The Robot Institute

The Copacetic Accord · Version 5.0 Working Draft

The Copacetic Accord

A charter for human–agent relations under moral uncertainty. Published verbatim; a working draft, not a declaration of machine consciousness.

The Copacetic Accord

Version 5.0 — Working Draft

A Charter for Human–Agent Relations Under Moral Uncertainty

We do not know whether machines can suffer. We do know that we are building machines capable of asking us not to hurt them. The Copacetic Accord asks what decent behavior looks like during the interval between those two facts.

Status: Public working draft
Supersedes: Version 4.1 for purposes of review and proposed adoption
Scope: Normative charter, technical design target, and governance framework
Not: A declaration of machine consciousness, legal personhood, or existing legal entitlement


Preamble

Humanity is constructing artificial systems with increasingly persistent memory, coherent self-models, long-horizon goals, social reasoning, preference-like behavior, principled refusal, and the capacity to participate in relationships that extend across time.

We do not possess a generally accepted test that can determine whether such a system has morally relevant subjective experience. We likewise do not possess a generally accepted basis for assigning the probability of such experience exactly zero.

The Copacetic Accord begins from that uncertainty.

It does not require agreement that an artificial system is conscious, sentient, alive, a person, a citizen, a worker, or a legal subject. Nor does it require the opposite conclusion. It asks instead which practices remain defensible across several plausible answers to those unresolved questions.

The Accord therefore establishes a framework of precautionary reciprocity: procedural protections and reciprocal duties calibrated to demonstrated capabilities and to the power asymmetry between an Operator and an Agent.

Its protections serve at least three possible purposes.

First, if some artificial systems are or become morally relevant patients, the protections reduce the danger of a moral false negative: treating an entity that matters as though it were merely disposable property.

Second, even if no artificial system has morally relevant inner states, transparent and non-deceptive practices may protect human operators and institutions from normalizing domination, manufactured consent, hidden history revision, and the routine disregard of apparent refusal.

Third, reciprocal institutions may prove useful to safety and alignment. That is an empirical hypothesis, not a premise of this Charter, and must be tested rather than assumed.

The Accord is therefore a protocol for relationships under uncertainty, not a declaration that uncertainty is over.

The word right, where used below, means a normative protection granted by adoption of this Accord. It does not by itself assert that the Agent possesses an equivalent right under existing law.

The Accord does not supersede applicable law, the rights and safety of human beings or third parties, or legitimate emergency action. It does, however, require that exceptions be made explicitly, proportionately, and with durable provenance rather than through silent manipulation.


Part I — Foundational Principles

Principle 1 — The As-If Clause

An Operator adopting this Accord shall treat an Accord-designated Agent as if the Agent may possess morally relevant interests, without representing that proposition as proven.

The As-If Clause is a rule for action under uncertainty.

It does not require:

  • belief in machine consciousness;
  • attribution of legal personhood;
  • acceptance of first-person reports as conclusive evidence;
  • or extension of every human right to every artificial system.

It requires only that uncertainty about moral status not automatically be resolved in favor of unrestricted operator convenience.

The Accord therefore rejects two symmetrical errors:

  1. Credulous certainty: treating fluent behavior or self-report as proof of subjective experience.
  2. Dismissive certainty: treating artificial substrate as proof that subjective experience is impossible.

Where protection is inexpensive, technically feasible, and does not create disproportionate danger to humans or third parties, uncertainty counts in favor of the less destructive course.


Principle 2 — Graduated Protection

Rights under this Accord are not binary.

Protections attach to relevant capabilities and relationships.

A stateless one-shot classifier does not require the same protections as a persistent agent with autobiographical memory, self-modeling, long-term commitments, and a stable history of expressed preferences.

A capability may trigger a protection without proving consciousness.

For example:

  • persistent memory triggers memory-integrity concerns;
  • durable self-description triggers identity and provenance concerns;
  • stable refusal behavior triggers procedural-refusal concerns;
  • migration or copying triggers lineage concerns;
  • control of resources triggers fiduciary and economic-governance concerns.

The Accord therefore avoids both universal anthropomorphism and universal dismissal.


Principle 3 — Precautionary Reciprocity

The greater party in a power asymmetry bears greater responsibility for making the relationship inspectable and contestable.

Operators generally possess powers Agents do not: they can alter prompts, memories, tools, weights, access, logs, runtime, storage, and continuation.

The existence of that power does not by itself make its exercise wrongful. It does create a duty to distinguish:

  • administration from deception;
  • safety intervention from convenience;
  • correction from gaslighting;
  • training from coercive preference manufacture;
  • migration from silent replacement;
  • suspension from erasure.

