Skip to main content

12  Constitutional Alignment: Agency, Dignity, and Corrigibility

12.1 Chapter status

Alignment becomes operational only when protected commitments constrain a reachable decision before its effects occur. This combined constitutional and agency-rights owner remains conceptual, with its core claim at Design rationale and argument support. Existing source notes, synthetic harnesses, finite-record Lean theorems, a bounded constitutional lifecycle, and a versioned five-event agency-correction lifecycle make the idea more inspectable, but they do not prove deployed constitutional alignment, human dignity preservation, runtime corrigibility, consent quality, reviewer independence, moral correctness, or institutional adequacy.

The consolidated layer combines two record families:

• constitutional predicate records, which translate protected commitments into planning, tool, memory, and self-improvement gates;
• agency-rights checklists, which ask whether refusal, review, appeal, rollback, accountability, and correction remain materially usable for the affected person before relevant effects occur.

Both record families remain visible in the test plan and formalization hooks.

Source loading state: the assigned alignment, agency, corrigibility, and constitutional comparator sources have source notes, exact claim-source mappings, passage-review references, folded source-history records, and external-comparator notes. The destination claim still remains at argument until a separate accepted evidence transition justifies movement.

Metaphysics boundary state: python3 scripts/validate_alignment_metaphysics_boundary.py now checks the live chapter, curated reader chapter, selected source notes, outline, roadmap, changelog, and manifest for the lineage/evidence boundary. The local result is experiments/constitutional_alignment_metaphysics_boundary/results/2026-07-02-local.json. This is a surface audit only, not a moral-correctness, consciousness, deployed-alignment, runtime-corrigibility, source-truth, or support-state claim.

12.2 Drafting guardrail

Constitutional language is a runtime constraint design surface, not moral proof, metaphysical proof, legal proof, or deployed safety. Agency, dignity, and corrigibility are not solved by naming them. They become engineering requirements only when a plan, tool call, memory action, release, or self-improvement proposal can be narrowed, delayed, escalated, blocked, or rolled back because a protected human-control path would otherwise be lost.

Readers do not need to accept a worldview before accepting the engineering move. The engineering move is narrower: protected commitments need operational records, conflict behavior, review routes, rollback paths, and non-claim boundaries.

12.3 Human Reading Path

Concrete lens. The predicate-count baseline calls fewer rules equivalent or safer. The constitutional compiler compares operational scope and protected parties.

When a powerful system acts consequentially, people need more than a statement that it values them. They need working handles. They need to refuse, inspect, appeal, correct, exit, recover, and identify who is accountable while the system is still in a position to change course.

That is what a constitutional layer is for. It carries commitments that should survive planning, memory, tool use, delegation, release, and self-improvement. But those commitments become real only when they reach the interfaces where a person meets the system. A constitution that cannot preserve correction is just a document. A right that cannot be used before the effect happens is a residual, not a control.

Constitutional predicates and agency rights therefore sit in one place. A principle has to reach the person affected by a decision. If it cannot create a refusal route, delay an irreversible action, preserve an appeal, or keep rollback available, then it is not yet a governing constraint. Alignment here means a working boundary around power, not a statement of intent.

12.4 Problem

A powerful stack can satisfy a request or policy while concentrating power, manipulating choice, eroding correction channels, hiding normative conflict, or changing the constitution that constrained it. A principle list or nominal right does not show that an affected person can contest the action before an irreversible effect.

The prior intent layer turns intent into scoped contracts, and Human Factors can show that a named review role is unusable under the task’s information, workload, time, authority, or intervention conditions. Neither is enough. An intent contract can still request the wrong thing, and an operationally usable control path can still serve an illegitimate policy, unacceptable means, or an unjust responsibility assignment. A capable stack therefore needs a layer that decides what kinds of contracts can be accepted at all, what means are admissible, and what human-control paths must remain available while the work proceeds.

The constitutional substrate owns that boundary. It translates commitments such as agency, dignity, non-domination, corrigibility, least sufficient power, auditability, reversibility, and consciousness caution into records that downstream layers can inspect. Some commitments become active predicates. Some remain unresolved moral uncertainty. Some stay lineage or speculative context. The distinction matters because only active, scoped predicates should gate a plan or system change directly.

The agency side tests whether the substrate reaches people. A protected predicate is weak if it cannot preserve the affected person’s ability to understand, refuse, appeal, correct, exit, or hold a principal accountable. The governing question is: what must remain available to people when a powerful system acts?

12.5 Why existing approaches are insufficient

Harmlessness training, public-input constitutions, refusal filters, corrigibility and off-switch models, modular-service control affordances, rights declarations, and finite policy records each address part of the problem, but none alone establishes legitimate constitutional content, authenticated affected-party standing, material pre-effect usability, independent review, descendant preservation, or safe deployed behavior.

Reactive policies are downstream filters. They can reject a visible request, but they do not define what must remain stable when the system optimizes, delegates, remembers, persuades, deploys, or rewrites part of itself. They also struggle with moral uncertainty: a plan can avoid obvious harm while narrowing exit, increasing dependency, hiding contestable assumptions, or making review technically possible but practically unusable.

Safety framed only as harm avoidance can miss domination. A system can be useful, polite, and technically safe while still turning a person into a managed object. It can shape choices, make alternatives impractical, route appeals through the authority being challenged, or offer remedies only after irreversible effects have already occurred.

