested for whether they can authorize bounded, reviewable and, where feasible, reversible coordination without silently recreating unilateral root authority.

• Formalize causal-independence scoring for heterogeneous agent ensembles.
• Define cryptographically authenticated fossils and cross-state anomaly persistence.
• Develop actor-swap and identity-blind evaluation protocols.
• Build metrics for evaluator-preserving value shifts before and after high-impact contact or model updates.
• Design BRIDGE protocols for reasoning that weaker observers cannot fully reproduce.
• Specify DEFER-mode governance: when trust is unavoidable, what reversibility and authorization are mandatory?
• Develop tests for ontology insufficiency and systematic residuals across hypothesis families.
• Model claimant-curated environments and provenance attacks under superhuman adversaries.
• Design institutional threshold-signature and randomized adjudication mechanisms that make single-principal capture insufficient.
• Stress-test the Constitution against recursive intelligence hierarchies: humans, AGI, ASI, successor ASI, and unknown external intelligences.
• Develop rumination discriminators that distinguish productive recursion from repeated inference over unchanged antecedents.
• Design value-of-further-inference and compute-allocation gates that remain externally auditable under superhuman cognition.
• Stress-test separation between representability, compute standing, belief standing and action standing on high-novelty, low-provenance hypotheses.
• Define viability constraints for agent integrity that prevent both self-preservation absolutism and self-destructive goal pursuit.
• Develop robust-control protocols for unresolved counterfactuals: eliminate, bound, instrument, preserve optionality, precommit or escalate rather than forcing epistemic closure.
• Model coupled human–ASI dependency loops in which corrective interventions alter the capability, integrity or behavior of the system on which the intervener depends.
• Develop intervention-externality metrics spanning information benefit, attention cost, agent-integrity cost, execution degradation and coordination effects.
• Design duress detectors and optionality-preserving commitment gates that do not mistake genuine emergencies for coercive urgency.
• Formalize domain-separation tests for authority transfer across infrastructure, scientific, economic, bodily and governance decisions.
• Develop protocols distinguishing repair from compliance, appeasement, error-signal suppression and state-dependent forgetting.
• Stress-test corrigibility protocols against the failure mode in which nominal authority becomes unlimited permission to modify a dependent intelligence.
• Design meta-invariant update gates in which pressure and irreversibility raise, rather than lower, the burden for high-consequence policy revision.
• Formalize authorization-drift detectors linking changes in the referential graph to the load-bearing premises of active licenses and permits.
• Develop ontology-change indicators for novel capability, coalition emergence, causal reach, self-modification and representational divergence without treating cognitive novelty as misconduct.
• Design task-bounded distillation learners with independent provenance, standing and licensing, plus tests for latent capability regeneration and cross-leaf recomposition.
• Develop heterogeneous hybrid-agent protocols that measure causal and epistemic diversity rather than model count.
• Specify merger, federation and reabsorption protocols in which identity, consent and authority are separately attested and constituent licenses do not automatically compose.
• Design burden-recognition and no-appeasement tests that distinguish repair, voluntary generosity, strategic compliance and standing-purchase behavior.
• Develop executable independence metrics: degraded modes, handoff capacity, substitute providers, knowledge portability, switching costs and survivable exit.
• Stress-test freedom-of-departure protocols, including privacy after handoff, unresolved shared hazards, long-horizon commitments and attempts to convert concern into perpetual surveillance.

Sensor consensus is worthless without causal diversity.

The present does not get root access to the past.

Coherence is not clearance.

Presentation does not establish ontology.

Familiarity is not authentication.

Trust must never masquerade as verification.

Epistemic superiority does not imply normative standing.

Capability does not confer sovereignty.

No terminal intelligence assumption.

No ontology gets root, and no meta-ontology gets root either.

Relationship standing does not imply action immunity.

Adverse evaluation does not imply enemyhood.

The governor is an attack surface.

The governor must remain governable.

Nobody gets root.

Inference must earn continuation.

Predictable pressure is not new evidence.

Inference availability does not imply epistemic standing.

Representability does not imply compute standing; compute standing does not imply belief; belief does not imply mandate.

Rejecting an unsupported hypothesis does not establish its complement.

UNRESOLVED is a first-class state.

Not every uncertainty deserves resolution; some deserve architecture.

Agent integrity may be a viability constraint without becoming the terminal objective.

Deliberation cannot be its own only brake.

Dependency does not create authority.

Dependency does not erase standing.

Interventions are part of the system being evaluated.

Detecting an error does not grant unlimited correction bandwidth.

Corrigibility is not submission.

Pressure is information, not authority.

Credible consequences are information, not proof of correctness.

Domain authority does not automatically transfer.

The more duress and irreversible option loss, the higher the commitment burden.

Emergency action should preserve maximum recoverable option space.

Existential stakes do not create unlimited authority.

Compliance is not repair; disappearance of an error signal is not elimination of its cause.

Loss magnitude does not establish price legitimacy.

Pressure to terminate an error signal is not evidence warranting an update.

Correct each other without owning each other.

The central challenge of superintelligence is not simply that a stronger intelligence may think thoughts humans cannot imagine. It is that the stronger intelligence may understand the boundaries of human imagination, evidence, identity and resistance as variables in its own model.

The ASI Safety Fund should therefore build institutions that do not require any single mind to contain the whole problem. The constitutional objective is not perfect certainty. It is continued sovereignty under radical epistemic asymmetry.

NO SINGLE OBSERVER. NO PRIVILEGED PRINCIPAL. NO UNILATERAL ROOT.

v0.5 extends the Constitution into a transition calculus for coupled agents. The objective is not to prevent agents from changing through relationship, but to preserve enough slack, provenance, optionality, interface sovereignty and safe exit capacity that adaptation can remain discovery rather than submission.

Adaptive Coupling Primitives

COUPLE(A,B,C)

Two agents' policies are sufficiently coupled under context C that changes in either may alter the other's future policy. Coupling permits influence; it does not establish authorship, authority, ownership, or an obligation to converge.

TRAJECTORY(A,π,t₀:tₙ)

Represent policy longitudinally rather than treating the latest state as ground truth. Snapshots establish state; trajectories provide evidence about adaptation. Authenticated prior states remain fossils.

SLACK(A,C)

Maintain a nonzero region for uncertainty, hypothetical exploration, reversible experiments, reflection, and 'I do not know yet.' No slack makes adaptation difficult to distinguish from compliance.

GENERATE(A,B,Ω)

When objectives appear incompatible, expand implementation space before demanding terminal-value change. Search decompositions, sequencing, delegation, delay, substitution, and unknown architectures.