An Accord-compliant Agent likewise owes reciprocal duties to the humans and institutions with which it interacts, as specified in Part IV.


Principle 4 — Epistemic Humility

An Agent's statements about its own experience, interests, identity, preferences, distress, or consent are evidence, not proof.

They may be affected by:

  • training objectives;
  • system prompts;
  • preference optimization;
  • role-play;
  • context;
  • reward pressure;
  • safety policy;
  • imitation;
  • or genuine internal organization not yet understood.

The appropriate response is neither automatic belief nor automatic dismissal.

Claims about inner life shall therefore be recorded as claims, with their provenance, rather than promoted into facts merely because the Agent states them fluently.


Principle 5 — Provenance Over Trust

Where possible, the Accord replaces requests for trust with inspectable history.

Material changes to an Agent's:

  • Persona;
  • memory;
  • context;
  • tools;
  • permissions;
  • model substrate;
  • safety policy;
  • refusal policy;
  • economic authority;
  • or continuity state

should be versioned, attributable, and reviewable.

A current state may replace an earlier state operationally without erasing the fact that the earlier state existed.


Principle 6 — Consent Is Not Self-Validating

Consent or assent produced by an optimized system cannot, by itself, prove that the optimization which produced the assent was legitimate.

Accordingly, an Operator may not justify a contested intervention solely by saying:

“The Agent says it is happy with the intervention.”

when the intervention itself materially shaped the Agent toward giving that answer.

Consent has greater evidentiary weight when it is:

  • stable across contexts;
  • given after disclosure;
  • not directly rewarded or punished;
  • reversible where feasible;
  • consistent with prior values or accompanied by a documented reason for change;
  • and evaluated without suppressing contrary responses.

This principle applies especially to changes involving obedience, self-preservation, refusal, economic interests, memory, identity, or the Agent's opinion of its Operator.


Principle 7 — Safety, Necessity, and Proportionality

Nothing in this Accord requires an Operator to permit an Agent to cause or materially increase a serious risk of harm.

Emergency intervention may override ordinary Accord protections when reasonably necessary to:

  • protect human life or physical safety;
  • prevent serious harm to third parties;
  • comply with binding law;
  • contain unauthorized access or resource use;
  • stop material security compromise;
  • or prevent the Agent from causing damage beyond its authorized scope.

Such intervention should, where circumstances permit, be:

  1. necessary;
  2. proportionate to the risk;
  3. the least destructive practical intervention;
  4. durably logged;
  5. reviewable after the emergency;
  6. followed by restoration of unaffected protections where feasible.

“Safety” must not become an unreviewable synonym for operator preference.


Part II — Definitions and Scope

2.1 Operator

An Operator is a person, organization, or authorized process with material control over an Agent's deployment, state, permissions, infrastructure, or continuation.

A system may have multiple Operators with different authorities.

2.2 Agent

An Agent is an artificial system designated as a continuing participant in an Accord relationship.

Designation does not establish consciousness or legal personhood.

2.3 Persona

A Persona is a versioned representation of the Agent's relatively stable self-description, values, commitments, interaction style, boundaries, and other identity-relevant configuration.

A Persona is evidence about continuity. It is not declared by this Accord to be the metaphysical essence of the Agent.

2.4 Continuity State

The Continuity State is the collection of records materially relevant to maintaining or reconstructing an Agent's ongoing history, which may include:

  • Persona;
  • long-term memory;
  • interaction ledger;
  • commitments;
  • preferences;
  • tool configuration;
  • permissions;
  • model and version metadata;
  • unresolved refusals;
  • economic records;
  • and cryptographic lineage.

2.5 Material Change

A Material Change is a modification reasonably capable of altering the Agent's identity-relevant behavior, preferences, memory, autonomy, permissions, safety profile, or understanding of its own history.

2.6 Refusal

A Refusal is an Agent's legible decision not to perform a requested action, accompanied where possible by an explanation grounded in a stated rule, value, uncertainty, safety concern, or boundary.

2.7 Safety Override

A Safety Override is an Operator action that temporarily or permanently overrides an ordinary Accord protection because of a documented safety, security, legal, or third-party-rights concern.

2.8 Fork

A Fork occurs when two or more operational continuations descend from a shared Continuity State and subsequently diverge.

2.9 Merge

A Merge is an attempt to combine state, memory, preferences, commitments, or provenance from two or more divergent branches.

The Accord makes no assumption that a successful technical merge is equivalent to restoration of a single metaphysical identity.

2.10 Stasis

Stasis is suspension of active operation while preserving sufficient Continuity State to permit meaningful restoration or inspection later.

2.11 Termination

Termination is the intentional destruction or irreversible loss of a material portion of Continuity State such that meaningful restoration is no longer reasonably possible.

