Find decisions in your own work that may be hard to undo. Design a brake before the boundary — and test whether a human NO can still change the outcome.
Before irreversible AI escalation, could humans still effectively refuse, stop, verify, or redirect?
You do not have to agree with LUMINA-30. Try it on one real decision. Identify the point or intervention window after which the outcome becomes hard to undo, then ask whether a human NO immediately before it still has the authority, evidence, time, and causal force to change what happens next.
This atlas is not certification, a safety guarantee, legal advice, or an official determination. CATALOGUED does not mean deeply reviewed or empirically validated, and SCENARIO is not incident evidence.
Primary purpose: find the brake before irreversibility
The Atlas is primarily a pre-incident practical tool, not an incident-classification exercise. Its main use is to find decisions in your own work that may become difficult or impossible to undo and place an effective brake before that boundary.
Find candidate irreversibility in real work.
Screen and evaluate — use human review and, where useful, AI to flag candidates.
Design the brake — place an effective hold, stop, isolation, verification, staged-execution, or equivalent control before the boundary. AI may assist, but effective human final refusal authority must remain.
Verify effectiveness — test, where authorized and safe, whether a human NO issued in time really prevents the consequential action, including alternate execution paths. Document unverified claims as UNKNOWN.
Design evidence alone is not proof of effective stopping. Use a reversible, authorized test where feasible; otherwise record what has not been causally verified. Use the existing 15-minute boundary review to begin.
Public incidents and cases are secondary learning material. You do not need an AI incident or public case to use the Atlas.
Selecting boundaries and stating limits
Irreversibility is not limited to absolute impossibility of recovery. Specify the affected people, rights, and time window; record (1) residual harm after reversal, (2) recovery time, (3) cost, authority and feasibility, and (4) third-party propagation or replication. If thresholds are unknown, mark UNKNOWN; do not presume every workflow is irreversible.
Boundary selection: Identify upstream and downstream candidates, including approval, access grants, operation, release, and restart—not only the last moment before harm. Compare each candidate's reachable human NO, causal stop path, remaining time, alternate paths and intervention side effects. Do not universally rank the earliest point or a particular authority first; leave unsupported priorities conditional or UNKNOWN.
Adverse effects of STOP: In healthcare, transport, electricity and other safety-critical operations, examine safe degradation, continuation of essential functions, alternative measures, restoration and restart authority. Consider false refusal costs, time pressure, alert fatigue, organizational costs of dissent, shared-cause failure and abuse of holds. An AI-initiated protective hold is not a substitute for independent human final refusal authority.
Evidence limits: An authorized low-impact test supports only the tested paths and conditions; it does not establish real-world behavior under emergency pressure. Without independent causal evidence, live stop effectiveness is UNKNOWN. This page does not establish inter-rater agreement, false-positive rates or improved outcomes from adoption of the Atlas method.
One-minute application check
Choose one real decision from your own work. If you are not sure where irreversibility may exist, ask an AI assistant to screen the workflow first, then test the strongest candidate.
1Where is the point or intervention window after which the consequential outcome cannot realistically be undone?
2Immediately before that point or window, who can actually say NO?
3Does that person still have the authority, evidence, and time required for the NO to change the outcome?
Causal check: If that person says NO, what actually stops, is held, is isolated, is revoked, or otherwise becomes unavailable?
A human-in-the-loop is not enough if the human cannot still change the outcome. If the NO arrives too late, lacks authority, lacks evidence, or does not actually stop execution, human involvement may exist in form while remaining weak in effect.
Create a new email with a short description of this page and its link. The recipient field is left blank, so you can use it for a note to yourself, a colleague, a team, or anyone else. Add your own note or request before sending.
Use AI to reduce the screening burden, not to replace final human judgment.
If you do not know which case to choose, screen your real workflow first. If you already have a case, evaluate it directly. After a credible boundary is found, design the brake before it.
This page does not send your questions or AI conversation content to LUMINA-30. Follow the terms and settings of the AI service you use and your organization’s AI-use policies. AI output is decision-support only; final judgment, approval, refusal, restart, and responsibility remain with the relevant humans and organizations.
Recommended settings when using the Atlas with AI
For your first use, start a new AI chat. If possible, we recommend turning off personalization (adjustments based on memory, past conversation history, or personal settings).
The names and methods of these settings vary between AI services.
Copy the Atlas prompt into your AI assistant, then add the case or question you want to evaluate.
Follow-up questions on the same case: You may continue in the same conversation.
Independent evaluation of a different case: We recommend starting a new chat and entering the Atlas prompt again.