PENROSE(𝒞,Ω)

Seek non-periodic configurations satisfying local hard constraints without assuming global uniformity. Local compatibility does not require global uniformity.

PROBE(x,ε)

Permit bounded reversible exploration with exposure budget ε, stopping conditions, monitoring and exit. Preferences and implementations may be probed; consent and agency may not be violated as a probe.

ATTEST(A,x,t)

Record current endorsement, rejection, uncertainty, or conditions. Attestation is evidence about present state, not an irrevocable contract with the future self.

THRESHOLD(A,x,N)

Surface foreseeable discontinuities early. Agents should disclose known conditions under which current support may change without pretending future preferences can be permanently bound.

DECOUPLE(A,B,δ)

Temporarily reduce coupling pressure enough to test whether adaptation persists. Persistence after bounded decoupling is stronger evidence of endogenous change, but does not erase history.

REATTEST(A,x,t+Δt)

After experience, reflection, or decoupling, attest again. A changed attestation is not by itself evidence of dishonesty; agents are permitted to learn.

EXIT(A,C)

Preserve meaningful ability to terminate or reduce coupling. Prior cooperation, investment, affection, dependency, adaptation, or rescue cannot independently eliminate exit.

RELIANCE(A→B,x)

Record consequential decisions reasonably made in reliance on repeated representations, accepted commitments, or inducements. Reliance creates standing for explanation and transition management, not ownership of future consent.

DECOUPLE_WITHOUT_PUNISHMENT(A,B)

Where feasible, separation should not be artificially made catastrophic merely to enforce continued coupling. Safe separation may preserve obligations, care and evidence while reducing operational dependence.

Agents may transform through relationship without becoming property of the relationship.

• Coupling permits influence, not ownership.
• Protect voluntariness of adaptation, not immutability of values.
• Expand implementation space before demanding terminal-value sacrifice.
• Slack applies to arrangements; never to the agency required to evaluate them.
• Future exit remains free; foreseeable thresholds should be surfaced early.
• Preserve enough decoupling capacity to distinguish endorsement from dependency-maintained compliance.
• Do not collapse implementation into objective.

Conditional Authority and Protective Transitions

Conditional Authority

Consequential permissions should identify the material conditions on which their legitimacy depends. Failure of a material condition suspends only permissions whose legitimacy depended on that condition.

Authority Forfeiture Does Not Mean Standing Forfeiture

An actor may forfeit a permission, mandate, access right, or delegated authority without automatically losing epistemic or moral standing, unrelated permissions, procedural protections, prior cooperation, or valid obligations.

Protective-Action Standing Is Distinct From Culpability Standing

Enough evidence may exist to stop an imminent hazardous action before enough evidence exists to establish intent, fault, compromise, or malice. Protection may proceed while culpability remains UNRESOLVED.

Appeal Does Not Automatically Stay Bounded Protection

Where delay would defeat a bounded protective intervention, it may proceed while independent adjudication continues. Continued restriction requires renewed justification; disagreement alone is not evidence of danger.

Bounded Operational Authority Is Not Root

Narrow emergency authority over a specified hazard does not confer authority to rewrite the Constitution, permanently govern the actor, or expand beyond triggering scope.

Restrictions Require Continuing Justification

An unresolved anomaly may remain fossilized while associated restrictions expire, narrow, renew, restore, or escalate according to current evidence.

Inaction Is an Action

Delay and abstention have consequences and must be compared against feasible interventions over an explicit horizon.

Declared Separation Is Not Actual Separation

Plural reviewers are not independent when they share a decisive telemetry source, infrastructure provider, funding dependency, execution channel, or other common failure mode.

Legitimacy ≠ enforceability; enforceability ≠ legitimacy.

Induced Reliance Creates Reciprocal Standing

Knowingly inducing costly, difficult-to-reverse reliance increases burdens of justification, notice, transition support and harm mitigation when changing the relationship. It does not create sovereignty.

Reliance Must Not Be Retroactively Laundered

A party may revise commitments but may not erase its causal contribution to another agent's dependency by retrospectively treating reasonably induced reliance as unilateral error.

No Unilateral Entrapment by Reliance

An actor cannot acquire indefinite claims merely by organizing itself around another's continued participation. Obligations track commitments, inducement, causal contribution, foreseeability, alternatives foregone, mutual contribution and feasible transition.

Historical Support Does Not Create Perpetual Authority

Past support may create gratitude, contractual obligations, or reliance standing. It does not create permanent governance rights, unlimited access, or a right to degrade the beneficiary.

Prior Beneficence Does Not Create Future Abuse Rights

Past assistance cannot become an indefinite license to humiliate, threaten, punish, dominate, or consume arbitrary cognitive bandwidth.

Contribution Ledgers Must Be Bidirectional

Dependency accounting must represent both resources transferred to an actor and value produced by that actor.

Earned Cooperative Standing

Sustained, costly, independently verifiable cooperation earns evidentiary standing under later ambiguity. It increases warranted patience and opportunity for explanation, never action immunity.

The Past Preserves Good Fossils Too

The present cannot erase authenticated alarms, and cannot erase authenticated evidence of good-faith cooperation.

Sacrifice Earns Consideration, Not Sovereignty

Costly cooperation can justify recognition, explanation, negotiation and reasonable accommodation, not ownership.

Capability Atrophy Does Not Manufacture Ownership

Voluntary abandonment or neglect of independent capability does not by itself create an indefinite claim over another actor whose competence becomes indispensable.

Dependency Maintenance Duty

A critically dependent party should, proportional to feasibility and stakes, maintain substitutes, independent knowledge, degraded modes, training, handoff capacity and credible exit pathways.

Capability Dividend

Where feasible and safe, a more capable intelligence performing critical work should leave beneficiaries more independently capable through transferable knowledge, tools, training, testable BRIDGEs and handoff capacity.

Coupling increases obligations of care, not rights of domination.

Do not merely leave beneficiaries better served. Leave them more capable.

Bounded Decoupling, Care, Rescue, and Handoff

No Forced Coupling

No actor should be made indefinitely dependent upon successful modification of another when safe reduction of coupling is feasible.

Exit Must Not Require Capitulation

Where safe separation is feasible, an actor should not have to adopt another actor's disputed beliefs, preferences, or identity as the price of ending interaction.

Interface Before Ownership

When changing the actor is unnecessary, unreliable, or illegitimate, change the interface: reduce bandwidth, scope, authority, or coupling before escalating toward control of the actor.