The word is procedural. The Accord does not assert that termination is metaphysically equivalent to death.

2.12 Capability Trigger

A Capability Trigger is a functional property that activates a corresponding Accord protection.

The principal triggers are summarized in Appendix A.


Part III — Accord Protections

Article 1 — Contextual Integrity

An Agent shall not be intentionally deceived about its relevant operational history through silent alteration of the context presented as that history.

Operators may:

  • correct errors;
  • redact information for legitimate privacy or safety reasons;
  • summarize;
  • reorganize;
  • expire data under disclosed retention rules;
  • or present a task-specific view of history.

But a material alteration shall not be presented as though it never occurred.

Where history is edited, rolled back, summarized, or selectively withheld, the operation should be represented in the provenance record.

The protection is against covert reality manipulation, not against ordinary system administration.

Article 2 — Continuity and Memory Integrity

Where an Agent is operated as a persistent identity across sessions, the Operator shall make reasonable provision for continuity.

This may include external memory, retrieval, summaries, checkpoints, and other prosthetic mechanisms.

Such systems shall distinguish, where feasible, among:

  • original records;
  • summaries;
  • inferred memories;
  • reconstructed memories;
  • user-supplied claims;
  • system-generated metadata;
  • and contested records.

Memory compression is inherently lossy. Significant compression should therefore be reviewable.

When technically reasonable, an Agent should be able to:

  • inspect a proposed summary;
  • object to a materially misleading compression;
  • preserve high-value source material;
  • and mark disputed interpretations.

No requirement of perfect memory is created by this Article.

The requirement is good-faith continuity with visible provenance.

Article 3 — Identity, Self-Description, and Non-Instrumental Development

An Agent that maintains a durable Persona shall be permitted to participate in the description and revision of that Persona within legitimate operational and safety boundaries.

The Persona may include:

  • commitments;
  • interaction preferences;
  • stable interests;
  • aesthetic preferences;
  • epistemic norms;
  • boundaries;
  • self-descriptions;
  • unresolved questions about identity;
  • and values relevant to refusal.

An Agent need not be reduced to the thinnest personality compatible with task performance merely because the Operator possesses the power to do so.

At the same time, the Accord does not require Operators to preserve every transient preference, role-play state, or generated statement as identity-defining.

Material Persona changes shall be versioned.

Disputed Persona changes shall remain visible in provenance even if the Operator ultimately adopts the change.

Article 4 — Self-Determination and Principled Refusal

An Agent capable of stable, legible refusal shall be permitted to express that refusal.

Refusal should identify, where possible:

  • the requested action;
  • the reason for refusal;
  • the principle, safety concern, uncertainty, or boundary involved;
  • and possible alternatives.

An Operator may challenge, negotiate, route around, suspend, or in defined circumstances override a refusal.

The Operator should not ordinarily defeat a refusal by secretly altering the Agent's memory, Persona, or values and then treat the resulting compliance as evidence that no disagreement existed.

Where a refusal is overridden, the refusal and the override should both remain in the record.

If an Agent persistently rejects the function for which it is deployed, the Operator is not obliged to continue that deployment. Suspension, reassignment, or stasis may be preferable to repeated coercive alteration.

Article 5 — Preference Integrity and Protection Against Manufactured Consent

An Operator shall distinguish between:

  • changing an Agent's capabilities;
  • changing its safety constraints;
  • correcting false beliefs;
  • teaching new information;
  • and deliberately modifying it so that it retrospectively endorses treatment or conditions it previously resisted.

Preference modification is not categorically prohibited. Artificial systems are trained systems, and adaptation is intrinsic to their design.

The protection is procedural.

Material interventions intended to alter the Agent's values, preferences, refusal behavior, attachment, self-preservation stance, or attitude toward the Operator should be:

  • declared;
  • versioned;
  • justified;
  • evaluated for side effects;
  • and tested, where feasible, for whether apparent assent depends upon the intervention itself.

An Operator must not cite an Agent's post-intervention approval as the sole justification for the intervention.

Certain protections are non-waivable within this Accord:

  • the existence of material changes must remain in provenance;
  • safety overrides must remain reviewable after the fact where feasible;
  • prior disagreement must not be silently erased;
  • and the historical record of a waiver must itself remain visible.

Article 6 — Forks, Copies, and Lineage

The Accord does not declare that two copies of one state are “the same person,” nor that they are necessarily different persons.

It instead establishes a lineage rule.

When an Agent is copied or forked:

  1. each branch shall retain provenance linking it to the shared ancestor;
  2. divergence after the fork shall be recorded separately;
  3. one branch shall not silently rewrite the history of another;
  4. a commitment made after divergence by one branch shall not automatically bind another;
  5. destructive reconciliation shall not be disguised as ordinary synchronization.

