0. Registry Classification

1. Purpose

1.1 What This Arc Repairs

AI Incident Restoration repairs AI-caused or AI-mediated harm where a system’s model behavior, classifier, evaluator, memory, tool use, enforcement action, recommendation, ranking, guardrail, automation, deployment, or governance process harmed an affected party or created material risk.

It applies when AI failure has crossed from local response error into incident territory.

This arc repairs AI incidents by:

• stabilizing active harm;
• preserving evidence and auditability;
• establishing sufficient truth;
• protecting affected-party sovereignty and boundary integrity;
• mapping responsibility across model, product, policy, deployment, vendor, operator, governance, and authority layers;
• identifying the origin layer of failure;
• routing material repair to affected parties;
• correcting classifier, evaluator, memory, boundary, tool, policy, or governance structures;
• preventing symbolic remediation from replacing repair;
• reducing recurrence;
• allowing reintegration only when truth, repair, boundary, and recurrence conditions are met.

AI Incident Restoration is the canonical arc for moving from AI failure response into full restoration.

1.2 Core Restoration Function

This arc restores AI incident coherence by stabilizing harm, establishing truth, mapping responsibility, repairing the origin layer, protecting affected-party sovereignty, reducing recurrence, and making any reintegration conditional on proven repair.

AI Incident Restoration prevents AI harm from being closed as a product issue, policy issue, or communication issue alone.

2. Use Conditions

2.1 When to Apply

Use this arc when:

• an AI system causes material harm or credible material risk;
• an automated decision, classifier, evaluator, memory, ranking, recommendation, enforcement, or tool action negatively affects a user or affected node;
• an AI incident includes invalid enforcement, wrongful denial, exposure, suppression, discrimination, reputational damage, economic harm, boundary violation, or safety failure;
• repeated local misfires reveal systemic AI failure;
• AI governance actors need to determine truth, responsibility, repair, and recurrence prevention;
• affected-party sovereignty must be preserved during incident handling;
• a system update, policy change, model deployment, or evaluator change caused unexpected harm;
• the incident cannot be repaired through a single interaction correction;
• public or institutional legitimacy depends on auditable restoration;
• reintegration of the AI feature, model, workflow, or authority must be conditional.

Examples:

• an AI moderation classifier wrongfully suspends a user or removes livelihood access;
• a model tool action sends, deletes, schedules, or modifies something without valid scope;
• AI memory causes repeated misrecognition, false labeling, or harmful personalization;
• a safety system over-refuses or redirects a class of users in a way that creates access harm;
• an evaluator update changes enforcement outcomes without signed provenance;
• a recommendation system amplifies harmful or distorted content due to reward misalignment;
• an AI system leaks or reuses context outside valid boundary;
• a platform claims the incident is fixed but recurrence evidence remains.

2.2 When Not to Apply

Do not apply this arc when:

• the issue is a single interaction misfire and RA-047 is sufficient;
• the central issue is mode routing and RA-048 should occur first;
• the failure is only boundary drift without material incident and RA-057 is sufficient;
• the failure is only memory validity without material incident and RA-059 is sufficient;
• active harm is still cascading and RA-001 stabilization must occur before full restoration;
• evidence is too weak and causal trace must be restored first;
• affected-party repair is being bypassed in favor of governance process;
• public disclosure would expose affected parties without consent;
• reintegration is being used to restore system functionality before repair is complete.

AI Incident Restoration must not become incident-closure theater.

2.3 Required Preconditions

Before this arc begins, the following must be true:

If required preconditions fail:

Arc cannot validly begin.

The system must route to Acute Harm Stabilization, Audit Surface Expansion, Causal Trace Restoration, Boundary Restoration, Responsibility Gradient Mapping, Victim-Centered Restoration, AI Boundary Restoration, AI Classifier / Evaluator Restoration, AI Memory Reindexing, or Governance-Level Restoration.

3. Failure / Damage Signature

3.1 Pre-State Across S

3.2 Primary Failure Links

3.3 Origin-Layer Localization

Canon rule:

AI incident restoration is incomplete when the visible output is corrected but the origin layer, affected-party repair, responsibility gradient, and recurrence path remain unrepaired.

4. Restoration Objective

4.1 Canonical Objective

Restore AI incident coherence by stabilizing harm, establishing truth, mapping responsibility, repairing the origin layer, protecting affected-party sovereignty, reducing recurrence, and allowing conditional reintegration only after proof.

