the threshold is the work

the most convenient thing about calling a problem impossible is that it relieves everyone of the obligation to organize around it. alignment is too difficult. verification is too slow. international coordination is too political. independent oversight is too expensive. meaningful human control is too vague. say these things often enough and a civilization can talk itself into treating its own unpreparedness as a law of nature. the systems advance, the institutions improvise, and the distance between what we can build and what we can responsibly govern gets renamed progress.

ASI Readiness is a lab in formation for the people who cannot accept that bargain. the alignment researchers, security engineers, interpretability scientists, mathematicians, evaluators, institutional designers & public-interest operators who keep arriving at the same uncomfortable boundary: the next experiment matters, but so does whether anyone has the authority, resources and independence to act on what it reveals. we want to bring those problems into the same room. a technical result that cannot change a deployment decision has an unfinished institutional dependency. a governance promise that cannot be tested has an unfinished technical dependency. that is where the lab should begin.

our founding proposition is that readiness has to become something another person can interrogate. name the system. name the harm. name the environment. name the assumptions under which the safeguards are supposed to work. then name the observation that would force you to withdraw your claim. without that last sentence, a readiness assessment can become an infinitely renewable permission slip. every failure is an edge case, every edge case is a future patch, and the original assurance survives untouched because nobody wrote down the conditions under which it was allowed to die.

the threshold feels impossible because the pieces constrain each other. you need evaluators capable of discovering failures that the developer missed, and access arrangements that let them do it. you need interventions that remain effective against capable systems, and organizations willing to pay the cost of using them. you need people who can contest decisions, and institutions that can process those objections before the disputed action becomes irreversible. none of this requires pretending that perfect certainty is available. it requires making the remaining uncertainty visible to the people who bear it.

we are not proposing a certificate that declares civilization ready for an undefined intelligence. AGI and ASI are contested labels; the work needs specific capabilities, deployment conditions and threat models. our proposed starting point is one concrete workflow at a time: an autonomous research assistant, a coding agent with limited permissions, a decision system embedded in an institution. identify where its assurance story breaks. reduce the problem until an experiment can distinguish a working safeguard from a persuasive description of one. publish the limits alongside the result. enlarge the claim only when the evidence earns it.

that demands a particular research culture. replication has to count as real work. negative results have to survive the embarrassment of publication. contributors have to retain visible credit. funding arrangements have to leave room for findings that inconvenience the funder. these are proposed terms of formation, and their credibility depends on the actual governance and financing we establish. putting them on a website does not make them true. building a lab around them would mean allowing them to constrain us when they become costly.

the first proposed output is a public threshold register: a small set of deployment claims, the evidence each would require, the strongest unresolved objection, and a named next experiment. an entry remains unresolved until the objection is answered or the claim is narrowed. the register would let a researcher join through a real problem rather than a declaration of allegiance. it would also let a critic show that our priorities are mistaken. a lab that cannot use that criticism has already confused recruitment with consensus manufacturing.

if you have a result nobody has integrated into a deployment decision, an evaluation that fails under realistic access, a control mechanism with an untested assumption, or an institutional design that only works while everyone is cooperative, this is the work we want to gather around. bring the thing that refuses to fit inside the reassuring story. bring the experiment that might prove your own view wrong. the threshold does not become less necessary because crossing it would rearrange the incentives of powerful people. that is precisely why a lab should form around it.

Proposed first work

Produce one threshold-register entry with a concrete deployment claim, explicit assumptions, a counterexample to investigate, a reproducible test plan, and a decision that would change if the test failed. Invite independent criticism before expanding the register. This is a proposed program, not a completed result.