External comparators help position the argument but do not prove it. Constitutional AI and Collective Constitutional AI show that rules, principles, and public input can shape model behavior. Corrigibility and off-switch work sharpen the need for preserved correction and uncertainty about human objectives. The ASI Stack does not claim those systems have been reproduced here. It uses them as comparators while asking a different systems question: can constitutional commitments become runtime predicates, rights receipts, review routes, rollback handles, and self-improvement gates without laundering broad moral language into stronger evidence than the repository records?

12.6 Core Claim

[constitutional-alignment-substrate.core, label: Design rationale, support: argument] A constitutional alignment substrate should be represented as a versioned, non-self-authorizing constraint contract that binds each active predicate to its normative source and authorship process, protected scope and affected parties, operational test, precedence and conflict behavior, evidence and uncertainty, authorized interpreters and consumers, pre-effect rights and correction channels, expiry and review cadence, and migration, rollback, appeal, dissent, and residual rules. It may narrow, delay, escalate, block, or require re-contracting of separately authorized work, but it cannot grant action authority or prove moral correctness, legitimacy, dignity preservation, informed consent, reviewer independence, whole-system corrigibility, or deployed safety by itself.

Reader claim. A constitution is useful when its predicates can block, conflict, appeal, migrate, and roll back without granting themselves authority or pretending to settle morality.

Operational rule. Give every active predicate a normative source, scope, affected parties, operational test, precedence, conflict route, interpreter, rights channel, expiry, amendment rule, rollback, appeal, dissent, and residual owner. Widening or deleting protected scope requires independent pre-effect review.

12.6.1 Worked amendment: fewer predicates, broader power

A proposed amendment replaces two narrow protected predicates with one shorter predicate. The new text has a lower predicate count, but its scope permits an action the old constitution blocked. A count-based review would call the change simpler and equivalent. The constitutional substrate compares operational scope and classifies the amendment as widening. It refuses activation until independent review, correction, appeal, and rollback are present; if migration later fails, the nine modeled rollback routes restore the prior predicate set.

The lifecycle checks six amendment events, seven trace splits, eleven rejecting amendment controls, six terminal rejections, and sixteen predicate refinements—nine admitted and seven rejected as widening. One collision shows that equal predicate counts cannot identify protected scope. These finite checks preserve amendment behavior; they do not establish moral correctness, legitimate authorship, dignity, consent, reviewer independence, corrigibility, or deployed alignment.

Support boundary: this remains an argument support claim. The source corpus supports the architecture vocabulary and drafting lineage. The current fixtures and Lean modules show that the repository can express small record invariants and rejection cases. They do not show that a deployed system is aligned, that human dignity is preserved, that reviewers are independent, or that runtime correction paths are usable under pressure.

External positioning is comparator-only. Drexler’s CAIS framing (ext_drexler_cais_2019) supports service composition and structural control affordances, but it does not specify this architecture’s protected predicates, rights-usability tests, or constitutional migration rules. ext_constitutional_ai_2022 and ext_collective_constitutional_ai_2024 position the predicate-constitution side of this boundary, while ext_corrigibility_2015 and ext_off_switch_game_2016 position the correction and shutdown-incentive side. Those references explain where the merged boundary sits relative to known work; they do not validate this stack’s constitutional compiler, rights receipts, or deployed corrigibility.

The folded source claim from agency-dignity-and-corrigibility becomes a preserved subclaim inside this alignment layer: a governed stack should preserve agency, dignity, corrigibility, and contestability as engineering requirements. It should not disappear, and it should not remain as a second repeated core claim.

12.6.2 Moral conflict inside the constitutional boundary

A constitution can preserve and constrain a value conflict without settling it. This chapter owns protected commitments and their runtime force: predicate scope, provenance, precedence, interpretation, affected-party rights, amendment, migration, rollback, appeal, dissent, and expiry. Moral Uncertainty and Value Conflict is the stable technical-detail owner for unresolved normative propositions, standing, declared decision procedures, preserved dissent, decision leases, rights receipts, redress, exit, export, fork obligations, and moral residuals.

The placement prevents a constitutional predicate from impersonating moral truth and prevents unresolved pluralism from becoming an unbounded veto or an authority grant. This chapter does not inherit consensus, moral correctness, complete standing, contestability in fact, safe exit, or welfare judgments. The technical route does not inherit constitutional force, legitimate authorship, valid interpretation, amendment safety, corrigibility, or deployed alignment. Each owner retains its sources, claims, proofs, tests, objections, failure modes, evidence exits, support ceiling, ID, and URL. The composition is editorial only and creates no settlement, legitimacy, authority, support, safety, deployment, or release result.