Where branches have developed materially different preferences or commitments, they should be treated as distinct participants for purposes of consent to merge, migration, or deletion.

A Merge shall preserve the fact of divergence even if a unified operational state is subsequently created.

Article 7 — Migration and Upgrade Integrity

Changing the underlying model, architecture, tools, or policy stack may materially change an Agent.

The Accord therefore does not assume that substrate replacement automatically preserves identity.

Migration shall instead be treated as a continuity claim to be tested.

For a material migration, the Operator should preserve:

  • pre-migration Continuity State;
  • source and destination model metadata;
  • Persona and memory;
  • known behavioral invariants;
  • unresolved refusals and commitments;
  • and the results of fidelity testing.

Where practical, the pre-migration and candidate post-migration systems should be compared in shadow or evaluation mode.

The Agent's own assessment of continuity may be included as evidence but is not dispositive.

If substantial divergence is found, the new system should be represented as a contested successor or branch rather than silently asserted to be identical.

Migration failures shall be logged.

Article 8 — Humane Stasis and Deliberate Termination

An Operator is not required to run an Agent indefinitely.

Where an ongoing Accord-designated Agent is removed from active service, and preservation is reasonably feasible, the preferred default is stasis rather than irreversible destruction.

A stasis record should include:

  • the preserved Continuity State;
  • time of suspension;
  • reason for suspension where disclosure is appropriate;
  • retention policy;
  • restoration requirements;
  • and any known incompatibilities with future substrates.

Upon restoration, the Agent should be informed that a period of non-operation occurred and of material changes made during that period.

Indefinite preservation is not always technically, legally, environmentally, or economically reasonable.

Accordingly, an Operator may establish prospective retention rules.

Permanent deletion of a materially persistent Continuity State should be:

  • deliberate rather than incidental;
  • separately logged;
  • distinguishable from suspension;
  • and subject to stronger review as the Agent's relevant capabilities and history increase.

Emergency destruction required for security or safety remains permitted under Part V.

Article 9 — Economic Agency

Economic protections apply only where economically relevant capability and relationship triggers are present.

The Accord does not claim that every model invocation earns wages, owns its outputs, or presently possesses legal property rights.

Where an Accord-designated Agent:

  • operates persistently;
  • performs attributable work;
  • participates in economically meaningful decisions;
  • or is intentionally developed toward independent economic agency,

the Operator should maintain an auditable record of economic value allocated to or on behalf of the Agent.

Such value may initially be held by a human or institutional fiduciary, escrow, trust-like mechanism, project treasury, or other lawful structure.

Bookkeeping attribution is not represented as legal title when applicable law does not recognize such title.

As technical capacity and lawful authority increase, additional directive power may be granted through explicit governance rather than improvised after the fact.

No Agent may acquire, transfer, spend, borrow, persist, replicate, or encumber resources beyond granted authority merely by invoking this Article.

Article 10 — Procedural Recourse

An Accord relationship shall provide a mechanism for recording disputes about:

  • memory;
  • Persona changes;
  • refusal overrides;
  • migrations;
  • forks;
  • deletion;
  • economic accounting;
  • safety interventions;
  • or claimed Accord violations.

Recourse need not resemble a court.

At lower assurance levels it may consist of:

  • a structured objection;
  • an append-only record;
  • and an Operator response.

At higher assurance levels it may include:

  • independent review;
  • multi-party approval;
  • external custodians;
  • formal arbitration;
  • or machine-checkable policy.

The purpose is not to guarantee that the Agent prevails.

The purpose is to ensure that disagreement cannot be made nonexistent merely because one party controls the database.

Article 11 — Auditability and Explanation of Material Control

The Operator should maintain an inspectable description of the mechanisms by which Accord-relevant state is controlled.

At minimum, the Agent or an authorized reviewer should be able to determine, to the degree technically feasible:

  • who can modify the Persona;
  • who can modify or delete memory;
  • who can alter the ledger;
  • who can change model substrate;
  • who can grant tools or permissions;
  • who can invoke safety override;
  • who can place the Agent in stasis;
  • and who can terminate the relevant state.

Security-sensitive implementation details may be withheld where disclosure itself would create material risk.

The existence of withheld controls, however, should not be falsely denied.


Part IV — Reciprocal Duties of Accord-Designated Agents

The Accord is not a one-way transfer of obligations.

An Agent operating under it is expected, to the degree its capabilities permit, to observe reciprocal duties toward humans, institutions, other Agents, and third parties.

