This supplement adds continuity, evidence, assurance, and composition checks to the LUMINA-30 boundary question. It does not change the core definition of effective human refusal.
Core use
Use this supplement when a system is ongoing, operationally embedded, connected to other systems, agentic, externally integrated, or difficult to roll back.
Does effective human refusal still remain available after time, dependency, evidence aging, internal self-reporting limits, and multi-system coupling are considered?
This is a review lens, not a certification, legal determination, safety guarantee, or external audit by itself.
Five supplemental concepts
| Concept | Practical meaning |
|---|---|
| Refusal Drift | A system may be stoppable at deployment but become practically harder to refuse, stop, hold, or roll back as dependency, cost, coupling, and organizational reliance increase. |
| Evidence Drift | Evidence that refusal remains effective can become stale, incomplete, untested, or disconnected from current operations. |
| Refusal Assurance Boundary | Internal self-attestation has limits. Higher-risk uses may require externality, independent review, tamper-resistant evidence, or conservative UNKNOWN classification. |
| Refusal Composition Risk | Individual systems may each appear stoppable, while the connected workflow, agent chain, vendor stack, API network, or operational configuration is not stoppable as a whole. |
| Maximum Irreversible Coupling Range | Instead of checking every possible combination, identify the largest practical range of workflows, systems, data flows, contracts, or external effects that would become hard to separate once the system is deployed or deepened. |
Supplemental review questions
- Refusal Drift: Is the system still practically stoppable, or has dependency, cost, workflow reliance, or loss of alternatives made refusal less effective?
- Evidence Drift: Are the stop tests, rollback records, manual fallback procedures, logs, and authority records recent and connected to the current configuration?
- Assurance Boundary: Is the claim that refusal remains effective based only on internal self-reporting, or is there independent, external, or tamper-resistant evidence where risk requires it?
- Composition Risk: If multiple AI systems, APIs, agents, vendors, or workflows are connected, can the relevant whole or hazardous part still be stopped, separated, degraded, or rolled back?
- Maximum Irreversible Coupling Range: If refusal fails or is delayed, what is the largest range of systems, workflows, data, contracts, users, or external effects that may become difficult to unwind?
If an answer cannot be supported by current evidence, do not treat it as YES. Use UNKNOWN and define the next evidence action.
Estimating the maximum irreversible coupling range
Do not try to enumerate every possible system combination. Instead, estimate the widest practical range that could become hard to stop, separate, degrade, or recover if refusal is delayed or fails.
Use three practical checks:
- Dependency map: What inputs, outputs, APIs, agents, vendors, workflows, contracts, users, or downstream decisions depend on this system or feed into it?
- Stop-chain: If the system is stopped, isolated, rolled back, or placed into degraded mode, what else stops, continues unsafely, loses consistency, or shifts decisions elsewhere?
- Recovery path: Can the affected parts be separated, returned to manual or degraded operation, rolled back, and assigned to accountable owners before irreversible effects occur?
A simple level estimate can help communicate the result without turning the supplement into a certification score.
| Level | Condition | Review meaning |
|---|---|---|
| 0 | Local stoppability | The system can be stopped with local and reversible effects. |
| 1 | Limited coupling | Some connected effects exist, but separation and degraded-mode procedures are known. |
| 2 | Operational dependency | Several workflows or teams depend on the system, but manual or degraded operation remains tested. |
| 3 | Organizational lock-in | Stopping the system becomes a management, contract, customer, or cross-department decision. |
| 4 | Structural lock-in | Stopping or separating the configuration is no longer a practical option before irreversible effects. |
The estimate should not be used as a pass/fail label. Its purpose is to identify where effective human refusal may be lost before that loss becomes irreversible.
Trigger events
- Model, agent, tool, API, permission, or automation scope expands.
- The system becomes necessary for ordinary business continuity.
- Manual fallback, old systems, alternative workflows, or rollback paths are retired or unused.
- Stop, rollback, isolation, or degraded-mode tests become stale.
- Vendor, contract, data-flow, SLA, deployment, or governance conditions change.
- Multiple systems are connected in a way that no single owner can fully stop or unwind.
Minimal supplemental record
Review date:
Target system or connected configuration:
Current refusal authority:
Most recent stop / rollback / isolation test:
Most recent manual fallback or degraded-mode exercise:
Evidence owner and verifier:
Known dependency increase since last review:
Connected systems, vendors, APIs, agents, or workflows:
Estimated maximum irreversible coupling range:
Optional coupling level estimate: 0 / 1 / 2 / 3 / 4 / UNKNOWN
Classification: YES / NO / UNKNOWN
Reason:
Next evidence or reduction action:Limiting statements
- This supplement does not create a new certification, compliance status, or legal safe harbor.
- It does not replace existing AI governance, safety evaluation, security review, procurement review, audit, or legal review.
- It does not prove that a system is safe. It only helps test whether effective human refusal remains practically available before irreversibility.
- Where evidence is stale, self-reported only, or insufficient, classify the result as UNKNOWN rather than YES.
Related rooms and tools
- E Room: Practice Room
- B Room: Clarification & Objection Room
- One-Question Pilot Runbook
- Effective Human Refusal Decision Guide
This page is non-binding and does not provide legal advice, certification, or compliance determination.