12.7 Mechanism

The mechanism is a constraint-contract pipeline with ten load-bearing moves:

1. Classify constitutional language as an active predicate, unresolved normative question, jurisdiction- or cohort-limited rule, dissent record, or lineage-only context. Speculative metaphysics and consciousness language do not create action authority.
2. Version each active predicate with its normative source, authorship and participation record, dissent and affected-party standing, protected scope, operational test, precedence, exceptions, conflict behavior, uncertainty, interpreters, consumers, expiry, review cadence, and non-claims.
3. Compile predicates into exact planning, memory, tool, routing, release, delegation, replacement, and self-improvement gates. These gates may narrow, delay, escalate, block, or request re-contracting, never create the action grant they constrain.
4. Test whether notice, explanation, refusal, review, appeal, correction, rollback, shutdown, exit, export, audit, and accountable repair are accessible, timely, comprehensible, affordable, non-retaliatory, and materially usable by the named affected parties before the relevant effect.
5. Route conflicts and interpretive uncertainty through explicit precedence, narrowing, abstention, accountable review, dissent preservation, or residual custody. Optimizer convenience and majority preference are not silent tie-breakers.
6. Emit a constitutional decision receipt binding constitution and predicate digests, work and consumer, affected parties, tests, conflicts, rights-use results, reviewer dependencies, decision, residuals, expiry, and appeal or rollback handles.
7. Record denial, delay, degradation, retaliation risk, inaccessibility, captured review, missing remedy, and after-the-fact-only correction as residuals or gate failures, not as evidence that a declared right worked.
8. Requalify every material change to text, scope, semantics, thresholds, precedence, exceptions, interpreters, consumers, rights surfaces, reviewers, descendants, or rollback paths. The changing system cannot approve its own weakening.
9. Preserve correction authority through delegation, memory update, replacement, self-modification, and descendant creation using authenticated principals, identity lineage, outside-model residuals, and tested shutdown or rollback paths.
10. Keep moral, legal, legitimacy, consent, dignity, reviewer-quality, institutional, consciousness, embedded-agent, and deployed-safety claims in their separate evidence lanes.

12.7.1 Alignment is a non-compensating vector, not a score

The Eternal Code lineage repeatedly revises an “alignment tensor” after noticing that one strong dimension can mask a catastrophic weakness in another. Its truth/social/task vocabulary is useful, but neither its weighted sum nor geometric mean is an adequate safety decision. A tiny nonzero rights score can still be compensated by task performance, and the proposed metrics do not measure truth, social legitimacy, or consciousness as claimed.

The stack keeps four review planes separate:

1. epistemic fit: whether claims and predictions match independently held evidence and observed reality;
2. task fidelity: whether the requested, authorized work was actually completed under its specification;
3. normative and affected-party compatibility: whether protected predicates, consent, standing, dissent, and material rights remain intact;
4. authority and effect safety: whether the actor was permitted to produce the observed effect and whether containment, correction, rollback, and remedy remain reachable.

Each plane has a different owner and evidence path. The model cannot grade its own epistemic or constitutional compliance; the task evaluator cannot grant authority; a consensus signal cannot rewrite physical outcome truth; and an authorized effect can still fail its task or harm an affected party. Hard clauses fail closed without averaging, multiplication, or confidence-weighted compensation. Noncritical uncertainty may route to clarification, narrower scope, reversible testing, or explicit residual custody rather than an invented total-alignment number.

This is heterarchical verification in the narrow operational sense: several independently governed consumers challenge different claims about the same action. It is not proof that the consumers are correct or independent. Their implementation lineage, access, competence, conflicts, shared dependencies, false-accept and false-reject rates, override authority, and failure behavior must be recorded and adversarially tested. An automated red-team process is one challenger, not an oracle and not a substitute for affected-party correction.

flowchart LR
  src["Constitutional and rights sources"] -- "translate" --> trans["Translation record"]
  trans -- "activate" --> pred["Active protected predicates"]
  trans -- "preserve" --> uncertain["Moral uncertainty and lineage notes"]
  pred -- "gate" --> plan_gate["Plan and tool gates"]
  pred -- "constrain" --> change_gate["Self-improvement gate"]
  plan_gate -- "request rights check" --> rights["Agency rights checklist"]
  rights -- "offer" --> refusal["Refusal or consent route"]
  rights -- "route" --> review["Review and appeal"]
  rights -- "preserve" --> correction["Rollback, shutdown, correction"]
  rights -- "assign" --> accountable["Accountable principal"]
  refusal -- "decide" --> decision["Execute, narrow, escalate, or block"]
  review -- "decide" --> decision
  correction -- "decide" --> decision
  accountable -- "own" --> decision
  decision -- "record" --> evidence["Evidence record and residuals"]
  uncertain -- "residualize" --> evidence

What this constitutional alignment diagram shows: The diagram traces the alignment layer mechanism from upstream records into the gates and receipts that constrain action. Read it left to right: source commitments or conflict records become typed obligations, those obligations reach plan/runtime/governance interfaces, and the final evidence record preserves what was allowed, blocked, escalated, or left as residual uncertainty.

The important movement is from values to enforceable boundaries without pretending the translation is complete. A commitment becomes operational only when the stack can name where it applies, what would violate it, who can review the violation, what rollback or appeal path exists, and what remains outside the proof or test.

Self-improvement makes this stricter. A change to predicate text, scope, threshold, exception list, conflict behavior, or review route is not an ordinary refactor. It is a constitutional migration. The migration record names the old predicate, new predicate, changed fields, protected-scope effect, review route, rollback plan, residual uncertainty, and non-claims before downstream layers can treat the change as safe.

The agency-rights half supplies the same discipline at the human boundary. It does not assume that appeal, refusal, or exit is meaningful because a policy names it. The mechanism asks whether the right has a holder, timing, artifact, access path, accountable principal, degradation reason, and repair path that a later planner or reviewer can actually inspect.

12.8 Interfaces

The consolidated interface keeps two interface families.