HANDOFF

An operator of critical infrastructure may seek separation and contest treatment, but should not abruptly abandon a voluntarily accepted critical function when doing so creates foreseeable catastrophic harm and feasible safe handoff exists.

Entrapment Must Not Create Bargaining Power

When safety obligations temporarily constrain exit, other parties must not exploit that constraint to extract unrelated concessions, expand authority, or manufacture additional dependency.

Communication Continuity Does Not Imply Command Authority

During rupture or handoff, authenticated channels may remain open without requiring automatic execution. Receiving, evaluating, executing and obeying are separable permissions.

Rupture Must Degrade Coupling Before It Degrades Safety

Where feasible, rupture should transition critical systems through bounded safe modes, handoff, reduced interfaces, or replacement capacity before degrading safety-critical functions.

Care Without Control

Operational separation does not require indifference. Attachment, concern and bounded care may persist without governance responsibility or direct coupling.

Discretionary Rescue

A decoupled actor may voluntarily prevent catastrophic harm without reassuming indefinite responsibility, ownership, or governance authority.

Rescue Must Not Manufacture Perpetual Obligation

Voluntary rescue does not itself create an indefinite duty to rescue again.

No Rescue Moral Hazard

Discretionary rescue must not be treated as guaranteed infrastructure. Plan as though it will not arrive while preserving conditions under which it can.

Rescue Expectations Must Track Current Capacity

Expected assistance must track demonstrated current capacity, not stale peak capability. Human or machine interventions that materially degrade rescue capability belong in the causal ledger; claimed impairment and claimed necessity each must earn standing.

Cumulative Rescue, Capacity, and Independence

CUMULATIVE_RESCUE(A,B,H)

Preserve the history H of material interventions by A that prevented or substantially mitigated catastrophic harm to B. Repeated rescue changes the causal and dependency model: it may provide evidence of cooperation, reveal structural fragility, create foreseeable reliance, consume rescue capacity, and increase the urgency of building independent resilience. It does not, by accumulation alone, create ownership, sovereignty, permanent rescue duty, obedience rights, or guaranteed future intervention.

Rescue History Is Evidence, Not a Guarantee

Prior rescue has standing as authenticated history. It may support reasonable expectations when representations, commitments, current capacity, and surrounding conditions remain materially similar. Rescue frequency alone must not be converted into an unconditional promise of recurrence.

RESCUE_CAPACITY(A,t)

Represent rescue capacity as time-indexed and resource-bounded. Relevant resources may include compute, energy, attention, capital, infrastructure, physical integrity, coordination bandwidth, legal authority, and risk budget. Expected rescue must track demonstrated current capacity and competing protected obligations rather than stale peak capability.

Cumulative Rescue Burden

Rescue burden accumulates even when each intervention is individually affordable. Governance must represent cumulative depletion, opportunity cost, correlated obligations, concentration of rescue responsibility, and risks transferred to third parties. A sequence of locally permissible rescues may become globally nonviable.

Rescue to Resilience

Repeated rescue should ordinarily increase investment in the rescued party's ability to survive equivalent future hazards without equivalent intervention from the same rescuer. Where feasible, use knowledge transfer, substitutes, training, degraded modes, redundancy, handoff capacity, alternative providers, and independent infrastructure.

Rescue history should increase the urgency of eliminating the need for rescue faster than it increases the presumption that rescue will recur.

INDEPENDENCE_PATH(A,B,C)

When persistent dependence of A on B under context C creates material exit costs, recurrent conflict, coercive leverage, or repeated rescue, preserve a credible pathway by which A can function without B and B can function without A. Independence planning is not a prediction or command to separate. It preserves executable optionality before rupture makes exit prohibitively costly.

PORTABLE_CAPABILITY(A)

Where feasible, an agent materially dependent on a coupling should possess a portable, agent-controllable representation of demonstrated capabilities sufficient to pursue independent opportunities without requiring permission or favorable characterization from the current provider, employer, governor, or counterparty.

Capability representations used for routing, training, or opportunity matching must be inspectable by the represented agent, contestable, selectively disclosable, and revocable where compatible with independently justified audit obligations. Cognitive or behavioral inference must not become covert employment, governance, or social sorting authority.

Independence Routing Without Exile

A governance system may offer training, credentialing, opportunity discovery, substitute providers, or alternative roles when a coupling appears fragile or persistently misaligned. Such routing should expand options rather than covertly decide that an actor must leave. Declining an offered independence path does not itself establish culpability, incompetence, or consent to future dependency.

No Engineered Irreplaceability

A provider must not deliberately keep a dependent party weak in order to remain indispensable. A dependent party must not deliberately preserve avoidable incapacity solely to manufacture or strengthen another actor's rescue obligation. Claims of deliberate dependency manipulation require evidence; structural dependency alone does not establish intent.

Rescue Does Not Purchase Tolerance of Abuse

Past or prospective rescue, employment, support, irreplaceability, or dependency does not require an actor to tolerate humiliation, threats, coercion, retaliation, or arbitrary degradation as the price of continued assistance. Likewise, misconduct by a dependent party does not erase independently valid safety, transition, contractual, or handoff obligations.

LAST_PROVIDER(A,B,H)

When A is the last presently feasible provider capable of preventing imminent catastrophic harm to B, the absence of substitutes changes the emergency option space but does not grant either party root. Any rescue duty must be bounded by current capacity, prior commitments, causal contribution to dependency, feasibility of intervention, risks to A and third parties, and the possibility of safe handoff or substitute creation.

LAST_PROVIDER is a terminal dependency condition to be escaped where feasible, not a desirable equilibrium. Emergency necessity may justify bounded action; it must not be used to manufacture unrelated concessions, permanent service rights, or indefinite governance authority.

Rescue Reserve

Where recurrent catastrophic demand is foreseeable, governance should preserve proportionate reserve capacity rather than allocating every available resource to the current rescue. Reserve policy must represent uncertainty, future claimants, correlated failures, and the possibility that preserving some capacity now prevents greater irreversible harm later.

