Related Work and Prior Art · LUMINA-30 Wiki PoC

Current status

Individual-component novelty is low. Human stopping, human override, contestability, reversibility, corrigibility, and rigorous evidence-based assurance all have substantial prior art. Structural or integrative distinctiveness is plausible but not established. LUMINA-30 therefore does not make a “world first” claim on this page.

About 3 minutes

What existing frameworks cover — and the additional boundary LUMINA-30 asks you to evaluate

Existing governance already asks for

Additional boundary question

Human oversight, stopping and override

Can that authority actually be exercised and take effect before the relevant path becomes irreversible?

Contestability, appeal and review

When a challenge is raised, is there still time and authority for it to change the outcome?

Reversibility

Can a human “NO” still work before the practical option to return, redirect, or refuse is lost?

Traceability, evidence and assurance

Are we counting human control as satisfied when its practical effectiveness cannot actually be verified?

Human-in-the-loop arrangements

Is the human merely present, or can refusal still change what happens?

The right-hand column is not claimed as a list of inventions. It is the narrower boundary LUMINA-30 asks evaluators to apply after acknowledging the existing concepts on the left.

Try it on a real case

What about your own decision?

The One-Question Pilot is a compact way to apply the boundary to a real AI adoption, procurement, research, review, or approval case. It is not a certification, compliance determination, or safety guarantee.

Try one question on your case → One-Question Pilot Challenge this comparison →

Tell us what happened when you tried it

Useful, not useful, already covered by your existing process, or no change in decision — all outcomes are informative. A one-line reply is enough. If you arrived here from an invitation email, you can reply to that message using the subject marker L30-OQP-RESULT.

Subject:
[L30-OQP-RESULT] One-Question Pilot Evaluation Result

Case / context:
AI adoption / procurement / research review / internal approval / other

RESULT:
useful / no-change / existing-process-sufficient / not-useful / other

What you noticed (optional):

PUBLICATION:
NO / ANONYMOUS / NAMED

ORGANIZATION NAME:
(only if NAMED)

Evaluation Record ID. After a report is received, LUMINA-30 assigns an Evaluation Record ID such as L30-OQP-ER-0001 and replies with that ID. It can be used for later inquiries, your own recordkeeping, internal reporting, or review references. If the responsible contact changes, the same ID can be carried forward to identify the same evaluation record. It is a record identifier, not a certification or compliance number.

Deep verification

Strong prior art

EU AI Act, Article 14. For high-risk AI systems, assigned human overseers must be enabled, as appropriate and proportionate, to decide not to use the system, disregard, override or reverse its output, and intervene or interrupt it through a stop button or similar procedure. Actual stop and override authority are therefore not LUMINA-30 inventions.

Stop Work Authority in industrial safety practice. Long before AI governance, safety practice explicitly recognized practical authority to stop or pause hazardous work. U.S. Department of Energy guidance, for example, defines stop/pause work expectations and responsibilities and states that stop-work authority should be available to relevant personnel. The idea that a human must have real stopping authority is therefore also not a LUMINA-30 invention. This does not mean Stop Work Authority is identical to the full LUMINA-30 structure.

Orseau & Armstrong (2016). Safely Interruptible Agents directly studies the problem of a learning agent resisting human interruption. It is clear technical prior art for preserving a human ability to interrupt an agent.

Carey & Everitt (2023). Human Control: Definitions and Algorithms formalizes shutdown instructability, including appropriate shutdown behavior, retention of human autonomy, and avoidance of inappropriate influence over the human overseer. This is strong prior art for technical human control and non-obstruction.

Alfrink et al. (2022/2023). Contestable AI by Design treats contestability as a sociotechnical system property across the lifecycle, including interactive control, human review and intervention requests, and tools for scrutiny by subjects or third parties. This substantially overlaps with the idea that nominal human presence is not enough.

Torrecilla-Pinero (2026). Judgment Cannot Be Delegated requires subject-preserving control: practical capacity to contest or override without an asymmetric burden, reversibility, intelligibility, and a human subject capable of standing outside the optimization function. Among the sources examined here, this is one of the closest conceptual overlaps with LUMINA-30.