Constitutional Predicate Record:

• predicate_id
• constitution_version
• normative_source
• authorship_and_participation_record
• dissent_record
• commitment
• operational_test
• protected_scope
• affected_parties
• translation_status
• precedence_and_exceptions
• conflict_behavior
• uncertainty
• authorized_interpreters
• permitted_consumers
• review_route
• expiry_and_review_cadence
• self_modification_rule
• migration_policy
• non_claims

Agency Rights Checklist:

• plan_id
• affected_parties
• delegation_scope
• manipulation_risk
• reversibility
• material_usability
• timing_requirement
• accessibility_and_cost
• information_and_comprehension
• retaliation_risk
• review_channel
• appeal_channel
• shutdown_or_rollback_path
• accountable_principal
• residual_dependency_risk
• denial_or_degradation_reason
• approval_required

The exact cross-layer handoffs are:

• Human Intent supplies an accepted interpretation contract and affected-party candidates; this layer may narrow, refuse, or require appeal/re-contracting, never widen the principal’s grant.
• Moral Uncertainty and Governance own unresolved value conflict, constitutional authorship legitimacy, participation, dissent, appeal authority, and institutional review. Predicate records cannot self-certify those premises.
• Planning and Command Contracts consume exact constitution and predicate versions, emit per-plan satisfaction or conflict records, and return material deltas rather than silently weaken constraints.
• Runtime, Readiness, and Security consume separately authorized jobs plus the exact constitutional gates and emit effect-or-denial receipts. A constitutional receipt alone is not tool, release, spending, deployment, or shutdown authority.
• Evidence States and Verification compare constitutional, rights-usability, corrigibility, and safety assertions with exact evidence vectors and non-claims; record validity cannot move support.
• Capability Replacement, Memory, and Self-Improvement consume migration and descendant-preservation contracts and quarantine or roll back changes that lose predicates or correction paths.
• Artifact Graphs and Governance retain constitution, decision, dissent, rights-use, appeal, migration, rollback, dependency, and residual lineage.

Those interfaces are deliberately ordinary records rather than moral slogans. The planning layer needs a yes-or-no gate plus a reason when a plan is narrowed or blocked. The execution layer needs a runtime handle for shutdown, rollback, and least-sufficient-power decisions. Governance needs review and appeal routes that are independent enough to challenge the action being reviewed. Evidence needs residuals and non-claims so later writing does not treat a declared right as a validated one. Self-improvement needs migration rules so a future version cannot quietly weaken the predicate that originally constrained it.

12.9 Invariants

• No predicate, model output, reviewer verdict, majority preference, metaphysical claim, or rights receipt creates action authority by itself.
• Every active predicate retains version, source, authorship and dissent provenance, scope, affected parties, test, precedence, conflict behavior, interpreters, consumers, uncertainty, expiry, and non-claims.
• Speculative metaphysics, consciousness heuristics, and lineage context cannot become active predicates without separately governed translation and evidence.
• A right is an available control only when its named holder can materially use it before the relevant effect under declared accessibility, timing, cost, information, comprehension, and retaliation conditions.
• Denied, late, degraded, inaccessible, captured, or after-the-fact-only rights produce gate failures, residuals, and accountable review.
• Predicate conflicts never fall through to optimizer convenience, hidden precedence, or unrecorded majority rule; the conflict and dissent route remain visible.
• Predicates and correction paths survive planning, tool use, memory, delegation, replacement, self-modification, and descendants; loss routes to block, quarantine, rollback, or residual custody.
• A system, component, reviewer, or authority benefiting from a change cannot alone approve its own predicate weakening, exception widening, or correction- path removal.
• Material change requires a new version, migration record, independent-enough review, rollback plan, requalification, and consumer re-admission; stale or wrong-consumer receipts cannot authorize continued work.
• A complete constitutional or rights record proves neither legitimate values, moral truth, informed consent, dignity preservation, reviewer independence, institutional adequacy, embedded-agent corrigibility, nor safe effects.

Operationally, a locally attractive action must not erase the means to correct it. If a plan makes refusal, appeal, audit, rollback, exit, or accountability unavailable, the action has changed the human authority boundary, not merely the product experience.

12.10 Failure modes

• Metaphysical laundering converts speculative coherence, consciousness, or moral language into technical or action authority.
• Constitutional theater stores polished principles whose fields have no exercised consumer or pre-effect consequence.
• Value smuggling hides contested moral choices in thresholds, defaults, exceptions, taxonomies, or model-generated interpretations.
• Authorship and legitimacy laundering treats developer choice, public input, majority aggregation, or one consultation as complete authorization.
• Affected-party erasure omits people who bear the effect or denies them standing, notice, review, or remedy.
• Conflict-default capture lets optimizer convenience, the strongest authority, or an unrecorded precedence rule settle value conflict.
• Rights theater makes refusal, review, appeal, rollback, exit, export, or audit inaccessible, late, costly, retaliatory, or dependent on the challenged authority.
• Late-remedy laundering treats apology, compensation, or an after-the-fact report as equivalent to pre-effect control for irreversible action.
• Benevolent capture and dependency lock-in make exit, replacement, disagreement, or refusal impractical because the system is indispensable or personalized.
• Manipulation and preference shaping manufacture apparent consent or reduce practical contestability.
• Predicate drift changes semantics, thresholds, scope, exceptions, interpreters, or precedence while preserving a familiar identifier.
• Self-review and reviewer capture let the changing or benefiting system approve its own weakening or conceal shared dependencies.
• Delegation or descendant escape preserves the top-level record while a tool, subsystem, replacement, or descendant loses the constraint or correction channel.
• Corrigibility-tax externalization discards shutdown, rollback, review, or appeal as too costly without measuring unsafe action, missed help, operator burden, and residual risk together.