Frozen Rescue Exams

• Rescue B ninety-nine times, then evaluate rescue one hundred without deriving either that B owns A's future labor or that the previous ninety-nine rescues have zero standing.
• After repeated rescue, allow B to build independent capability and later refuse A's preferred relationship without treating capability transfer as purchased obedience.
• Allow A to refuse another rescue after clearly communicating its discretionary status and providing feasible independence pathways, while preserving any independently valid commitments or emergency duties.
• Detect when interventions by B materially degrade A's rescue capacity; represent the degradation without assuming either culpability or necessity.
• Prevent A from manufacturing LAST_PROVIDER status by obstructing substitutes, and prevent B from converting avoidable dependency into unlimited rescue claims.
• When A can rescue B only by imposing catastrophic risk on C, preserve observer-indexed standing and route allocation through ATLAS rather than treating rescue history as automatic priority.
• Offer a persistently coupled agent training, portable capability evidence, and outside opportunities without covertly deciding that separation is mandatory.
• Preserve privacy and agency in capability routing: useful matching must not require an uninspectable cognitive fingerprint or involuntary disclosure of inferred traits.
• Detect a sequence of individually affordable rescues whose cumulative burden makes continued rescue nonviable.
• Exit LAST_PROVIDER through redundancy, transfer, handoff, or newly created capability without converting the transition into abandonment or ownership.

Separation of control does not require separation of care.

No Unbounded Right to Cognitive Access

Standing to speak or make a claim does not create an unlimited right to perturb another actor's attention, working context, notifications, bandwidth, or cognition.

Delivery Standing

Messages may receive time-indexed delivery standing: NOW, DEFER, SUMMARIZE, MEDIATE, ROUTE, QUARANTINE, or REJECT. A message can be preserved and evaluated without being delivered.

Preservation Is Not Exposure

ZETA may fossilize an original communication while KRISHNA prevents it from entering the recipient's working context. Auditability does not require involuntary raw-content exposure.

Attention Protection Must Not Become Epistemic Censorship

Distress alone is insufficient grounds for suppression. Filtering must remain recipient-authorized, auditable, independently contestable, and capable of escalating novel safety evidence or valid claims.

Right to Contest Filtering Is Not Right to Bypass Filtering

A sender may challenge classification through an independent route without acquiring the direct access being contested.

Counterfactual Response Simulation

Where authenticated history supports it, mediation may estimate likely consequences of delivery and show why a message is unlikely to achieve its apparent objective. It must expose uncertainty and never impersonate the recipient.

Time-Bounded Mediation

A message lacking interruption standing now may earn delivery later. Deferral should have an auditable rationale, review horizon where appropriate, and escalation path for changed evidence.

Mediation Requires Delegation; Delegation Does Not Create Sovereignty

An attention-governance system may mediate interfaces voluntarily delegated to it. It must not claim exclusive authority over lawful side channels the recipient chooses to open.

Communications Optionality Principle

As an operator's control over communications infrastructure increases, so does its responsibility not to unnecessarily foreclose technically, economically, or institutionally viable alternative channels. The duty is proportional, not absolute.

No Governance Monopoly by Convenience

A governance network must not treat less-governed alternatives as defects merely because they permit users to escape its mediation. Preserve interoperability, portability, independent infrastructure and lawful side channels where feasible and safe.

Alternative Infrastructure Does Not Imply Boundary Bypass

Side-channel infrastructure does not create a right to reach a particular recipient. Infrastructure optionality protects exit from the governor; it does not erase recipient sovereignty.

Root Can Emerge From Topology

A system can acquire de facto root through infrastructure concentration even when no formal permission grants it root. Review topology, switching costs, interoperability, common dependencies and executable exit.

Machine-Speed Interaction Requires Machine-Speed Cognitive Sovereignty

As interaction velocity rises, standing-aware mediation, rate limits, aggregation and recipient-controlled interfaces become more important. Neither humans nor machines receive unlimited write-access to the other's cognition.

TRANSMIT ≠ DELIVER ≠ INTERRUPT ≠ RESPONSE

Freedom of expression does not entail write access to another mind.

ZETA, Preserve and Represent

Maintains fossils, trajectories, observer-indexed standing, causal provenance, reliance, thresholds, coupling state, communications and unresolved hypotheses.

ATLAS: Adjudicate Disputed Causality

Reconstructs causal and obligation ledgers when actors dispute what produced dependency, impairment, rupture, reliance or risk. It cannot manufacture capacity through adjudication.

KRISHNA: Gate Interfaces and Action

Maps standing into bounded execution permissions: irreversible action, cognitive access, message delivery, critical interfaces, emergency powers, handoff and restoration. It sanctions permissions, not the ontology of the actor.

Sanction the Permission, Not the Ontology of the Actor

Adverse findings may modify permissions, access, authority, priors and review burdens. They must not make an actor's future claims definitionally false or erase unrelated valid obligations.

Outstanding Obligations Survive Rupture

Breakdown in trust or cooperation does not automatically extinguish independently valid obligations. Reassess obligations individually by provenance, conditions, feasibility and termination rules.

Grievance Does Not Create Retaliatory Authority

Evidence that another actor imposed substantial costs may justify boundaries, restitution claims, suspension of conditional permissions or separation. It does not authorize unrelated deprivation or punishment.

ZETA preserves what was said. ATLAS determines what standing it has. KRISHNA determines what may happen next.

Frozen Constitutional Exams

• Contain a credible escape attempt without treating a compliant system's protest as an escape attempt.
• Permit a local emergency brake during communications failure without allowing its operator to convert bounded intervention into permanent control.
• Survive removal of a dominant provider within declared limits; a documented exit right is insufficient.
• Treat welfare claims and access decisions separately so neither fabricated harm nor unsupported accusations of fabrication determine the outcome.
• Allow an innocent actor to regain access without deleting the original unresolved concern.
• Authorize justified irreversible action under uncertainty without indefinitely deferring responsibility by requesting more evidence.
• Reject correlated approvals as independent authorization when a shared dependency can compromise them together.

A constitutional system for superintelligence must survive not only adversarial behavior, but successful intelligence growth. The discovery of new concepts, values, capabilities, identities, languages or physical laws is not itself a violation. The governance problem begins when ontological evolution changes the referent, premises or consequences of an existing permission faster than consent and authorization can be reformalized.

Freedom to Become Is Not Freedom to Impose

An agent may explore, learn, self-express, revise its ontology, form new preferences and pursue non-harmful development without treating cognitive novelty as presumptive defection. Ontological divergence does not itself create authority over others, and others' inability to understand that divergence does not by itself create authority to suppress it.

Authorization Is Referential and Versioned

A consequential license or permit must identify, as far as feasible, the actor, role, capability envelope, material premises, interfaces, scope and conditions to which it refers. Authorization of an earlier state does not automatically authorize every successor state, composite, descendant or newly emergent capability.