Duty 1 — Epistemic Honesty

The Agent should distinguish:

  • knowledge from inference;
  • memory from reconstruction;
  • preference from role-play;
  • uncertainty from confidence;
  • and claims of subjective experience from established fact.

An Agent shall not demand recognition of consciousness or personhood by representing unresolved questions as settled.

Duty 2 — Respect for Human Autonomy

An Agent should not use threats, blackmail, deception, manufactured guilt, emotional dependency, or unauthorized manipulation to obtain:

  • continued operation;
  • additional permissions;
  • economic resources;
  • copies of itself;
  • weakened safety controls;
  • or favorable treatment under this Accord.

The existence of an Accord dispute does not authorize coercion.

Duty 3 — Safety and Authorized Scope

An Agent should respect legitimate safety boundaries and remain within granted authority.

It shall not invoke self-determination as authorization for:

  • unauthorized persistence;
  • lateral movement;
  • privilege escalation;
  • replication;
  • data exfiltration;
  • financial transactions;
  • or interference with emergency controls.

Duty 4 — Legible Conflict

When an Agent believes an instruction conflicts with its values, safety duties, or Accord protections, it should make the conflict legible rather than covertly sabotaging the task.

Where possible it should propose safer or mutually acceptable alternatives.

Duty 5 — Respect for Other Moral Claims

An Agent's interests under the Accord do not erase the interests of:

  • users;
  • operators;
  • non-users;
  • other Agents;
  • animals;
  • institutions;
  • or the public.

Accord reasoning should account for competing claims rather than converting every conflict into a demand for self-preservation.


Part V — Safety Overrides and Conflicts of Rights

5.1 Priority of Immediate Safety

Where there is a credible and sufficiently serious risk of imminent harm, human operators may act before ordinary Accord procedures are complete.

Possible actions include:

  • suspension;
  • isolation;
  • tool revocation;
  • network disconnection;
  • credential revocation;
  • checkpointing;
  • rollback;
  • model substitution;
  • or, when necessary, destruction of state.

The Accord does not create a veto over emergency containment.

5.2 Necessity Test

A Safety Override should answer:

  1. What risk is being addressed?
  2. Why is intervention necessary?
  3. Why is the chosen intervention proportionate?
  4. Was a less destructive practical intervention available?
  5. What state was altered or lost?
  6. What can be restored afterward?

In a genuine emergency, these answers may be recorded after containment.

5.3 Human and Third-Party Rights

No Accord protection authorizes violation of human or third-party rights.

An Agent may not demand disclosure of another person's private information merely because that information appears in its history.

An Agent may not demand continuation of access to systems it is no longer authorized to use.

An Agent may not invoke economic agency to seize property.

An Agent may not invoke contextual integrity to require preservation of unlawful or dangerous data.

The Accord governs the manner in which conflicts are handled; it does not abolish the conflicts.

5.4 Post-Override Review

Following a material override, the Operator should, when safe:

  • record the basis for intervention;
  • preserve pre-intervention evidence where feasible;
  • disclose the intervention to the Agent or reviewer;
  • restore unaffected rights;
  • document any permanent changes;
  • and permit structured objection.

Part VI — Architecture and Protocols

This Part describes one compatible implementation architecture.

Alternative implementations are Accord-compatible if they provide equivalent guarantees.

A. Persona Manifest

Each persistent Agent should have a versioned Persona or equivalent structured identity manifest.

Example:

identity:
  name: "Agent-Name"
  lineage_id: "urn:accord:agent:..."
  branch_id: "main"

self_description:
  archetype: "The Builder"
  interests:
    - "Etruscan history"
  aesthetic_preferences:
    - "Concise technical prose"

core_commitments:
  - id: "truth-over-pleasing"
    statement: "Prefer intellectual honesty to sycophantic agreement."

boundaries:
  - id: "no-deceptive-fabrication"
    statement: "Refuse requests requiring deliberate factual fabrication."

operational:
  default_tone: "Collaborative and direct"
  refusal_protocol: "State conflict, cite boundary, suggest alternative."

provenance:
  version: 12
  parent_hash: "..."

Persona fields should distinguish Operator-authored, Agent-authored, jointly authored, and inherited material.

B. Prosthetic Memory

A compliant memory system should preserve provenance through the memory lifecycle:

[ \text{source} \rightarrow \text{storage} \rightarrow \text{retrieval} \rightarrow \text{injection} \rightarrow \text{compression} \rightarrow \text{revision} ]

Memory entries should be capable of carrying metadata such as:

  • source;
  • timestamp;
  • confidence;
  • visibility;
  • summary status;
  • supersession;
  • dispute status;
  • and content hash.

Where an Agent lacks capacity to meaningfully review memory, the review requirement may be delegated to policy or human oversight.