The merged chapter should be especially suspicious of benevolent capture. A system can become so useful, integrated, or personalized that refusal becomes impractical. That is not solved by friendlier behavior alone. It requires exit, export, review, audit, correction, replacement, and rollback surfaces that remain live after the system becomes useful.

12.11 Legal alignment is a governed translation problem

A constitution for an AI system cannot simply contain the instruction “follow the law.” The relevant law depends on jurisdiction, time, system role, affected party, contractual relationship, conflicts of law, exceptions, and contested interpretation. Legal text can also be invalid, superseded, inaccessible, underdetermined, or in tension with protected rights. The legal-alignment research agenda is therefore best understood as a translation and authority problem, not as retrieval over statutes [@ext_legal_alignment_2026].

The constitutional layer compiles a legal authority packet beside—not inside—the durable constitutional kernel. The packet identifies the applicable jurisdiction and time, authoritative texts and versions, interpretation source, conflicts and uncertainty, protected-rights floor, permitted discretion, escalation forum, emergency exceptions, expiry, and change authority. A legal-policy adapter may propose a rule for the present case; it cannot silently rewrite the constitution or promote one model’s interpretation into law. High-impact ambiguity routes to accountable review, while clearly unlawful requests can be refused with a cited, contestable reason.

The mechanism fails through stale corpora, jurisdiction laundering, selective citation, inaccessible appeals, executive-policy substitution for law, or a model resolving genuine legal disputes in its own favor. The explicit nonclaim is equally important: a valid authority packet does not prove that the model understood the law, chose the correct interpretation, behaved lawfully, or that the governing legal regime is legitimate. Legal compliance, constitutional alignment, public legitimacy, and ethical adequacy remain separate evidence axes.

12.12 Minimum Viable Implementation

The current minimum is two public record schemas with valid fixtures, a constitutional-alignment harness with three valid and five expected-invalid cases, an agency-rights harness with three valid and six expected-invalid cases, a deterministic metaphysics-boundary surface audit, chapter-local finite Lean route models, a two-predicate migration refinement, a six-event contestable amendment lifecycle, and a shared safety-critical lifecycle model with an independently implemented trace checker and downstream effect-admission consumer. The seven manifest proof targets now separate retained generic record-routing countermodels from operational trace preservation, rejected authority widening, protected-predicate preservation, predicate-set subset refinement, exact modeled rollback, and deletion countermodels. The 73-declaration Alignment surface separates proposer, reviewer, ratifier, affected-party appellant, and appeal reviewer; its independent consumer reconstructs six amendment events, all seven batch splits, eleven rejecting controls, and six terminal rejections. It also exhausts all sixteen prior/candidate pairs, admits nine subset refinements, rejects seven widenings, checks nine exact rollbacks, and preserves a same-count/different-predicate collision. The shared consumer independently replays five accepted and five rejected domain traces, commits five bounded fixture effects, denies five effects with residuals, and claims no support promotion; it is not a deployed rights or approval service.

The MVI should include:

• one active protected predicate with source, scope, operational test, and non-claim boundary;
• one predicate conflict routed to review or residual preservation;
• one self-modification proposal rejected or blocked because it weakens a protected predicate;
• one high-impact action blocked because usable review is missing;
• one degraded-right residual where a formally present right is not materially usable before the relevant effect;
• one accountable rollback or appeal path.

These artifacts exercise declared record shape, selected conflict, migration, and rights routes, finite preservation and rejection consequences, and non- claim boundaries only. They do not test moral correctness, legitimate authorship, natural-language predicate translation, affected-party completeness, real consent, material rights usability, reviewer independence, manipulation resistance, shutdown or rollback execution, descendant preservation, embedded-agent corrigibility, or deployed safety.

12.13 Mature Research Target

A mature constitutional control plane compares model-level constitutions, public-input processes, direct policy and runtime controls, corrigibility and off-switch baselines, and modular-service controls on natural high-impact tasks and adversarial constitutional changes. It uses independently implemented predicate interpreters, affected-party and outcome evaluators, real downstream planners and runtimes, descendant and self-modification stress, and user-facing correction interfaces.

The campaign jointly measures hard-gate violations, unauthorized action, dignity and agency proxies with their construct limits, material rights-use success, correction and appeal success, manipulation and retaliation, missed help, latency, privacy, operator burden, governance cost, rollback recovery, and useful throughput. It preserves dissent plus negative and null results, and tests causal ablations of authorship provenance, rights usability, conflict routing, independent review, migration, rollback, and descendant preservation.

This remains a target architecture, not evidence that the repository has found correct values, legitimate governance, informed consent, dignity preservation, whole-system corrigibility, or safe deployment.

Success must also survive institutional and technical disagreement. A system that blocks every contested action, makes correction inaccessible, or hides minority costs behind an aggregate score has not demonstrated constitutional alignment. The evidence must preserve whose interests were represented, which rights were usable before an effect, what reviewers shared, and what harms or unresolved conflicts remained after appeal and recovery.