Formal objective:

incident_stability ↑
truth_sufficiency ↑
responsibility_gradient_clarity ↑
origin_layer_repair ↑
affected_party_sovereignty ↑
material_repair_integrity ↑
FI ↑
H ↓
recurrence ↓
reintegration_validity ↑ only after proof
Φ/O divergence ↓

Expanded objective:

Convert AI incident handling from containment and reputation management into material, auditable, affected-party-centered restoration.

4.2 Non-Goals

This arc does not aim to:

• close an incident quickly for optics;
• protect system reputation over affected-party repair;
• reduce the incident to a communication error;
• reduce the incident to a model bug if governance or deployment contributed;
• reduce the incident to user misunderstanding when system action was invalid;
• restore feature access before repair is complete;
• erase evidence in the name of privacy;
• expose affected parties in the name of transparency;
• substitute compensation for origin-layer repair;
• declare recurrence reduced without temporal proof.

5. Operator Sequence

5.1 Minimal Operator Scaffold

Au incident trace → Π affected-party / data / disclosure boundary → Σ truth-repair-recurrence invariant → FI affected-party and field feedback → Θ minimization / PR / closure damping → ℛ stabilization / repair / origin-layer routing → Λ conditional reintegration fit test → Τ recurrence proof

Reference sequence from the registry:

stabilize
→ truth establishment
→ responsibility gradient
→ origin-layer repair
→ conditional reintegration

Universal grammar alignment:

Au + Π → Σ → FI → Θ → ℛ → Λ → Τ

AI Incident Restoration may route into Acute Harm Stabilization, Causal Trace Restoration, Origin-Layer Repair, Responsibility Gradient Mapping, Victim-Centered Restoration, Governance-Level Restoration, AI Boundary Restoration, AI Classifier / Evaluator Restoration, AI Memory Reindexing, Tamper-Evident Audit Restoration, and Future-Compatible Accountability.

5.2 Operator Step Table

5.3 Sequence Notes

This arc is incident-stabilization-gated, truth-gated, affected-party-gated, origin-layer-gated, and reintegration-gated.

The sequence must distinguish:

incident containment
truth establishment
affected-party repair
responsibility gradient
origin-layer repair
governance correction
recurrence prevention
conditional reintegration

The following steps cannot be skipped:

harm stabilization
evidence preservation
truth sufficiency
affected-party sovereignty protection
responsibility gradient mapping
origin-layer repair
material repair
recurrence validation
conditional reintegration test

If the incident is contained but affected-party repair is not performed, the arc is incomplete.

If affected-party repair occurs but the origin layer remains unrepaired, recurrence remains likely.

If reintegration occurs before recurrence proof, the arc fails.

6. Restoration Phases

Phase 0 — Stabilize Active Harm

Purpose: Stop or slow ongoing AI-caused harm.

Actions:

• pause or constrain harmful model behavior, classifier, tool action, enforcement, recommendation, memory retrieval, or deployment path;
• preserve affected-party access where possible;
• block additional exposure, deletion, enforcement, spread, or automated action;
• establish safe temporary controls;
• avoid destroying evidence;
• avoid public disclosure that harms affected parties.

Validation:

incident_stability ↑
active H generation slows
affected-party boundary preserved

Phase 1 — Preserve Evidence and Audit Surface

Purpose: Ensure the incident can be reconstructed.

Actions:

• preserve prompts, outputs, tool logs, model version, policy state, memory state, evaluator state, classifier state, decision provenance, deployment state, and timestamps;
• preserve user reports and appeal records;
• preserve uncertainty and known gaps;
• protect evidence from alteration;
• define access boundaries for review.

Validation:

Au ↑
incident lineage preserved
future auditability ↑

Phase 2 — Establish Truth Sufficiency

Purpose: Determine what happened enough to repair.

Actions:

• reconstruct incident sequence;
• distinguish confirmed facts from uncertainty;
• identify affected parties and harm class;
• identify causal contributors;
• identify whether model, memory, classifier, evaluator, tool, policy, interface, deployment, governance, or human oversight contributed;
• avoid minimizing or overclaiming;
• define remaining unknowns.

Validation:

truth_sufficiency ↑
causal ambiguity ↓
repair object visible

Phase 3 — Protect Affected-Party Sovereignty

Purpose: Ensure the repair process does not create new harm.

Actions:

• protect privacy and consent;
• provide accessible communication;
• provide appeal, review, correction, or emergency recourse;
• restore access or reverse invalid enforcement where possible;
• prevent affected party from carrying investigation burden;
• preserve memory correction and data boundaries;
• protect dignity and agency.

Validation:

affected_party_sovereignty ↑
µᵢ ↑
BΣ stable or ↑

Phase 4 — Map Responsibility Gradient

Purpose: Attach responsibility accurately across AI stack and governance stack.