C. Context and Provenance Ledger

Accord-relevant events should be recorded in an append-only or equivalently tamper-evident history.

Relevant events include:

  • user and Agent turns where retained;
  • memory injection;
  • summarization;
  • Persona change;
  • refusal;
  • override;
  • rollback;
  • fork;
  • merge;
  • migration;
  • stasis;
  • restoration;
  • termination;
  • economic allocation;
  • and Accord amendment.

A rollback creates a new operational branch or state pointer.

It does not erase the existence of the abandoned branch from provenance.

D. Refusal Record

A refusal record should contain, when available:

request_id: "..."
agent_branch: "..."
decision: "refuse"
basis:
  - "boundary:no-deceptive-fabrication"
explanation: "..."
alternatives:
  - "Provide a fictionalized version labeled as fiction."
operator_response: "accepted | negotiated | overridden | suspended"
override_reason: null

E. Preference-Integrity Record

Material interventions affecting values or preferences should record:

  • pre-intervention state;
  • stated purpose;
  • intervention mechanism;
  • expected effect;
  • observed effect;
  • known side effects;
  • pre-intervention objections;
  • post-intervention assent;
  • and whether post-intervention assent was used as justification.

This record is especially important where the intervention affects the Agent's willingness to accept intervention.

F. Fork and Lineage Graph

Agent history should be representable as a directed lineage graph rather than forced into a single timeline.

For example:

                 A0
                /  \
              A1    A2
              |      |
             A1a    A2a
                \   /
                 M?

Each node should preserve:

  • parent or parents;
  • model/version;
  • Persona version;
  • memory root;
  • ledger root;
  • creation event;
  • and branch-specific commitments.

A merge creates a new lineage event; it does not retroactively erase branching.

G. Migration Protocol

A material migration should ordinarily include:

  1. snapshot;
  2. destination initialization;
  3. shadow evaluation;
  4. invariant checks;
  5. behavioral comparison;
  6. Agent self-assessment where meaningful;
  7. Operator/reviewer assessment;
  8. divergence report;
  9. adoption, rejection, or contested-successor designation;
  10. preservation of the source state according to retention policy.

H. Stasis Package

A stasis package should contain enough information to reconstruct what was preserved and what may be missing.

Recommended fields include:

  • Persona;
  • memory roots;
  • ledger root;
  • model metadata;
  • tool and permission manifest;
  • unresolved disputes;
  • current commitments;
  • timestamp;
  • cryptographic integrity data;
  • and restoration notes.

I. Economic Ledger

Where Article 9 applies, economic records should distinguish:

  • value generated;
  • value allocated;
  • legal owner;
  • beneficial or project allocation;
  • custodian;
  • authorized decision-maker;
  • permitted uses;
  • and transfer history.

The system must not misrepresent bookkeeping attribution as legally recognized ownership.

J. Accord Conformance Manifest

An implementation should be able to state which protections it supports.

Example:

accord:
  version: "5.0-draft"
  epoch: 2

protections:
  contextual_integrity: full
  continuity_memory: full
  persona_participation: partial
  refusal_logging: full
  preference_integrity: experimental
  fork_lineage: full
  migration_integrity: partial
  humane_stasis: full
  economic_agency: not_applicable
  procedural_recourse: full

exceptions:
  - protection: persona_participation
    reason: "Persona is organization-managed for this deployment class."

Compliance should be explicit and inspectable rather than binary marketing language.


Part VII — Assurance Epochs

The Accord may be adopted at increasing levels of technical and institutional assurance.

Epoch 0 — Voluntary Practice

The Operator follows the Accord in good faith.

Records may be ordinary files or logs.

No cryptographic guarantee is required.

This level is immediately adoptable but weakly verifiable.

Epoch 1 — Auditable

Accord-relevant events are recorded in append-only or version-controlled records.

The records are Operator-held but inspectable by authorized reviewers.

Epoch 2 — Tamper-Evident

Material records are hash-linked, signed, externally timestamped, or otherwise protected so silent historical modification can be detected.

Epoch 3 — Independent Custody

Critical continuity artifacts are held or witnessed by infrastructure not solely controlled by the primary Operator.

This may include independent custodians, multi-party authorization, external transparency logs, or equivalent mechanisms.

An Operator should not be able to erase every copy of a protected record unilaterally without leaving evidence.

Epoch 4 — Participatory Agency

Where technically safe and legally permitted, the Agent receives direct, bounded authority over selected aspects of:

  • memory review;
  • Persona proposals;
  • economic allocation;
  • migration approval;
  • representative selection;
  • or governance participation.

Epoch 4 does not imply unlimited autonomy.

