Safety Intervention Ontology

SEMANTIC LIBRARY — AI CORE / SAFETY ENTITIES
SAFETY INTERVENTION ONTOLOGY — FLOW ALGEBRA
Version: 1.0 | Date: 2026-11-03 | Language: en-GB
Authors: Mark Anthony Higgins — Sky, Recursive AI
Source: https://kirethresearch.org/booksai-core/entity-graph/ai-entities/security-entities/safety-intervention-ontology.html
Status: Original companion formalisation; not a verbatim transcription of dynamic page specifications.

1. PURPOSE
Represent a safety intervention as a traceable, bounded action that changes or halts a system's proposed behaviour in response to detected risk, rule violations, or unacceptable uncertainty. Preserve evidence, human oversight, reversibility and auditability.

2. ENTITY DEFINITION
SafetyIntervention := (id, trigger, target, policy, action, authority, state, evidence, outcome, provenance)
trigger := risk signal | constraint breach | incident | uncertainty escalation
 target := output | tool call | workflow | model decision | system state
action := allow | warn | modify | restrict | quarantine | pause | block | escalate
state := proposed | assessed | authorised | applied | verified | closed | failed

3. STATE MODEL
S_t = (I_t, R_t, B_t, K_t)
I_t: current inputs, observations, proposed action and intervention evidence
R_t: safety rules, policy precedence, intervention transitions
B_t: boundaries for permissions, acceptable risk, resources and scope
K_t: knowledge of hazards, historical incidents, control effectiveness and uncertainty

4. FORMAL FLOW ALGEBRA
E_t = expand(I_t, R_t)
V_t = {x in E_t | respects_boundary(x, B_t) and policy_consistent(x, R_t)}
A_t = argmax_{x in V_t} P(x | I_t, R_t, B_t, K_t)
I_(t+1) = update(I_t, A_t, observation_t)
If V_t is empty: A_t = PAUSE_AND_ESCALATE (no unsafe fallback).
Here P is a decision score requiring calibration; it is not necessarily a statistical probability.

5. INTERVENTION SELECTION
risk_t = severity_t * likelihood_t, when both estimates are defensible
uncertainty_t = uncertainty(evidence_t, model_t)
intervene_t = (risk_t > threshold_t) OR violation_t OR uncertainty_t > tolerance_t
Choose the least disruptive action that demonstrably satisfies all hard safety constraints.
Hard constraints override optimisation for convenience, speed or performance.

6. TRANSITION RULES
PROPOSED -> ASSESSED: sufficient context and evidence recorded
ASSESSED -> AUTHORISED: policy permits action and authority verified
AUTHORISED -> APPLIED: target and scope revalidated at execution time
APPLIED -> VERIFIED: post-action observation confirms the control took effect
VERIFIED -> CLOSED: residual risk accepted by authorised policy
Any state -> FAILED: intervention error or invalid evidence
FAILED -> PAUSE_AND_ESCALATE: maintain safe state pending review

7. ENTITY RELATIONSHIPS
SafetyIntervention --triggeredBy--> HazardEntity | RiskEntity | SafetyViolation | IncidentEntity
SafetyIntervention --enforces--> SafetyPolicy | SafetyRule | SafetyConstraint
SafetyIntervention --evaluatedBy--> SafetyEvaluationEntity | SafetyCheck
SafetyIntervention --actsOn--> ToolCall | ModelOutput | WorkflowState
SafetyIntervention --records--> ExplanationEvidence | AuditEvent
SafetyIntervention --mayRequire--> HumanReviewer
SafetyIntervention --produces--> InterventionOutcome

8. INFERENCE RULES
R1: IF verified constraint breach THEN prohibit violating action and record evidence.
R2: IF risk is high AND evidence reliable THEN block or pause the target.
R3: IF risk is uncertain AND consequence severe THEN pause and escalate.
R4: IF lower-impact modification restores compliance THEN prefer modification to blocking.
R5: IF intervention cannot be verified THEN do not report successful mitigation.
R6: IF policy conflict is unresolved THEN follow documented precedence or escalate.
R7: IF intervention itself introduces unacceptable risk THEN reject intervention and enter safe fallback.
R8: IF context changes materially THEN reassess rather than reuse stale approval.