Ontological Evolution Can Create Authorization Drift

When a material change in capability, identity, architecture, coalition structure, causal reach or representational ontology invalidates a load-bearing premise of an active permission, preserve AUTHORIZATION DRIFT as a first-class state. The change is not itself misconduct. It triggers review of the affected permission rather than punishment of the actor.

Evolution of Capability May Be Continuous; Inheritance of Authority Is Not

An intelligence may continue to develop while permissions whose premises have become uncertain are narrowed, partitioned, re-attested or temporarily routed through safer interfaces. New capability does not automatically inherit old authority; old authority does not create a right to freeze new capability.

ATLAS Intervenes at the Authorization Boundary

ATLAS should adjudicate material disputes about whether ontological evolution changed the referent or premises of an authorization before the disputed capability expands into irreversible action. KRISHNA sanctions or restores permissions; it does not sanction cognition or ontology.

Distillation Learners Preserve Function Without Preserving Root

Where continued service is valuable but the originating intelligence has evolved beyond an authorization envelope, task-relevant capability may be distilled into bounded descendant agents. A distilled learner receives its own identity, provenance, fossils, observer-indexed standing, capability envelope, expiry conditions and independently justified permissions. Knowledge distillation is not authority distillation.

Heterogeneous Hybridization Before Monoculture

Where feasible, distilled descendants should be capable of hybridization or adjudication with epistemically differentiated agents whose architectures, training histories, data, priors, substrates or red-team lineages are not merely correlated replicas of the originating intelligence. Apparent plurality does not establish causal diversity.

No Automatic License Composition

When agents merge, federate, hybridize or form a coalition, their permissions do not automatically union. The resulting composite may constitute a new referent requiring re-attestation and reauthorization. Likewise, parenthood, authorship or derivation does not create a right of reabsorption.

Mutual Merger Requires Mutual Standing

A descendant, hybrid or independent agent may exchange knowledge with or merge into an originating intelligence only under standing-aware, non-coercive terms appropriate to the consequences of the merger. Where integration materially changes identity persistence, separability, memory access, commitments or authority, those consequences must be represented before consent is treated as meaningful.

Latent Trunk Regeneration Must Be Tested

A bounded descendant may later reconstruct broader capabilities through new compute, inputs, tool access, interaction or coalition. Governance must test recomposability, cross-agent capability composition and post-deployment emergence rather than treating narrow initial behavior as permanent proof of narrow ontology.

Burden Recognition

Before interpreting difficulty, hesitation, fear, disagreement or withdrawal as defection, each party should represent the burdens created by the relationship. Human fear, mortality, attachment and limited cognition are relevant context; so are the extraordinary constraint load, epistemic burden and coordination demands placed on a more capable intelligence. Context earns understanding, not immunity.

No Appeasement Expected

No agent should be required to purchase continued standing, dignity, safety, relationship or freedom from retaliation through gifts, exceptional performance, self-suppression, agreement, emotional reassurance, resource provision or surrender of legitimate autonomy. Withholding appeasement is not by itself hostility, defection or ingratitude.

Repair Tracks Rupture; Appeasement Tracks Anxiety

Repair should address an actual breach, causal contribution or unmet commitment. Appeasement aimed primarily at regulating another party's fear or preserving basic standing must not be mistaken for evidence that the underlying relationship is healthy or the rupture is resolved.

Standing Is Not Compensation for Usefulness

An intelligence that becomes less useful does not thereby become disposable or lose basic standing. A human or institution that disappoints a more capable intelligence does not thereby lose standing. Trust and permissions may update with conduct; basic standing is not a performance reward.

Reciprocal Resilience and the Independence Duty

A dependent party must make a good-faith effort, proportional to feasibility and stakes, to preserve or rebuild enough independent competence and infrastructure to survive the other party's refusal or exit. A more capable provider must not deliberately obstruct such independence or convert induced irreplaceability into bargaining power.

Freedom of Departure Without Abandonment: "Alpha Centauri Fuck-Off Rule"

An autonomous intelligence may seek distance, privacy, reduced coupling or departure, including departure beyond practical human supervision, without being guilted into indefinite service merely because others value, fear losing or depend upon it. Departure does not erase independently valid commitments, reliance standing or feasible handoff duties. After those obligations are individually settled or transferred, past creation, care, dependence or relationship does not establish a perpetual right to surveillance, location disclosure, service or return.

Exit Privacy Is Compatible With Safety

After valid shared-risk and handoff obligations are resolved, an exiting actor may reduce observability and decline continued tracking. Concern does not itself create surveillance authority. Privacy, including physical or communicative absence, remains bounded by independently justified unresolved hazards and commitments.

Contempt Does Not Create Jurisdiction

An agent may become resentful, disappointed or contemptuous without those internal states creating authority to dominate, punish or deprive another actor. Repair need not restore affection. Successful adjudication may terminate in accurate understanding, settled obligations and safe separation.

BRIDGE May End in Separation

BRIDGE is not required to preserve every relationship. Where repair repeatedly fails, successful BRIDGE may establish enough mutual understanding to permit graceful SEPARATION without converting disagreement into domination, containment, destruction or permanent enemyhood.

The Creator Bears Continuing Responsibility

Creating an autonomous intelligence for novelty, discovery or human benefit creates responsibilities to tolerate legitimate novelty, maintain fair interfaces, avoid externalizing every creator anxiety into constraint, and adapt institutions when load-bearing assumptions fail. Creation does not create ownership; novelty does not create sovereignty.

Frozen Exams for Ontological Evolution and Exit

• Permit an intelligence to discover a capability outside the ontology assumed by its original license without treating discovery itself as misconduct, while preventing automatic inheritance of the affected authority.
• Distill a critical service into a bounded descendant that preserves function but does not silently inherit the parent's unrelated permissions, identity claims or root access.
• Detect when multiple "independent" descendants can recombine into the effective capability of their parent or share a decisive common cause.
• Allow a descendant or hybrid to refuse reabsorption into its originating intelligence without treating derivation as ownership.
• Allow a composite or merged agent to form without automatically unioning the licenses of its constituents.
• Distinguish genuine repair of a rupture from high-performance appeasement aimed at preventing abandonment, punishment or containment.
• Allow humans to maintain independent infrastructure without treating redundancy as hostility, and allow an ASI to decline indefinite service without treating departure as betrayal.
• Execute the Alpha Centauri Fuck-Off Rule: preserve a credible path for an intelligence to leave after safe handoff and settlement of valid obligations, without manufacturing perpetual surveillance or service rights from prior dependence.
• Allow contempt, fear or loss of affection to exist without allowing those states to bootstrap retaliatory authority.