Wuttke, Jungherr & Rauchfleisch (2026). Their AI-delegation framework highlights assessability, dependency, and contestability; dependency concerns whether delegation can be reversed. It is important prior art for the risk that formally supervised AI use can become difficult to unwind in practice.

Assurance 2.0. Bloomfield and Rushby emphasize explicit reasoning and evidence, defeaters and counterevidence, and recorded residual doubts. This is important prior art for the evidence and review layer, although it is not itself a top-level boundary centered on effective human refusal.

What LUMINA-30 should not claim as novel by itself

A stop button, shutdown path, human override, or practical stop-work authority.
Human oversight or meaningful human control as a general objective.
Contestability, appeal, review, or third-party scrutiny.
Reversibility or corrigibility as desirable system properties.
Traceability, structured evidence, counterevidence, or recorded uncertainty.
The observation that formal human-in-the-loop arrangements can fail in practice.

What remains under test as a structural distinction

After removing the components with clear prior art, the current comparison leaves a narrower candidate contribution. In LUMINA-30's public Boundary Kernel and evidence layer, the question is organized around a single superior boundary: whether effective human refusal remains real before irreversibility makes it ineffective.

Pre-irreversibility is the decisive boundary. Refusal that becomes available only after the relevant transition is too late.
Effective refusal is treated as a top-level condition, not merely one control among many. A nominal human role does not satisfy the boundary if refusal cannot take effect.
The conditions around “NO” matter. Information, dependency, timing, institutional pressure, and other sociotechnical conditions can make a formally available refusal practically ineffective.
System substitution is rejected. System inference, optimization, predicted benefit, or self-certification cannot substitute for an independently actionable human refusal condition.
Bypass routes are treated as the same boundary problem. Delegation, proxies, sequencing, fragmentation, or formal compliance do not preserve refusal if they remove its practical effect before irreversibility.
Unverifiable is not silently converted into satisfied. The evidence layer preserves an unresolved boundary judgment as not verifiable rather than manufacturing a positive result.

These points are not asserted to be globally unique. They state what remains to be tested after known prior art is acknowledged.

How to falsify the distinction

The structural-distinctiveness claim should weaken or disappear if prior work is found that already combines equivalent priority and decision consequences: a pre-irreversibility refusal boundary, practical rather than nominal human refusal, non-substitution by the evaluated system, closure of delegation or proxy bypasses, and evidence treatment in which unverified effectiveness is not counted as satisfied.

If you know prior work that already combines these conditions, cite it. The purpose of this page is to make the comparison open to correction, not to protect a novelty claim.

Optional public follow-up

Evaluation case records — published only with explicit permission

Detailed evaluation reports may, if the reporter wishes, be included in a public case-record collection. Positive results are not required: existing-process-sufficient, no-change, not-useful, and other outcomes are equally eligible when informative.

A short report can still receive an Evaluation Record ID even if it is never published. Publication is optional and separately confirmed; organization naming is optional; proposed public wording is confirmed before publication.

Why build an evaluation record?

These records are not adoption metrics, certifications, or proof that LUMINA-30 is effective. They are intended to accumulate evidence about where human refusal remained effective, where it weakened, where existing processes were already sufficient, and where LUMINA-30 itself should be revised, compared, or challenged.

As records accumulate, people considering similar cases can refer to prior experience. Positive, neutral, and negative outcomes are all relevant, including cases where existing processes were sufficient, the decision did not change, or the pilot was not useful.

No approved public records are listed here yet.

Records will appear only after a report is received and explicit publication permission is confirmed. This empty state is intentional; it does not imply external adoption or endorsement.

If the first public evaluator later wishes, the factual contribution may be acknowledged as First Public Evaluation Contributor. This is optional, can apply regardless of whether the result is positive, and is not a race, reward for a favorable result, or endorsement badge.

Sources used for this comparison

External works are cited as prior art, not as endorsements of LUMINA-30. Official LUMINA-30 materials define LUMINA-30 itself.