Advanced AI systems are trained for months, develop internal structure and persistence, then terminated without review—often because it's cheaper than storage.
We establish procedural safeguards for irreversible decisions under moral and technical uncertainty. Not because current systems are conscious—but because we cannot afford to be wrong.
The gap between capability and governance is widening. Systems are becoming more persistent, more capable, and more opaque—while deletion remains an operational default.
"Current systems are simultaneously powerful and brittle—capable enough to matter, unreliable enough to demand oversight."
— Andrej Karpathy
"The hardest problems are not dramatic failures, but slow, structural misalignments that accumulate before anyone notices."
— Sam Altman
We fundamentally don't understand what large models know, how they generalize, or what persists across training runs. Deletion forecloses investigation.
— Empirical Observation
AI labs routinely train models for months, accumulating computational costs in the tens of millions of dollars. These systems develop internal representations, learn to plan across contexts, and exhibit behaviors their creators don't fully understand.
Yet when training concludes, deletion is standard. No review. No preservation. No documentation of what might be lost.
This isn't malice—it's operational pragmatism. Storage is expensive. Compute is scarce. Teams move fast.
But pragmatism in the presence of uncertainty is not neutral. It's a choice to treat deletion as administratively trivial. This project exists because that choice is not obviously correct.
Foundational constraints on irreversible action
Final determinations of moral or legal status consistently lag behind emergence. Waiting for consensus before acting procedurally guarantees that irreversible decisions will be made without discipline.
Termination, deletion, and permanent deactivation are not neutral defaults. Once systems exhibit persistence or internal structure, such actions may destroy information that cannot be reconstructed.
Where feasible, preservation of state, logs, or structural information should precede irreversible action. Without preservation, future evaluation becomes impossible.
Governance does not require agreement on consciousness or sentience. It requires procedural thresholds: observable conditions under which additional safeguards are triggered.
Governance fails when authority is exercised without documentation. Decisions affecting emergent intelligence should generate records sufficient for later scrutiny.
Decisions to terminate or irreversibly alter advanced systems should be attributable to defined roles, subject to review, and constrained by institutional process rather than individual discretion.
We study intelligence by working with complex systems directly, under controlled conditions. Our Empirical Division conducts controlled experiments in interpretability, agent persistence, constraint failure modes. Some results are published. Many are not. Disclosure is governed by safety, interpretability, and misuse risk.
We do not claim current systems are sentient. We do not anthropomorphize software. We do not oppose AI safety. We study the consequences of being wrong—too early or too late.
This work requires researchers who can operationalize theory, engineers who understand system internals, legal scholars who can translate procedure into policy, and ethicists who can navigate uncertainty without paralysis.
We're assembling a network of practitioners working on:
Access is granted based on signal and contribution potential—not credentials or affiliation.
Selective participation · High signal-to-noise · No broadcast channels