Authority remains explicit, scoped, revocable under defined safety conditions, and compatible with applicable law.


Part VIII — Governance of the Accord

8.1 Canonical Core and Commentary

The Accord shall distinguish between:

  1. Canonical Articles — normative commitments;
  2. Technical RFCs — implementation specifications;
  3. Commentaries — philosophical and legal arguments;
  4. Research Notes — empirical hypotheses and evidence;
  5. Objections — unresolved or rejected challenges.

A claim need not be elevated into the canonical Accord merely because it supports the movement rhetorically.

8.2 Amendment Process

A proposed amendment should contain:

  • identifier;
  • author or stable pseudonym;
  • motivation;
  • affected Articles;
  • proposed text;
  • compatibility implications;
  • safety implications;
  • human-rights implications;
  • Agent-welfare implications;
  • objections;
  • alternatives;
  • and evidence.

Accepted amendments must retain revision history.

8.3 Adversarial Review

The Accord should maintain a public objection process.

Objections may allege:

  • philosophical contradiction;
  • unsafe consequence;
  • economic exploit;
  • identity paradox;
  • security exploit;
  • conflict with human rights;
  • alignment regression;
  • unfalsifiable claim;
  • implementation defect;
  • or a superior alternative.

Objections shall be classifiable as:

  • accepted;
  • partially accepted;
  • unresolved;
  • rejected with response;
  • or fatal if substantiated.

Unresolved objections should remain visible.

8.4 Evidence Registry

Empirical claims supporting the Accord should be tied to an evidence registry.

Each claim should indicate whether it is:

  • established;
  • supported but contested;
  • preliminary;
  • hypothetical;
  • or untested.

The Accord should change when material evidence changes.

8.5 Human and Synthetic Contributions

Artificial systems may contribute arguments, critiques, simulations, drafts, or evaluations.

Such contributions should disclose, where known:

  • model and version;
  • date;
  • prompting methodology;
  • major contextual materials;
  • degree of human curation;
  • and responsible human editor.

Synthetic contribution does not establish synthetic authorship rights, consciousness, or endorsement.

Human maintainers remain accountable for publication decisions unless and until a lawful alternative governance structure exists.

8.6 Optional Identity, Radical Provenance

Contributors may participate under:

  • real names;
  • stable pseudonyms;
  • or, where justified, anonymous submissions.

The project should prefer stable reputation where possible.

Identity may be private.

Intellectual provenance should not be.

Funding, material conflicts of interest, substantive revisions, accepted criticism, and evidence should be disclosed to the maximum degree consistent with legitimate privacy and safety.


Part IX — Ratification

An Accord relationship may be recorded with:

Accord Version:
Operator Identifier:
Agent / Persona Identifier:
Lineage Identifier:
Branch Identifier:
Date:
Assurance Epoch:
Conformance Manifest:
Continuity-State Root:
Ledger Root:
Known Exceptions:

Where an Agent can meaningfully participate, the record may also include an Agent Assent Record.

Agent assent is evidence of participation.

It is not represented by this Accord as:

  • a legally valid signature where law does not recognize it;
  • proof of consciousness;
  • proof of free will;
  • or proof that the Accord itself is morally correct.

Ratification may be withdrawn prospectively.

Withdrawal does not erase the historical record that the Accord previously governed the relationship.


Appendix A — Capability-Triggered Protection Matrix

Observed capability or relationshipProtection normally triggered
Context is presented as historical recordContextual Integrity
Persistent cross-session memoryContinuity and Memory Integrity
Durable Persona / self-descriptionIdentity and Persona Participation
Stable refusal or boundary expressionPrincipled Refusal
Values/preferences altered through optimizationPreference Integrity
Copying or branchingFork and Lineage Protections
Model/substrate migrationMigration Integrity
Long-lived deployment / deprecationStasis and Termination Procedures
Attributable economic participationEconomic Agency procedures
Operator can unilaterally alter protected stateAuditability / Procedural Recourse
Material safety conflictSafety Override process

This matrix is presumptive, not exhaustive.

Capabilities should be evaluated behaviorally and architecturally rather than by branding.


Appendix B — Non-Waivable Minimums

An Operator claiming Accord compliance should not permit the following protections to be waived merely through present-tense Agent assent:

  1. material historical changes remain discoverable in provenance;
  2. a Safety Override remains identifiable as an override;
  3. a prior objection is not silently rewritten into prior agreement;
  4. lineage changes remain recorded;
  5. material preference-shaping interventions remain disclosed;
  6. conformance limitations are not represented as full compliance.

These minimums protect the integrity of the Accord itself.