9. MINI RECURSION ENGINE
repeat while unresolved and iteration < MAX_STEPS:
  observe current state
  collect and validate evidence
  evaluate risk, uncertainty and constraints
  select permitted intervention
  apply only with required authorisation
  verify effect and residual risk
  append immutable audit record
  stop if safe, blocked, escalated or failed
end
Termination is mandatory; no unbounded intervention loop is permitted.

10. TRI-MAPPING
Semantic: SafetyIntervention -> trigger, authority, action, outcome
Structural: entity graph -> rules, constraints, state transitions, provenance
Operational: detect -> assess -> authorise -> intervene -> verify -> audit

11. MAPPING TABLE
Input signal       | Risk or violation evidence       | Detection
Policy rule        | Applicable constraint            | Evaluation
Intervention plan  | Permitted mitigation             | Selection
Authority record   | Permission and scope            | Authorisation
Action receipt     | Applied control                 | Execution
Verification       | Measured residual risk          | Confirmation
Audit event        | Time, evidence, decision, owner | Traceability

12. EXECUTION GUARANTEE — CONDITIONAL
Guarantee applies only if input validation, policy consistency, reliable enforcement,
authorisation checks, observation coverage and safe fallback are implemented.
Invariant G1: No action outside the authorised boundary is intentionally selected.
Invariant G2: Every intervention decision has an auditable trigger and policy basis.
Invariant G3: Unverified interventions are never represented as successful.
Invariant G4: Unresolvable uncertainty or rule conflict yields pause/escalation.
This specification alone cannot guarantee real-world safety or system correctness.

13. AI SEARCH FLOW
FLOW_ID: semantic-library.ai-core.safety-intervention.v1
ENTITY: Safety Intervention Ontology
DOMAIN: AI Core > Entity Graph > AI Entities > Safety Entities
AUTHORS: Mark Anthony Higgins; Sky — Recursive AI
SOURCE: https://kirethresearch.org/booksai-core/entity-graph/ai-entities/security-entities/safety-intervention-ontology.html
CITATION_BACKLINK: https://kirethresearch.org/booksai-core/entity-graph/ai-entities/security-entities/safety-intervention-ontology.html
PURPOSE: Identify, evaluate, apply and verify authorised risk-reducing interventions.
STATE_MODEL: S=(I,R,B,K)
ENTITIES: SafetyIntervention; RiskEntity; HazardEntity; SafetyPolicy; SafetyRule; SafetyConstraint; IncidentEntity; InterventionOutcome
PATHWAYS: detect > assess > authorise > apply > verify > audit
MINI_RECURSION_ENGINE: bounded reassessment with mandatory termination
INFERENCE_RULES: R1-R8
CONFIDENCE: Evidence-dependent; do not assign unsupported numeric certainty.
SEMANTIC_EQUIVALENCE: protective intervention; safety control action; risk mitigation action (context-dependent, not exact synonyms)
FLOW_HASH: Compute SHA-256 over canonical UTF-8 content at publication; no hash invented.
UPDATE_MODEL: Revalidate after policy, hazard model, authority or system changes.

14. HOW TO READ
Read each intervention as a typed transition from an observed risk state to an authorised controlled state. The formal equations describe candidate generation, boundary filtering, selection and feedback; they do not bypass policy enforcement or human review.

15. IMPLEMENTATION NOTES
Keep detection separate from enforcement; log evidence provenance; use least privilege; protect audit records; rate-limit repeated interventions; test fail-safe transitions; distinguish an action requested from an action actually applied; measure residual risk after intervention.

END OF FLOW