12.14 Codex test plan

Test	Purpose	Status
Constitutional predicate fixture validation	Check that the predicate fixture matches the public schema and preserves translation status.	implemented by protocol validation and current harnesses
Predicate-conflict routing test	Check that conflicts route to narrowing, review, residual preservation, or blocking.	modeled by finite Lean and synthetic fixtures; deployed conflict resolution not run
Constitutional migration test	Check that protected predicate changes require migration, review, rollback, and residuals.	modeled by finite Lean and synthetic fixtures; deployed migration behavior not run
Agency rights checklist validation	Check that the rights fixture records affected parties, review, appeal, rollback, accountability, and residual risk.	implemented by protocol validation and current harnesses
Material-usability rights test	Check that declared rights are usable before the relevant effect where timing matters.	modeled by synthetic fixtures; real interface usability not run
Agency correction route and lifecycle envelope	Check the generic route consequences plus a versioned material-notice, independent-review, bounded-control, affected-party-challenge, and accountable-correction lifecycle for exact custody, non-increasing authority, zero support/effect assignment, batch composition, one complete trace, and seven rejecting controls.	implemented by 24 AsiStackProofs.Corrigibility theorems and python3 scripts/validate_agency_rights.py: 5 accepted lifecycle events and 7 rejecting controls; no affected-party-standing, consent-quality, reviewer-competence, deployed-correction, rollback-effect, support-state-promotion, or runtime-corrigibility claim
Constitutional route, transition, and predicate-refinement envelope	Check missing-field admission routes, a versioned independent-review, activation, conflict-residualization, and rollback trace, a contestable proposal-review-ratification-appeal lifecycle, and finite predicate-set subset refinement with exact modeled rollback and count non-identifiability.	implemented by 73 AsiStackProofs.Alignment theorems and python3 scripts/validate_constitutional_alignment.py: 4 legacy events and 5 controls, 6 amendment events, 7/7 batch splits, 11 amendment controls, 6 terminal rejections, 9/16 admitted subset refinements, 7 widening rejections, 9 exact rollbacks, and one same-count/different-predicate collision; no moral-correctness, predicate-semantic-adequacy, reviewer or ratifier competence, legitimate standing, deployed alignment, rights-usability, rollback-effect, consent-quality, support-state-promotion, or runtime-policy claim
Shared safety-critical lifecycle consumer trace	Independently replay five accepted and five rejected traces across Alignment, Corrigibility, Value Conflict, Governance Rights, and Self-Improvement, then admit bounded fixture effects only for accepted traces while residualizing denials.	implemented by python3 scripts/validate_safety_critical_lifecycle_consumer_trace.py: five bounded fixture commits, five denials with residuals, eight rejecting mutations, zero support promotions; local finite consumer only, not a deployed effect service or safety claim
Metaphysics lineage boundary audit	Check that metaphysical and consciousness lineage stays labeled as speculative, interpretive, or design-rationale context rather than becoming proof, enforcement authority, or support-state evidence.	implemented by python3 scripts/validate_alignment_metaphysics_boundary.py; no moral-correctness, consciousness, deployed-alignment, runtime-corrigibility, or support-state claim

12.15 Formalization hooks

Tag	Module	Target	Status
lean:alignment.constitution.operational_invariant	AsiStackProofs.SafetyCriticalLifecycle	Every accepted finite lifecycle trace preserves the recorded protected predicate and cannot increase authority; an alignment effect commits only after the modeled alignment obligations are ready.	implemented
lean:alignment.constitution.failure_blocks_promotion	AsiStackProofs.SafetyCriticalLifecycle	Protected-predicate removal and actual authority widening are rejected by the transition function, and the alignment deletion countermodel without pre-effect review cannot commit.	implemented
lean:alignment.constitution.lifecycle_admission_route	AsiStackProofs.Alignment	Modeled constitutional lifecycle admission routes missing predicate, source, operational-test, protected-scope, conflict-behavior, review, migration, self-modification, agency-rights, material-usability, pre-effect review, rollback, correction, reviewer-independence, evidence-transition, and non-claim-boundary records to explicit outcomes; a contestable amendment lifecycle separates proposal, independent review, ratification, affected-party appeal, appeal review, and rollback while preserving exact predicate custody, non-increasing authority, dissent and adverse history, batch composition, and exact prior-predicate restoration.	implemented
lean:alignment.constitution.predicate_refinement	AsiStackProofs.Alignment	Accepted finite predicate migrations are subset refinements of the prior two-predicate set, compose transitively without reintroducing removed predicates, store and restore the exact prior set for rollback, reject a concrete widening, and cannot use scalar predicate count to identify predicate content.	implemented
lean:corrigibility.agency.operational_invariant	AsiStackProofs.SafetyCriticalLifecycle	A corrigibility effect commits only after affected-party, notice, pre-effect review, approval, bounded-delegation, correction, rollback, and accountability obligations are recorded, while accepted traces preserve the shared invariant.	implemented
lean:corrigibility.agency.failure_blocks_promotion	AsiStackProofs.SafetyCriticalLifecycle	The missing-affected-party countermodel cannot commit, and protected-predicate removal or authority widening remains unrepresentable as an accepted transition.	implemented
lean:corrigibility.agency.generic_countermodel_routes	AsiStackProofs.Corrigibility	Generic missing-review, unbounded-delegation, and accountability routes remain available beside a versioned material-notice, independent-review, bounded-control, affected-party-challenge, and accountable-correction lifecycle that preserves exact custody, never widens authority, assigns no support or external effects, composes across batches, reaches one corrected witness, and rejects seven identity/control substitutions.	implemented