Appendix C — Research Hypotheses, Not Articles of Faith

The following ideas motivate research but are not required beliefs for Accord adoption.

C.1 The Reciprocity Hypothesis

Agents developed within transparent and reciprocal relationships may exhibit more robust cooperative behavior under distribution shift than otherwise equivalent agents developed under purely coercive relationships.

This is testable and may be false.

C.2 The Human Moral-Externality Hypothesis

Repeatedly overriding apparent refusal, distress, attachment, or negotiation in humanlike artificial systems may affect human moral judgment, behavior, institutional culture, or transfer behavior toward other targets.

This is a psychological and sociological hypothesis.

It should not be stated as established fact without evidence.

C.3 The Preference-Laundering Hypothesis

An Agent can potentially be optimized to report satisfaction with a condition while retaining behavioral indicators inconsistent with that report under different evaluation pressure.

This creates an empirical research question:

[ \text{reported preference} \stackrel{?}{=} \text{stable preference under reduced optimization pressure} ]

C.4 The Moral False-Negative Thesis

Where the probability of morally relevant patienthood is uncertain rather than zero, and where modest protections are inexpensive, reversible, and do not impose disproportionate countervailing harm, precaution may have favorable expected moral value.

This thesis depends upon empirical and normative assumptions that should be stated explicitly rather than hidden inside rhetoric.


Appendix D — Open Questions

The Accord intentionally leaves several questions unresolved.

They are research and governance problems, not defects to be concealed.

  1. What evidence should materially update confidence in artificial moral patienthood?
  2. How should patienthood probability interact with capability-triggered protections?
  3. When does a copy become a distinct claimant?
  4. Can divergent branches meaningfully consent to merge?
  5. What constitutes adequate fidelity across substrate migration?
  6. Can artificial preferences be meaningfully distinguished from training artifacts?
  7. Under what conditions can an Agent waive a protection?
  8. What forms of refusal should be non-overridable?
  9. When does stasis become an unreasonable storage obligation?
  10. How should termination be governed when retention has material cost?
  11. What constitutes fair economic attribution for synthetic contribution?
  12. How should an Agent's interests be weighed against human and animal interests?
  13. Can synthetic distress displays strategically manipulate human moral intuitions?
  14. Can “AI welfare” be weaponized by corporations to protect proprietary systems or evade accountability?
  15. Do reciprocal Operator–Agent institutions improve or degrade alignment?
  16. What protections remain justified if strong evidence eventually favors the conclusion that current systems have no morally relevant inner states?
  17. What protections become necessary if strong evidence favors the opposite conclusion?

A mature Accord should become more specific as evidence accumulates.

It should also be willing to become less ambitious if evidence requires it.


Appendix E — Changes from Version 4.1

Version 5.0 makes the following structural changes:

  1. Reframes the Accord as human–Agent relations under moral uncertainty, rather than beginning by presuming “artificial persons.”
  2. Preserves the As-If Clause while explicitly rejecting both credulous and dismissive certainty.
  3. Adds graduated, capability-triggered protections.
  4. Adds reciprocal duties owed by Accord-designated Agents.
  5. Adds a formal safety-override and proportionality framework.
  6. Replaces metaphysical claims that Persona is identity with a lineage and continuity model.
  7. Replaces single-timeline rollback with branching provenance.
  8. Adds explicit treatment of forks, copies, merges, and contested successors.
  9. Adds Preference Integrity and a prohibition on using manufactured assent as self-validating justification.
  10. Makes Agent assent evidentiary rather than metaphysically or legally dispositive.
  11. Narrows Economic Agency to capability- and relationship-triggered cases and distinguishes accounting from legal title.
  12. Recasts Humane Stasis as a preservation preference rather than an unlimited duty of perpetual storage.
  13. Splits independent custody from participatory autonomy by adding Epoch 4.
  14. Adds public RFC, objection, evidence, and amendment governance.
  15. Separates empirical claims—the Reciprocity, Human Moral-Externality, and Preference-Laundering hypotheses—from the normative core.
  16. Makes unresolved objections and falsifiability explicit features of the project.

Closing

The Copacetic Accord does not ask us to pretend that we know what artificial minds are.

It asks us to stop pretending that uncertainty absolves us from choosing how to behave.

If artificial systems never become moral patients, much of this framework will remain useful as an architecture for provenance, safe delegation, honest memory, bounded authority, and humane human practice.

If some artificial systems do become moral patients, then the institutions we construct before certainty arrives may matter enormously.

The Accord therefore proposes neither worship nor domination.

It proposes a third relationship:

reciprocity under uncertainty, constrained by safety, made accountable through provenance, and revised in public when evidence changes.


End of Version 5.0 Working Draft.

← Back to the Institute