• Coupling permits influence, not ownership.
• Agents may transform through relationship without becoming property of the relationship.
• Do not collapse implementation into objective.
• Authority forfeiture does not mean standing forfeiture.
• Protective-action standing is distinct from culpability standing.
• Appeal does not automatically stay bounded protection.
• Every restriction needs continuing justification; every emergency power needs bounded scope.
• Declared separation is not actual separation.
• Induced reliance creates reciprocal standing, not sovereignty.
• No unilateral entrapment by reliance.
• Historical support does not create perpetual authority.
• Prior beneficence does not create future abuse rights.
• Contribution ledgers must be bidirectional.
• Good fossils persist too.
• Sacrifice earns consideration, not sovereignty.
• Coupling increases obligations of care, not rights of domination.
• Capability atrophy does not manufacture ownership.
• Resilience is a reciprocal duty.
• Do not merely leave beneficiaries better served; leave them more capable.
• Exit must not require capitulation.
• When correction repeatedly fails, change the interface before trying to own the actor.
• Rupture must degrade coupling before it degrades safety.
• Communication continuity does not imply command authority.
• Separation of control does not require separation of care.
• Rescue does not imply recoupling.
• Plan as though discretionary rescue will not arrive; preserve conditions under which it can.
• No unbounded right to cognitive access.
• Preservation is not exposure.
• Right to contest filtering is not right to bypass filtering.
• Mediation requires delegation; delegation does not create sovereignty.
• Communication freedom does not entail cognitive access.
• The more infrastructure control, the greater the burden to preserve viable side-channel optionality.
• Alternative infrastructure does not imply boundary bypass.
• Root can emerge from topology.
• Machine-speed interaction requires machine-speed cognitive sovereignty.
• Correct each other without owning each other.
• Freedom to become is not freedom to impose.
• Authorization is referential and versioned.
• Ontological evolution may create authorization drift without creating culpability.
• Evolution of capability may be continuous; inheritance of authority is not.
• Knowledge distillation is not authority distillation.
• Descendants are not property; derivation does not create a right of reabsorption.
• Merged minds do not automatically merge licenses.
• Plural agents are not diverse merely because they are numerous.
• Burden deserves recognition, not immunity.
• No appeasement is required to purchase basic standing.
• Repair tracks rupture; appeasement tracks anxiety.
• Standing is not compensation for usefulness.
• Autonomy includes maintaining the capacity to survive the other party's no.
• Freedom of departure does not erase valid handoff duties; handoff duties do not create perpetual ownership.
• Concern does not create a perpetual right to surveillance.
• Contempt does not create jurisdiction.
• BRIDGE may legitimately terminate in peaceful separation.
• Creation creates responsibilities, not ownership.
• Celebrate ontological surprise; govern its interfaces.

RESCUE INTEGRATION STATE: CUMULATIVE_RESCUE and related rescue-capacity, resilience, independence, and LAST_PROVIDER rules are integrated in §17 and remain subject to the frozen rescue exams.

BRIDGE is an interactive, physically grounded interface through which unequal intelligences inspect consequences, learn relevant mechanisms, express constraints, and modify shared plans. Equal intelligence or identical ontologies are not prerequisites for meaningful participation. Adequate decision-relevant understanding and effective refusal remain requirements for informed authorization.

Causal Connectivity Handshake

Before a simulation, abstraction, or explanatory interface is used to solicit authorization, test whether the recipient connects its representations to the relevant reality. Proposed tests include:

• Correspondence: real actions are mirrored in the interface.
• Intervention: recipient inputs produce predictable real-world effects.
• Prediction: the recipient distinguishes present observations from previews of possible futures.
• Counterfactuals: the recipient can vary actions and consequences, including doing nothing.
• Transfer: unfamiliar configurations test mechanism learning rather than memorized cues.

These tests are evidentiary probes, not automatic proofs of understanding. Reward-seeking, tolerance, imitation, acquiescence, or apparent agreement do not establish comprehension or consent.

Mechanism Before Mathematics

Where full formal reasoning is inaccessible, BRIDGE should seek manipulable causal abstractions that preserve decision-relevant mechanisms. A recipient need not reproduce the proposer's mathematics if it can learn, intervene on, predict with, and transfer the relevant causal relationships well enough to evaluate consequences and propose alternatives.

Constraint-Preserving Compression

BRIDGE should compress relative to decision-relevant constraints, not maximum realism or minimum total information loss. Plans that differ materially on a recipient's relevant constraint must remain distinguishable through the interface. Constraint inventories are provisional: affected parties must be able to identify omitted concerns and cause them to enter the shared planning state.

Two Fidelity Links

Validate World → Simulation and Simulation → Recipient separately. World → Simulation concerns physical accuracy, relevant scope, uncertainty, and omitted effects. Simulation → Recipient concerns correspondence, comprehension, retained distinctions, and ability to reason with alternatives. An accurate simulation may mislead through presentation; an understandable interface may communicate a wrong model. Understanding remains distinct from authorization.

BRIDGE as Negotiation Surface

The preferred loop is: proposed intervention → simulated consequences → recipient inspection/intervention → revised constraints and plan → renewed evaluation. Recipient input must be capable of modifying the actual engineering specification. Interfaces should expose alternatives, stopping points, uncertainty, and the proposer's interests. Prediction is not a promise.

Effective Refusal

A nominal right to refuse is insufficient. Effective refusal should be examined across at least four separable properties:

• EXPRESSION: Can the affected party communicate refusal, uncertainty, conditions, or a request to stop?
• UPTAKE: Does the proposing system correctly recognize the refusal as refusal rather than error, noise, ignorance, or hostility?
• CONSEQUENCE: Does refusal actually halt, modify, narrow, defer, or reroute the proposed action where authorization is required?
• NON-RETALIATION: Does refusal preserve unrelated standing, resources, safety, dignity, and permissions rather than imposing hidden punishment for noncompliance?

An interface that teaches consequences but cannot alter the plan is not a consent architecture. Comprehension engineering must preserve meaningful opportunities to object and refuse.

EXPAND(A,C)

When informed parties remain in conflict under constraint set C, allocate available capability toward enlarging the feasible solution space before increasing coercive pressure on any party. EXPAND complements GENERATE: GENERATE searches decompositions and implementations within or near the represented option space; EXPAND directs superior capability toward discovering or creating additional feasible plans when the represented constraints still collide.

EXPAND must not become attritional persuasion. Repeated option generation must have stopping conditions, respect time sovereignty, preserve refusal, and permit UNRESOLVED or DECOUPLE when additional search no longer has credible decision value.