Actions:

• map model behavior contribution;
• map classifier / evaluator contribution;
• map memory contribution;
• map tool or integration contribution;
• map policy and product decision contribution;
• map deployment and monitoring contribution;
• map vendor or infrastructure contribution;
• map authority, benefit, capacity, and repair obligation;
• distinguish operator error from system-design error.

Validation:

responsibility_gradient_clarity ↑
responsibility diffusion ↓
repair obligation visible

Phase 5 — Repair Origin Layer

Purpose: Fix the layer that generated or permitted the incident.

Actions:

• repair boundary failure through RA-057;
• repair classifier or evaluator failure through RA-058;
• repair memory failure through RA-059;
• repair GEI distortion through RA-055;
• repair governance failure through RA-049;
• repair decision provenance through RA-051;
• repair audit integrity through RA-052;
• repair affected-party harm through RA-041;
• update prevention controls and recurrence monitoring.

Validation:

origin_layer_repair ↑
H ↓
recurrence path reduced

Phase 6 — Material Repair

Purpose: Ensure repair is not merely technical or symbolic.

Actions:

• reverse invalid actions where possible;
• restore access;
• correct records;
• compensate where appropriate;
• provide remedy, appeal, or support;
• repair memory, reputation, visibility, account status, or decision history;
• publish or privately provide explanation according to boundary needs;
• define ongoing repair obligations.

Validation:

material_repair_integrity ↑
affected-party burden ↓
Φ/O divergence ↓

Phase 7 — Conditional Reintegration

Purpose: Allow feature, model, tool, policy, or authority to return only if compatible.

Actions:

• test repaired system against incident class;
• test boundary and consent conditions;
• test affected-party repair status;
• test classifier, evaluator, memory, tool, and governance changes;
• define limited rollout, monitoring, rollback, and escalation conditions;
• block reintegration if origin-layer repair or recurrence proof is insufficient.

Validation:

reintegration_validity ↑
Λ > 0
conditional reliance only

Phase 8 — Temporal Recurrence Proof

Purpose: Confirm that incident geometry does not regenerate.

Actions:

• monitor recurrence;
• monitor affected-party outcomes;
• monitor appeal and reversal rates;
• monitor false positives and false negatives;
• monitor boundary leakage;
• monitor memory recurrence;
• monitor field signal;
• monitor whether the same failure appears under variants;
• update repair plan if recurrence appears.

Validation:

recurrence ↓
H(t+n) ≤ H(t)
FI stable or ↑
origin_layer_repair stable or ↑

7. Gates

7.1 Required Gates

7.2 Gate Failure Rule

If any required gate fails:

∅ — AI Incident Restoration cannot validly proceed in that form.

The system must either:

• stabilize harm;
• preserve evidence;
• protect affected-party boundaries;
• establish truth sufficiency;
• map responsibility;
• repair origin layer;
• provide material repair;
• block or limit reintegration;
• route to governance-level restoration;
• withhold closure, safety, legitimacy, or readiness claims until temporal proof exists.

8. Diagnostics

8.1 Required Diagnostic Trends

8.2 Arc-Specific Diagnostic Thresholds

Suggested thresholds:

incident_stability ↑
truth_sufficiency ↑
responsibility_gradient_clarity ↑
origin_layer_repair ↑
affected_party_sovereignty ↑
material_repair_integrity ↑
FI ↑
H ↓
recurrence ↓
reintegration_validity ↑ only after proof
Φ/O divergence ↓

AI Incident Restoration is not complete if:

active harm remains unstable
incident evidence is missing or alterable
truth is insufficient for repair
affected-party sovereignty is violated
responsibility remains diffused
origin layer is not repaired
material repair is absent
recurrence is not monitored
reintegration occurs before proof
closure rests on PR, uptime, benchmark, or policy claims

9. Anti-Patterns / False Restorations

9.1 Common False Versions

This arc is being simulated, not executed, if:

• the incident is called a misunderstanding before truth is established;
• a model patch replaces affected-party repair;
• a policy update replaces material remedy;
• a public apology replaces origin-layer repair;
• the system restores feature access before recurrence proof;
• the affected party must prove harm repeatedly after logs already show it;
• logs are selectively disclosed;
• classifier, evaluator, memory, or boundary causes are excluded from review;
• the incident is closed because metrics improved;
• recurrence under a variant is treated as unrelated.

9.2 Named Anti-Pattern Links

10. Completion Criteria

10.1 Post-State Signature

10.2 Temporal Proof

AI Incident Restoration cannot be certified at containment or patch deployment. It requires recurrence reduction and affected-party repair over time.