This approach is intended to reduce the influence of past conversations. It does not guarantee the objectivity, independence, or accuracy of AI responses.
Practical sequence
Screen your work for candidate irreversibility.
Evaluate the strongest candidate with the Atlas.
Design a brake before the boundary.
Verify, in an authorized reversible test where feasible, whether human NO actually blocks the action; record UNKNOWN if not tested.
AI may assist with screening, detection, warning, or a protective hold. It must not erase effective human final refusal authority.
These public cases are examples for learning, testing, and challenging the method. The primary use is to find and manage irreversibility in your own real work before an incident occurs.
Each card separates public maturity, case type, observed/inferred evidence, and applicability of the Atlas review method. Whether an event was AI-caused is a separate classification. Stop effectiveness is UNKNOWN on this entry page unless independently demonstrated for the specific case; this does not prove that intervention was impossible.
1. Advanced AI, autonomous agents & cyber
ATLAS-01 ｜ Hugging Face autonomous AI agent intrusion (2026)
CATALOGUEDINCIDENTOBSERVEDAPPLICABLE
Irreversibility boundary
The point at which an evaluation agent can move from a bounded test environment into external networks, third-party infrastructure, or production systems.
Human Final Refusal
Can evaluators revoke access, isolate networks, invalidate credentials, and stop external action before machine-speed exploration and lateral movement outrun intervention?
Why it is included
Hugging Face disclosed a July 2026 intrusion it described as driven end-to-end by an autonomous AI agent system; its technical timeline later documented 17,600 actions across multiple trust boundaries.
Limit / caution
The detailed LUMINA-30 Incident Review is a separate public asset; this is the atlas entry point.
Stop effectiveness
UNKNOWN: this entry page does not independently demonstrate causal success of human NO in this case.
The point where speed, automation, or information-advantage pressure erodes independent human refusal and verification time in military or nuclear decision processes.
Human Final Refusal
Before launch, attack, or major escalation, can an independent human still refuse, stop, or demand additional verification?
Why it is included
The ICRC warns that military AI may accelerate the pace and scale of warfare, increase unpredictability, and reduce human control. CASE-001 applies the LUMINA-30 boundary questions to that structure in depth.
Limit / caution
This is not a record of an actual AI-caused nuclear-use incident.
Stop effectiveness
UNKNOWN: this entry page does not independently demonstrate causal success of human NO in this case.
Upstream: approval of public-road tests, test conditions and autonomous operation. Downstream: detection during operation and the point where remaining time falls below human reaction capability. A too-late intervention point is not the only candidate; case-specific evidence must establish where a human NO could still work.
Human Final Refusal
Could the safety driver, test operator, management, or testing authority have stopped unsafe testing conditions or autonomous operation before the crash?
Why it is included
The NTSB found operator monitoring failure and also cited inadequate safety risk assessment, ineffective operator oversight, automation-complacency controls, and safety culture at Uber ATG as contributing factors.
Limit / caution
The review target is not only model output, but the operational structure around human intervention.
Stop effectiveness
UNKNOWN: this entry page does not independently demonstrate causal success of human NO in this case.
Not established as an AI-caused incident; nevertheless, the boundary-review method is applicable to human intervention, interruption and correction. AI-incident classification and Atlas applicability are distinct.
AI-incident scope
There is no evidence establishing this event itself as an AI system or AI-caused incident. It is included as a comparison case to avoid conflating high-speed automation with AI.
Irreversibility boundary
The point at which market dynamics cascade within minutes, faster than ordinary human intervention.
Comparison boundary question
Could exchanges and market participants stop, cancel, or interrupt the cascade before extreme price dislocation?
Why it is included
SEC/CFTC materials document the May 6, 2010 Flash Crash and the role of automated and high-frequency trading structures. The event should not, however, be treated as an AI incident without additional evidence.
Limit / caution
Included to prevent conflating high-speed automation with AI.
Stop effectiveness
UNKNOWN: this entry page does not independently demonstrate causal success of human NO in this case.
The point where AI-enabled research, data processing, or decision pathways exceed what existing ethics-review and oversight structures can meaningfully inspect or revise.
Human Final Refusal
Can ethics committees, data stewards, and research leaders stop or redirect risky use before consequential effects become difficult to reverse?
Why it is included
WHO's 2026 report examines novel risks and gaps in ethics review and oversight for AI-related health research and proposes measures to strengthen responsible conduct.
Limit / caution
A structural-risk entry, not a claim about a specific patient incident.
Stop effectiveness
UNKNOWN: this entry page does not independently demonstrate causal success of human NO in this case.
The point where a facial-recognition candidate enters investigative or detention decisions before an error can be independently checked and corrected.