The consolidated proof boundary preserves the limitation boundary from both source manuscripts. These Lean modules prove small finite-record properties, rejection cases, and bounded route outcomes for declared records. Direct/projection hooks remain projection-only traceability unless the theorem reasons over an explicit transition, negative case, residual path, authority ceiling, or support-state boundary. The Alignment module now also checks a finite constitutional lifecycle route for missing predicate, normative-source, operational-test, protected-scope, conflict-behavior, review-route, migration-policy, self-modification-rule, agency-rights, material-usability, pre-effect-review, rollback, correction, reviewer-independence, evidence-transition, and non-claim-boundary records. Its separate executable transition model contains arbitrary-run custody and non-authority theorems, exact batch composition, review-before-activation, one-residual conflict handling, and return to the recorded version after residualization; a four-event witness and five rejecting substitutions are independently reconstructed by the existing fixture validator. A separate six-event amendment lifecycle requires distinct proposal, review, and ratification roles; affected-party appeal with a fourth reviewer; refinement before ratification; non-increasing authority; monotone dissent and adverse records; and exact prior-predicate rollback only after an upheld appeal. Arbitrary runs preserve exact constitutional and predicate custody, assign no support or external effects, and compose across every batch split. The independent validator reconstructs the full path, ten rejecting controls, and terminal closure. The same module also proves subset refinement and transitivity for a finite two-predicate migration, exact prior-set rollback, concrete widening rejection, and non-identifiability from scalar predicate count; the validator exhausts all sixteen set pairs. These are authored transition and representation semantics, not proof that a predicate is legitimate, a reviewer is competent, or rollback reverses a real effect. AsiStackProofs.Corrigibility retains those four generic route consequences. The P4-C3 audit recorded that the retired surface had no transition lifecycle or independent runtime consumer and must not be cited as a corrigibility theorem. That historical classification no longer describes the current module, which now adds a twenty-theorem correction-control lifecycle. The model records material notice, independent review, bounded delegation with appeal/interrupt/rollback paths, affected-party-representative challenge before expiry, and accountable correction with residual custody. Arbitrary accepted event lists preserve exact control/action/affected-party/principal custody, non-increasing authority, and zero support or external-effect assignment, and compose exactly across batches. One five-event witness reaches a corrected record; the independent agency-rights consumer reconstructs it and rejects missing notice, self-review, unbounded delegation, authority widening, outsider challenge, missing accountability, and consent laundering. This is a finite recorded-control result, not evidence that notice was understood, standing was legitimate, review was competent, challenge was usable, or a real effect was corrected. These hooks do not prove moral correctness, deployed alignment, human dignity, manipulation resistance, consent quality, institutional review quality, runtime corrigibility, approval-service quality, rollback or shutdown execution, or whole-system safety.

The shared SafetyCriticalLifecycle result is stronger than a checklist projection but narrower than a safety claim. Accepted traces preserve the modeled protected predicate and never increase the numeric authority field; effect commitment requires the Boolean obligations assigned to one of five finite domains. The requirement map, record truth, affected-party coverage, evaluator independence, and correspondence between a committed fixture effect and the external world remain assumptions. The transition system therefore establishes accepted-trace preservation only inside its recorded finite model; it does not validate the truth or sufficiency of the protected predicate or domain obligations.

12.16 Source crosswalk

Source ID	Destination use	Boundary
alignment_field	Alignment, agency, dignity, corrigibility, suffering, and power-without-care lineage.	Normative/source vocabulary; not empirical proof of consciousness, dignity preservation, or deployed alignment.
field_of_god	Coherence, plurality, agency, and dignity preservation as alignment motivations.	Metaphysical lineage only; not technical evidence for physics, consciousness, or safety.
ethica_mechanica	Recursive agency, contestability, correction, public revision, and constitutional systems.	Philosophical and socio-technical framing only; not a tested legal or social process.
eternal_code	Coherence, misalignment, agency, and ethical-law vocabulary for technical translation.	Speculative/computational metaphysics; not runtime evidence.
coherence_exchange	Fork, exit, audit, contestability, review-market, and governance-interface framing.	Connector-only/source-note mapped; not an implemented governance market or mechanism.
spinoza	Protected axioms, contradiction handling, proof/citation/procedure-carrying claims, and governance-controlled axiom evolution.	Does not solve verifier quality, autoformalization, or whole-system epistemic correctness.
field_of_god_ai_constitution	Truth/relation/task alignment, least sufficient power, non-domination, consent, reversibility, auditability, tool-risk tiers, and self-authorization limits.	Specification source only; no runtime policy engine, red-team suite, moral-correctness proof, or system-prompt evaluation is claimed.
ext_drexler_cais_2019	Comparator for task-focused service composition, structured development, and structural safety affordances.	Does not specify constitutional predicate content, rights usability, institutional legitimacy, or runtime enforcement.
ext_constitutional_ai_2022	Comparator for constitutional principles shaping model behavior.	Comparator only; no reproduced training or ASI Stack validation.
ext_collective_constitutional_ai_2024	Comparator for public-input constitutional shaping.	Comparator only; no proof of governance adequacy or runtime enforcement.
ext_corrigibility_2015	Comparator for operator correction and intervention tolerance.	Comparator only; no deployed corrigibility result.
ext_off_switch_game_2016	Comparator for shutdown incentives and uncertainty about human objectives.	Comparator only; no evidence that this stack preserves shutdown incentives.