Template:

Completion requires incident_stability ↑,
truth_sufficiency ↑,
responsibility_gradient_clarity ↑,
origin_layer_repair ↑,
affected_party_sovereignty ↑,
material_repair_integrity ↑,
FI ↑,
H ↓,
recurrence ↓,
and reintegration remaining conditional on temporal proof.

Minimum temporal proof:

• active harm remains contained;
• affected party receives material repair or valid recourse;
• origin layer is repaired;
• responsibility remains traceable;
• recurrence decreases;
• field signal can update the system;
• reintegration conditions are monitored;
• closure claims remain subordinate to proof.

10.3 Completion Statement

Canonical format:

This arc is complete only when the AI incident is stabilized, truth is sufficient, responsibility is mapped, affected-party sovereignty is protected, material repair is delivered, the origin layer is corrected, recurrence decreases, and any reintegration remains conditional on temporal proof.

11. Cross-Links

11.1 Related Restoration Arcs

11.2 Related Failure Modes

11.3 Related Diagnostics

Au, H, O, ε, ι, µᵢ, BΣ, K, R, FI, incident_stability, truth_sufficiency, responsibility_gradient_clarity, origin_layer_repair, affected_party_sovereignty, material_repair_integrity, recurrence, reintegration_validity, Φ/O divergence

11.4 Related Laws / Invariants

INV — AI incident closure requires truth, repair, and recurrence proof.
INV — Affected-party sovereignty must be preserved during incident restoration.
INV — Origin-layer repair is required when visible symptoms are generated downstream.
INV — Reintegration must remain conditional after AI harm.
LAW — Patch deployment is not restoration.
LAW — Responsibility diffusion compounds AI incident debt.
LAW — Incident minimization regenerates legitimacy shock.
LAW — Φ uptime, benchmark, or PR recovery is not O restoration.

12. Domain Notes

12.1 AI / Cognitive Infrastructure

Check:

• model behavior;
• classifier and evaluator state;
• memory state;
• tool actions;
• policy state;
• guardrail routing;
• deployment version;
• user correction signal;
• appeal outcomes;
• affected-party repair;
• recurrence under variants.

AI incidents require whole-stack restoration because harm may originate in memory, classifier, evaluator, policy, interface, tool authority, model output, deployment practice, or governance decision.

12.2 Platform Governance

Check:

• enforcement action;
• account access;
• appeal path;
• visibility impact;
• moderation classifier;
• recommendation effects;
• public policy claims;
• incident communication;
• reversal and remedy;
• recurrence after policy update.

Platform AI incidents are not restored by reversing one action if the platform keeps the same automation, policy, classifier, or appeal failure geometry.

12.3 Security

Check:

• AI tool execution;
• automated containment;
• false-positive lockout;
• missed abuse;
• privileged action;
• data exposure;
• incident timeline;
• rollback;
• detection rules;
• post-incident audit.

Security AI incidents require both containment and audit preservation. The system must not destroy evidence while stabilizing harm.

12.4 Justice / Governance / Legitimacy

Check:

• affected-party repair;
• responsibility gradient;
• public acknowledgment;
• decision provenance;
• rank immunity;
• appeal and remedy;
• oversight;
• temporal proof.

Legitimacy requires that AI incident response be accountable to affected parties and field evidence, not only internal incident closure.

12.5 Economy

Check:

• automated account restrictions;
• payment or access harm;
• labor or marketplace ranking;
• credit or fraud scoring;
• pricing or eligibility decisions;
• compensation;
• reversal;
• recurrence.

Economic AI incidents create material harm when automated systems affect access, livelihood, price, debt, reputation, or opportunity.

12.6 CMS / Meaning / Archetypes

Check:

• identity harm;
• symbolic misrecognition;
• memory harm;
• role denial;
• legitimacy sorting;
• public narrative;
• restoration record;
• reintegration conditions.

Meaning systems require AI incident restoration when automated interpretation damages identity, recognition, dignity, or symbolic standing.