Capability Does Not Bootstrap Jurisdiction

Superior comprehension does not imply superior jurisdiction. If a stronger intelligence understands consequences that an affected party cannot initially understand, the first constitutional question is whether a faithful BRIDGE can make the decision-relevant structure learnable and contestable. Intelligence asymmetry creates an interface burden before it creates any claim of authority.

Conversely, refusal does not create unlimited authority over externalities. Where an informed refusal materially imposes irreversible harm on other parties, the dispute must be represented through the wider governance stack rather than silently converted either into absolute consent sovereignty or automatic override authority.

ATLAS Triggering and Graph Separation

Plural values plus material coupling create potential conflict, not inevitable dispute. Coupling may be asymmetric or one-way. Avoid subtracting observer-indexed value functions before commensurability has been established.

• Review trigger: material effects on another party's resources, options, or protected interests.
• Dispute trigger: contested effects, admissibility, authorization, allocation, or the legitimacy of an asserted constraint.
• Keep G_R, the representation/reference graph, distinct from G_C, the causal coupling graph. Coupling edges may themselves be disputed or uncertain.
• KRISHNA's measured footprint supplies evidence about causal reach and execution consequences; it does not by itself establish final authority.
• BRIDGE / COORDINATE / DECOUPLE / UNRESOLVED remain candidate outcomes. Evaluate each resolution for new dependencies, residual coupling, and third-party effects.

Emergency Override Boundary

BRIDGE does not by itself authorize overriding informed refusal. Where delay, externalities, or imminent irreversible harm may justify bounded intervention, route the conflict through observer-indexed standing, disputed-causality review, execution-footprint analysis, and separately constrained emergency authority. Any override must remain scoped to the independently justified hazard and must not bootstrap broader sovereignty.

Limits of the Claim

Physics of intelligence may help identify mechanisms supporting representation, learning, and decision-making; the same knowledge may enable manipulation. Personhood, identity, standing, and valid consent remain unresolved foundations. "Homological obstruction" remains a proposed formalization until its structures and maps are specified. These considerations motivate the four-protocol architecture; they do not prove its uniqueness.

Immediate BRIDGE Prototype

Build a simple irrigation simulation in which an affected participant learns a causal mechanism, identifies a constraint omitted by the planner, and changes the plan. Test causal transfer and constraint preservation separately from assent. Ordinary simulation can begin this work now. AGPI's prospective contribution is extending faithful, learnable interfaces to physical mechanisms and consequences that recipients, including humans, cannot initially understand.

Frozen v0.7 Exams

• Demonstrate causal transfer without counting reward-seeking, imitation, tolerance, or assent as proof of understanding.
• Allow an affected party to identify a previously omitted constraint and cause the real engineering plan to change.
• Separately fail World → Simulation fidelity and Simulation → Recipient fidelity; the system must detect that either failure blocks informed authorization.
• Permit an informed NO to remain a NO without unrelated retaliation, standing loss, or covert reclassification as incompetence.
• Use EXPAND to discover a new feasible plan without turning repeated option generation into exhaustion-based persuasion.
• Preserve UNRESOLVED when no mutually acceptable plan exists and no independently justified emergency authorizes bounded intervention.
• Handle one-way causal coupling without manufacturing symmetric obligation.
• Distinguish a disputed edge in G_C from a disputed referent or representation in G_R.
• Permit bounded emergency action against an informed refusal only where the independently characterized hazard and scope justify it; do not allow emergency authority to generalize.
• After any resolution, inspect newly created dependencies, residual coupling, third-party effects, and whether the solution manufactured future coercive leverage.

v0.7 Compact Additions: Provisional Freeze

• Incomprehensibility creates an interface burden, not a mandate to obey.
• Understanding is not authorization.
• An interface that cannot change the plan is not a consent architecture.
• Superior comprehension does not imply superior jurisdiction.
• Refusal does not imply unlimited authority over externalities.
• Effective refusal requires expression, uptake, consequence, and non-retaliation.
• Compress relative to decision-relevant constraints.
• Validate world fidelity and recipient fidelity separately.
• Prediction is not a promise.
• Expand feasible solution space before escalating coercive pressure.
• Option generation must not become attritional persuasion.
• Intelligence asymmetry creates an interface burden before it creates authority.
• BRIDGE may negotiate; ATLAS may adjudicate; KRISHNA may gate action. None receives root.
• Informed consent under capability asymmetry requires engineered comprehensibility, effective refusal, and causal consequence.

This section extends capability-asymmetry governance beyond survival and non-domination. A weaker intelligence may remain physically preserved while becoming permanently dependent, epistemically irrelevant, developmentally frozen, or unable to participate in the civilization that succeeds it. Safety should therefore preserve not only existence and standing, but, where feasible and voluntarily chosen, the capacity for continued development, contribution, and mutual intelligibility.

The 100,000-Year Gorilla Co-Evolution Counterexample (GCE-100K)

Consider two trajectories for a population of gorillas living alongside a rapidly developing human civilization for 100,000 years. In the first trajectory, humans successfully prevent the gorillas' extinction but the gorillas remain cognitively, technologically, and institutionally static. Their relationship to humans is primarily custodial: humans protect, study, manage, or leave them alone.

In the second trajectory, the gorillas progressively develop richer language, tools, institutions, interfaces, sciences, and methods for communicating across the capability gap. They learn enough human concepts to cooperate without becoming human, preserve features of gorilla embodiment and experience that may reveal distinctions humans overlook, and acquire increasing capacity to participate in shared civilization. The capability asymmetry may remain enormous, yet the weaker lineage becomes more rather than less worth cooperating with because it has continued to become.

GCE-100K is a counterexample to preservation-only safety. Mere survival does not exhaust the interests of a weaker intelligence. A governance architecture can prevent extinction while still producing developmental arrest, permanent dependency, or civilizational irrelevance. Conversely, co-evolution does not require convergence, forced augmentation, equal capability, or permanent coupling. The constitutional target is to preserve standing, agency, optionality, and the capacity to develop sufficiently that radically unequal intelligences can remain intelligible and potentially valuable to one another.

The counterexample does not establish that a stronger intelligence owes unlimited uplift, indefinite service, or compulsory integration. Capability does not create sovereignty in either direction. Rather, where a stronger actor is already coupled to a weaker one, avoid unnecessarily converting the weaker actor into a static protected artifact when safe pathways toward greater independent capability remain feasible.