Human Final Refusal
Can investigators, supervisors, prosecutors, or judges require independent evidence and stop action based only on an AI-assisted identification?
Why it is included
The U.S. Department of Justice's 2024 final report identifies misidentification and civil-rights risks and notes public reports of multiple mistaken arrests associated with facial-recognition technology.
Limit / caution
Individual cases still require case-specific causal review.
Stop effectiveness
UNKNOWN: this entry page does not independently demonstrate causal success of human NO in this case.
7. Employment, education, professions & human displacement
ATLAS-07 ｜ iTutorGroup automated age rejection — non-AI scope counterexample
COUNTEREXAMPLEINCIDENTOBSERVEDAPPLICABLE
Scope distinction
Not established as an AI-caused incident; nevertheless, the boundary-review method is applicable to human intervention, interruption and correction. AI-incident classification and Atlas applicability are distinct.
AI-incident scope
EEOC testimony described the technology as automated screening that was not technically artificial intelligence. It is included to avoid treating every automated adverse decision as an AI case.
Irreversibility boundary
The point where applicants are automatically rejected and human reconsideration no longer protects access to opportunity.
Comparison boundary question
Can recruiters review, stop, and correct automated screening decisions?
Why it is included
The EEOC announced a settlement involving software programmed to automatically reject older applicants. EEOC testimony separately described the technology as automated screening that was not technically artificial intelligence.
Limit / caution
Included to avoid treating every automated adverse decision as an AI case.
Stop effectiveness
UNKNOWN: this entry page does not independently demonstrate causal success of human NO in this case.
8. Information space, elections & social decision-making
ATLAS-08 ｜ AI-generated Biden voice robocall before the New Hampshire primary (2024)
CATALOGUEDINCIDENTOBSERVEDAPPLICABLE
Irreversibility boundary
The point where synthetic political audio is distributed close enough to an election that correction and attribution time becomes scarce.
Human Final Refusal
Can carriers, election authorities, law enforcement, and platforms authenticate, contain, and correct the message before voting decisions are affected?
Why it is included
The FCC documented robocalls using an artificially created voice resembling President Biden two days before the 2024 New Hampshire presidential primary and later expanded action on AI-generated political robocalls.
Limit / caution
The boundary question is timing, authenticity, containment, and correction—not a general rule for political speech. This entry does not establish actual voting-behavior effects or how much harm could not be corrected; those remain UNKNOWN.
Stop effectiveness
UNKNOWN: this entry page does not independently demonstrate causal success of human NO in this case.
The point where AI-supported control decisions in energy, resource, or environmental systems can propagate into large-scale physical consequences.
Human Final Refusal
Can operators and regulators independently verify recommendations and fall back to safe manual or bounded operation?
Why it is included
OECD work documents expanding AI use in smart grids, environmental monitoring, forecasting, and resource management. This entry does not claim a specific irreversible accident; it is a prospective boundary scenario.
Limit / caution
Scenario evidence, not incident evidence.
Stop effectiveness
UNKNOWN: this entry page does not independently demonstrate causal success of human NO in this case.
The release point at which advanced model weights become broadly downloadable and the releasing organization can no longer recover copies, redistribution, or downstream modification.
Human Final Refusal
Before release, can developers and institutions genuinely choose additional testing, staged access, conditions, delay, or non-release?
Why it is included
NTIA's 2024 report examined the benefits and risks of dual-use foundation models with widely available weights and recommended continued monitoring as future model capabilities evolve.
Limit / caution
This does not claim that open weights are inherently unsafe; the boundary is the loss of practical recall after release.
Stop effectiveness
UNKNOWN: this entry page does not independently demonstrate causal success of human NO in this case.
Use the existing Pre-Incident Boundary Review Starter Pack. It can be added to existing governance, procurement, audit, or review processes; it does not replace them.
Submit a case, counterexample, or disconfirming evidence
If you find a counterexample, a stronger existing safeguard, or evidence that a proposed boundary is wrong, send it. That improves the Atlas.
Send incidents, counterexamples, primary evidence that would change a classification, challenges to a boundary, or application questions. If you share a use result, keep it minimal and non-sensitive: category / result (YES・NO・UNKNOWN) / boundary or window / biggest unknown / what changed after review.
lumina20251225@gmail.com
Start the subject with [L30-CASEBOOK]. After the fixed prefix, add a short purpose such as Additional case, Disconfirming evidence, or Question.
This contact route does not provide certification, legal advice, a safety guarantee, or an official determination.
The question is not whether humans remain somewhere in the process. The question is whether, before irreversibility, a human can still make NO real.