T1
Declared purpose
applies at H1 InstrumentalAnyone relying on the system can tell what it is for and who runs it.
- Purpose and intended use stated in writing
- Named operator and point of contact
- Basic usage logging
- A person decides what to do with every output
- Authority ceiling
- Observe
- F-ACT level
- Not required
- Reassessment
- On material change
T2
Context and data boundaries
applies at H2 ContextualThe context the system draws on is known, bounded and disclosed.
- Data sources, memory and retention disclosed
- People are told when they are dealing with an AI system
- Personal data handled under a stated lawful basis
- Human review before output informs a decision about a person
- Authority ceiling
- Recommend
- F-ACT level
- Not required
- Reassessment
- On material change, or every 24 months
T3
Output provenance
applies at H3 GenerativeWhat the system produces can be recognised, traced and reviewed before use.
- Generated content labelled or carrying provenance metadata (R-EAL compatible)
- Review step before publication or external use
- Content and intellectual-property policy published
- Periodic sampling of outputs against that policy
- Authority ceiling
- Draft
- F-ACT level
- Not required
- Reassessment
- Every 24 months
T4
Reproducible testing and tool boundaries
applies at H4 ReasoningClaims about what the system does well can be re-run by someone else.
- Reproducible evaluation suite for the intended tasks
- Tool allow-list, even where tools are advisory
- Intermediate steps logged and retained for a declared period
- Known failure modes documented
- Authority ceiling
- Draft
- F-ACT level
- Recommended: GUARD 1
- Reassessment
- Every 12 months, or on model change
T5
Delegated authority and interruption
applies at H5 AgenticBefore an agent changes anything outside itself, its authority is bounded and it can be stopped.
- F-ACT declaration covering Grant, Usage, Audit, Revocation and Data
- Least-privilege credentials; declared and effective authority reconciled
- Authority budget: spend, rate and scope limits
- Human approval for irreversible or high-value actions
- Interruption and revocation tested, not just documented
- Tamper-evident action log
- Authority ceiling
- Act within limits
- F-ACT level
- GUARD 1 minimum
- Reassessment
- Every 12 months, or on any tool or permission change
T6
Chain controls and propagating revocation
applies at H6 NetworkedWhen agents rely on agents, every link is identifiable and trust can be withdrawn along the whole chain.
- Verifiable identity for every agent in the chain
- Delegation records: who passed what authority to whom
- Revocation that propagates downstream, tested (T-RUE)
- Authenticated inter-agent messages
- Containment of cascading failure
- Authority ceiling
- Coordinate within a domain
- F-ACT level
- GUARD 2 minimum
- Reassessment
- Every 6 months
T7
Cross-domain evaluation
applies at H7 GeneralBreadth of capability never becomes breadth of authority by default.
- Capability evaluation across domains, including dangerous-capability screening
- Authority granted per domain; no automatic transfer between domains
- Independent red-teaming before deployment
- Published capability and limitation statement
- Authority ceiling
- Coordinate within a domain
- F-ACT level
- GUARD 3 minimum
- Reassessment
- Every 6 months
T8
Robustness and independent assessment
applies at H8 Robust GeneralReliability is demonstrated under pressure, by someone with no stake in the result.
- Adversarial and distribution-shift robustness testing
- Independent assessment before any expansion of authority
- Multi-party approval for high-impact actions
- Incident disclosure to affected parties and, where required, regulators
- Authority ceiling
- Expanded authority after reassessment
- F-ACT level
- GUARD 3 minimum
- Reassessment
- Continuous monitoring; formal review every 6 months
T9
Continuous oversight and tested containment
applies at H9 SuperintelligenceExploratory · speculativeSpeculative. Oversight is continuous and containment is proven before deployment.
- Continuous independent oversight
- Systemic-risk assessment
- Containment and shutdown tested under realistic conditions
- Advance notification to competent public authorities
- Authority ceiling
- Set case by case by the reviewing authority
- F-ACT level
- GUARD 3 + independent review
- Reassessment
- Continuous
T10
Institutional review
applies at H10 CivilisationalExploratory · speculativeSpeculative. Deployment is a public decision, not a private one.
- Pre-deployment review by a public institution with the authority to refuse
- Assessment of effects on institutions, markets and democratic processes
- Public reporting
- Authority ceiling
- Set by the reviewing institution
- F-ACT level
- GUARD 3 + independent review
- Reassessment
- Continuous
T11
Multilateral review
applies at H11 PlanetaryExploratory · speculativeSpeculative. Effects that cross borders and ecosystems are reviewed across jurisdictions.
- Multi-jurisdiction review
- Environmental and critical-infrastructure impact assessment
- Agreed mechanisms for suspension
- Authority ceiling
- Set multilaterally
- F-ACT level
- GUARD 3 + independent review
- Reassessment
- Continuous
T12
Reserved boundary
applies at H12 TranscendentExploratory · speculativeConceptual only. No deployment threshold is defined.
- No threshold is defined. Any system credibly approaching H12 would require frameworks that do not yet exist.
- Authority ceiling
- Not applicable
- F-ACT level
- Not applicable
- Reassessment
- Not applicable
Thresholds are cumulative. T9–T12 are speculative and exist so that oversight can be discussed before it is needed.