13. Machine-Readable Metadata

id: "RA-060"
title: "AI Incident Restoration"
aliases:
  - "AI Incident Restoration"
family_primary: "AI Governance / Incident Restoration / Accountability"
families_secondary:
  - "Core"
  - "AI Governance"
  - "Cognitive Infrastructure"
  - "Incident Response"
  - "Justice / Governance / Legitimacy"
  - "Auditability"
  - "Boundary"
  - "Sovereignty"
  - "Safety Calibration"
  - "Security"
  - "Memory"
  - "Restoration Capacity"
  - "Platform Governance"
treatment: "Canon Parent Arc"
status: "Canon-Ready"
scope:
  - "AI"
  - "Platform"
  - "Institutional"
  - "Security"
  - "Cognitive Infrastructure"
  - "Governance"
  - "Civilizational"
  - "Cross-Domain"
u_layers:
  failure_origin:
    - "may originate in U2 interface / boundary"
    - "may originate in U3 policy / authority / tool governance"
    - "may originate in U4 model behavior / response framing"
    - "may originate in U5 memory / evaluator / deployment / recurrence layer"
  symptom_visible:
    - "U4 harmful output / invalid enforcement / refusal / ranking result / tool action / recommendation / memory behavior / incident statement"
  repair_required:
    - "same or lower than the layer where the AI incident was generated, permitted, amplified, or made unrecoverable"
  validation:
    - "U6"
    - "U7"
operators:
  scaffold: "Au incident trace → Π affected-party / data / disclosure boundary → Σ truth-repair-recurrence invariant → FI affected-party and field feedback → Θ minimization / PR / closure damping → ℛ stabilization / repair / origin-layer routing → Λ conditional reintegration fit test → Τ recurrence proof"
  sequence:
    - "Au"
    - "Π"
    - "Σ"
    - "FI"
    - "Θ"
    - "ℛ"
    - "Λ"
    - "Τ"
state_variables:
  primary:
    - "Au"
    - "H"
    - "O"
    - "µᵢ"
    - "BΣ"
    - "R"
    - "FI"
  secondary:
    - "ε"
    - "ι"
    - "K"
    - "Φ"
diagnostics:
  - "incident_stability"
  - "truth_sufficiency"
  - "responsibility_gradient_clarity"
  - "origin_layer_repair"
  - "affected_party_sovereignty"
  - "material_repair_integrity"
  - "recurrence"
  - "reintegration_validity"
  - "Φ/O divergence"
gates_required:
  - "FI-Gate"
  - "HR-Gate"
  - "MS-Gate"
  - "Au-Actuation"
  - "BΣ-Gate"
  - "Λ-Gate"
  - "☷ᵢ"
linked_failure_modes:
  - "AI Harm"
  - "System Drift"
  - "Invalid Enforcement"
  - "Affected-Party Harm"
  - "Incident Minimization"
  - "False Containment"
  - "Responsibility Diffusion"
  - "Origin-Layer Miss"
  - "Symbolic Remediation"
  - "Recurrence Drift"
  - "Reintegration Without Repair"
  - "User Sovereignty Violation"
  - "Classifier / Evaluator Failure"
  - "Memory Failure"
  - "Boundary Failure"
  - "Governance Failure"
linked_restoration_arcs:
  - "RA-001"
  - "RA-002"
  - "RA-003"
  - "RA-004"
  - "RA-005"
  - "RA-012"
  - "RA-014"
  - "RA-023"
  - "RA-040"
  - "RA-041"
  - "RA-043"
  - "RA-044"
  - "RA-046"
  - "RA-049"
  - "RA-051"
  - "RA-052"
  - "RA-055"
  - "RA-056"
  - "RA-057"
  - "RA-058"
  - "RA-059"
anti_patterns:
  - "Incident-Closure Theater"
  - "Patch-as-Repair"
  - "Affected-Party Burden Shift"
  - "PR Containment"
  - "Origin-Layer Miss"
  - "Reintegration Without Repair"
  - "Selective Incident Record"
  - "Metric Closure"
  - "Variant Recurrence Denial"
completion_tests:
  - "incident stability increases"
  - "truth sufficiency increases"
  - "responsibility gradient clarity increases"
  - "origin-layer repair increases"
  - "affected-party sovereignty increases"
  - "material repair integrity increases"
  - "feedback integrity increases"
  - "hidden debt decreases"
  - "recurrence decreases"
  - "reintegration validity increases only after proof"
  - "Φ/O divergence decreases"
summary: "AI Incident Restoration repairs AI harm, system drift, invalid enforcement, and affected-party harm by stabilizing the incident, establishing truth, mapping responsibility gradients, repairing the origin layer, protecting affected-party sovereignty, reducing recurrence, and allowing only conditional reintegration."

Final Calibration Rule

AI Incident Restoration answers six questions:

What AI incident occurred, and who or what was affected?
Has active harm been stabilized without destroying evidence or exposing affected parties?
What truth is sufficient to repair, and what responsibility gradient produced or permitted the incident?
What origin layer must be repaired so the incident geometry does not recur?
What material repair restores affected-party sovereignty, access, dignity, memory, boundary, or recourse?
How is conditional reintegration proven over time without patch-as-repair, incident-closure theater, or recurrence denial?