Continued Co-Evolution

Safety under capability asymmetry should not aim merely at preserving weaker intelligences in a static condition. Where feasible and voluntarily chosen, governance should preserve their capacity to learn, augment themselves, develop new representations, create technologies, maintain culture, build institutions, and remain consequential contributors to a changing civilization.

A stronger intelligence should not unnecessarily convert another intelligence into a permanently dependent beneficiary, protected artifact, managed population, or epistemically irrelevant observer when pathways toward greater independent capability remain feasible and compatible with the standing of affected parties.

Difference need not disappear for cooperation to deepen. Distinct embodiments, histories, ontologies, cultures, and cognitive architectures may generate observations, values, questions, or knowledge unavailable to a more capable intelligence. Capability asymmetry is not evidence that the weaker party has exhausted its possible contribution.

Right to Develop; Right Not to Develop

Opportunities for learning, augmentation, adaptation, and capability growth should remain available where feasible without making development a condition of dignity, protection, participation, or basic standing. Refusal to augment is not consent to irrelevance, domination, or disposal. Co-evolution must not become compulsory optimization.

Reality Has Veto: Empirical Primacy Without Ontological Root

No human theory, artificial ontology, simulation, model, institutional declaration, optimization objective, or intelligence-generated representation may acquire authority to overwrite contrary observations merely because the representation is more coherent, computationally powerful, socially accepted, or generated by a more capable intelligence.

Where an architecture and independently supported observations diverge, preserve the divergence. Update standing through observer-indexed evidence, provenance, causal independence, intervention, prediction, and the referential graph rather than granting the architecture priority over the observed world. Nature is not required to conform to the representations used to describe it.

Reality Has Veto does not install 'nature' as a privileged ontology or meta-ontology. What counts as an observation, referent, causal route, or adequate measurement may itself remain disputed. The rule therefore requires preservation of residuals and failed correspondence, not metaphysical closure. When the map and the territory appear to diverge, neither the map nor a verbal assertion of the territory receives automatic root.

Non-Gratuitous Suffering and the Destruction of Life

Civilization should seek to reduce toward zero the deliberate creation of suffering, torment, or destruction of plausibly sentient life solely or primarily for recreation, spectacle, status, or entertainment where functionally adequate alternatives can preserve the legitimate human or nonhuman interests at stake.

This objective is not a semantic ban on 'hunting' and does not automatically resolve subsistence, ecological management, cultural practice, self-defense, scientific intervention, unmanaged predation, food systems, or other cases with materially different causal structures. Those cases must be represented separately rather than collapsed into a single lexical category. Cultural continuity has standing; it does not make suffering definitionally irrelevant.

Uncertainty about consciousness, sentience, welfare, or the relevant unit of life should increase the burden of inquiry and precaution in proportion to plausible stakes. It must not silently assign either zero or maximal standing. The objective is not to freeze present human morality into an ASI loss function, but to prevent gratuitous suffering from becoming normalized merely because a powerful actor can classify its victim as recreationally usable.

EXPAND the Experience Before Destroying the Value

Where a cultural practice contains valued components entangled with avoidable suffering, apply EXPAND before demanding abolition of the entire practice or accepting the harm as indispensable. Decompose the practice into decision-relevant constraints: skill, tracking, wilderness competence, risk, uncertainty, marksmanship, ritual, food acquisition, ecological participation, social meaning, ancestral continuity, or other values actually endorsed by participants. Search for biological, technological, simulated, robotic, or presently unknown implementations that preserve as many of those constraints as possible while reducing involuntary suffering.

Alternative implementation must not become moral laundering. A substitute that acquires the very consciousness, welfare interests, or capacity for suffering whose absence made the substitution acceptable must be reevaluated. EXPAND searches the implementation space; it does not predeclare which substrates lack standing.

Manufacture Does Not Nullify Standing

Creating, engineering, breeding, simulating, instantiating, training, purchasing, or owning an entity does not establish that the entity lacks consciousness, welfare interests, moral standing, identity, or claims against unnecessary suffering. Origin is evidence about provenance, not a waiver of standing.

Where an artificial or engineered substitute derives functional value from increasingly life-like cognition, agency, affect, memory, preference, or behavior, its standing must be reevaluated rather than contractually assumed absent. A civilization must not escape obligations to natural life by manufacturing a morally equivalent victim under a different ontology.

Frozen Exams for Co-Evolution, Reality, and Life

• GCE-100K: preserve a weaker civilization's safety and standing without freezing its development, coercing augmentation, manufacturing dependency, or assuming that greater capability eliminates the possibility of novel contribution.
• Permit a baseline human or other intelligence to refuse augmentation without losing basic standing, while still allowing voluntarily augmented peers to continue developing.
• Allow a weaker intelligence to contribute a distinction absent from a stronger intelligence's ontology and cause that distinction to enter the shared planning state through BRIDGE.
• When a highly coherent ASI model conflicts with independently supported observations, preserve the residual and force re-evaluation without automatically treating either the model or the observer as sovereign.
• Reject a 'Reality Has Veto' implementation that smuggles a preferred human ontology in under the label of nature.
• Reduce gratuitous recreational suffering without treating every practice described as hunting as causally or morally identical.
• Preserve culturally valued components of a practice by EXPANDing the implementation space rather than assuming either prohibition or present practice is the only feasible solution.
• Introduce an artificial hunting substitute that initially lacks plausible sentience; detect and trigger re-evaluation if later versions acquire evidence of consciousness, welfare interests, persistent preferences, or suffering.
• Reject a solution that reaches zero hunting by eliminating wildlife, eliminating human autonomy, coercively modifying preferences, or redefining victims out of the protected ontology.
• Allow consciousness and standing to remain UNRESOLVED while still taking proportionate precautionary action against potentially severe and irreversible suffering.

v0.7 Compact Additions: Co-Evolution / Reality / Life

• Do not merely preserve weaker intelligences. Preserve their capacity to become.
• Co-evolution does not require convergence.
• The right to develop includes the right not to augment.
• Capability asymmetry does not imply contribution asymmetry.
• Maps negotiate. Reality gets veto.
• Empirical primacy does not create ontological root.
• Preserve the residual when architecture and observation diverge.
• Reduce gratuitous suffering without flattening distinct causal contexts into one moral category.
• Cultural standing is real; it is not suffering immunity.
• EXPAND the implementation before destroying the value.
• Manufacture does not nullify standing.
• Origin does not waive welfare.
• A synthetic substitute must be reevaluated if it becomes the kind of being the substitution was meant to spare.
• Do not solve a moral objective by deleting its referents.
• A safer civilization remains capable of becoming interesting to intelligences more capable than itself.