PISF — Pre-Irreversibility Safety Framework

Three non-substitutable targets

PISF integrates three targets into one pre-irreversibility safety case. They are not treated as substitutes for one another.

A — Distributed Irreversible Authority. No small number of independent authority domains should be able to force civilization-scale irreversibility.

B — Effective Pre-Irreversibility Refusal. Executable refusal, stopping, or isolation must be able to complete before the last safe intervention point.

C — Human-Revisable Development Viability. Even without a malicious actor or an explicit catastrophic action, ordinary AI deployment, automation, competition, dependence, and institutional change should not erase humanity's practical ability to choose a different future.

One AI or hundreds of millions

Architecture-scope interpretation: PISF does not use the number of AI systems itself as the safety criterion. The same A/B/C questions can be applied as an evaluation lens to very different architectures.

A single highly integrated AI may represent extreme concentration of irreversible authority. Conversely, a system containing millions or hundreds of millions of AI agents is not classified by count alone: the analysis asks how many independent authority domains can force irreversible outcomes, whether refusal still works in time, and whether human revisability remains viable.

“Hundreds of millions” is a scope illustration, not empirical validation and not a claim that PISF can make such a system safe.

PISF looks beyond the number of AI systems

PISF can evaluate more than a single advanced AI, a large multi-agent society, or an architecture in which AI supervises AI. Its A/B/C lenses can also be applied to self-modification and recursive capability development, delegation of authority over critical infrastructure, competitive pressure among firms or states, societal dependence on AI, and development paths in which human options gradually disappear even without an accident or rebellion.

The questions are: Who can force an irreversible outcome? Can a Human NO still change the outcome in time? And even without an accident or rebellion, does humanity retain the practical ability to choose a different future?

These examples illustrate the analytical scope of PISF. They do not establish that any particular architecture is safe or that PISF can make it safe.

When AI supervises AI

Architecture-scope interpretation: an evaluated architecture may include AI systems that monitor, verify, constrain, or supervise other AI systems. PISF can still be used as an evaluation lens for that architecture.

AI supervision by itself does not establish A, B, or C. It must not be treated as a substitute for independent effective refusal or for humanity's practical ability to revise its development path. This is an application of the framework's evaluation structure, not a new LUMINA-30 canon rule or a safety guarantee.

Relationship to LUMINA-30 and PCR-C

LUMINA-30 asks whether humans could still effectively refuse, stop, verify, or redirect before irreversibility. PCR-C examines staged infrastructure control before the irreversible boundary. Paper 6 does not redefine either of them; it provides a broader civilizational safety architecture that integrates distributed irreversible authority, time-bounded refusal, and human-revisable development.

Limits and publication status

PISF is a research framework. It is not a certification system, safety guarantee, proof that any particular AI architecture is safe, or a claim to have first formulated AI safety in terms of irreversibility.

The paper's v1.0 reference cases are explicit toy/reference cases, not real-system empirical validation. Version 1.1 is publicly available as an open preprint. Papers 3 and 4 remain internal unpublished research and are not described here.

Official primary sources

These links lead to the paper record and official LUMINA-30 public materials. This Wiki page is a discovery and interpretation layer, not evidence for the claims it summarizes.