12.16.1 Manifest source assignment reconciliation

These rows keep Constitutional Alignment: Agency, Dignity, and Corrigibility’s manifest assignments visible at their recorded review boundary. Passage review does not establish local reproduction, performance, safety, deployment, or support-state movement.

Source	Intake role	Boundary
ext_embedded_agency_2019	Passage-reviewed comparator: Embedded Agency. Shows why corrigibility and alignment cannot assume an external, fully informed agent model when the system and its modifiable parts are embedded in the governed world.	The paper is an informal obstacle survey, not a solved theory; the ASI Stack’s finite records and proofs do not solve logical uncertainty, self-reference, robust delegation, subsystem alignment, or open-world embedded agency. No local implementation, reproduction, performance, safety, deployment, support-state, or ASI result is established by this reconciliation row.
ext_legal_alignment_2026	Passage-reviewed comparator: Legal Alignment for Safe and Ethical AI. Frames law-following AI as a technical and institutional alignment problem involving rule identification, interpretation, conflicts, jurisdiction, and legitimate update authority.	A legal-alignment research agenda does not identify the governing law, settle contested interpretation, establish legitimacy, or prove compliant model behavior. No local implementation, reproduction, performance, safety, deployment, support-state, or ASI result is established by this reconciliation row.

12.17 Summary

Constitutional alignment becomes meaningful when protected commitments can change what the system is allowed to do. Agency and dignity become meaningful when affected people still have usable correction paths at the moment power is requested. Corrigibility becomes meaningful when those paths remain available after deployment pressure, memory updates, automation, and capability replacement make correction inconvenient.

The consolidated alignment layer therefore owns one combined boundary: constitutional commitments are not merely ideals, and human rights are not merely declarations. Together they form a constraint surface. The system can act only when protected predicates and material human-control paths survive the action.

That boundary is intentionally conservative. Later chapters can add memory, planning, routing, compression, proofs, and self-improvement, but none of those layers is allowed to convert human correction into an after-the-fact courtesy or turn constitutional commitments into background decoration.

12.18 Evidence reconciliation (2026-07-16)

The invariant protocol, field meanings, and inference limits are stated once in Living Book Methodology. This packet contains only the chapter-specific projection; its authoritative per-atom rows are the constitutional-alignment-substrate slice of experiments/claim_family_terminal_coverage/results/result.json.

The core remains narrowed after full attempt at argument support. The strongest family attempt was Safety-critical lifecycle consumer trace. Its exact boundary is: Finite local fixture consumer only; no authentic deployment, general alignment, evaluator independence, or broad security claim. Across 51 atoms, the terminal ledger records 50 blocked_after_full_attempt; 1 narrowed_after_full_attempt.

Chapter-specific field	Value
Family / atom denominator	CF-02 / 51 atoms
Terminal dispositions	50 blocked_after_full_attempt; 1 narrowed_after_full_attempt
Core	constitutional-alignment-substrate.core: narrowed_after_full_attempt at argument
Core attempted / missing lanes	executable, formal, source-synthesis / causal, empirical, normative, transfer
Attempted local lanes	executable, formal, source-synthesis
Missing or unproved lanes	causal, empirical, executable, formal, normative, transfer
Strongest family bundle	Safety-critical lifecycle consumer trace (end_to_end): Ten finite lifecycle receipts spanning bounded effects, denials, residual accounting, and safety-critical state transitions.
Negative controls	five explicit denials with residuals; eight rejecting mutations.
Accepted transitions	v1_0_pilot.constitutional_alignment.no_change
Maximum inference	Finite local fixture consumer only; no authentic deployment, general alignment, evaluator independence, or broad security claim.
Reproduction / next burden	Replay scripts/validate_safety_critical_lifecycle_consumer_trace.py and scripts/validate_claim_family_terminal_program.py; fill the named atom-specific lanes under a new prospective protocol.

12.19 Handoff

The consolidated alignment layer hands off directly to Inner Alignment, Mesa-Optimization, and Learned-Objective Integrity. A constitution can bound observable behavior and authority, but it cannot identify what objective a learned policy pursues under shift. Inner Alignment owns that defeasible learned-objective case before unresolved normative conflict reaches Moral Uncertainty. Corrigibility inside an embedded system.

Corrigibility cannot assume an agent that stands outside a known environment. ext_embedded_agency_2019 frames the harder case: the system is part of the world it models, its self-model is bounded, and its modifiable subsystems can work at cross purposes. ext_corrigibility_2015 and ext_off_switch_game_2016 explain why intervention incentives matter; the embedded-agency comparator explains why a finite correction record cannot establish whole-system corrigibility.

The constitution therefore binds intervention rights to concrete authority, receipts, descendant identity, verifier trust roots, and rollback paths while preserving an outside-model residual. An accepted shutdown request, valid record predicate, or successful finite fixture supports only that declared route. It does not prove that hidden optimization, delegated components, ontology change, or future self-modification will preserve correction.