AI safety, alignment and governanceEden Research · Collaboration
Begin with one question you can challenge
The first request is specific: review the P21 analysis bound for unusable replies. Does it cover every later input that an unusable reply could change, and does the statistical decision stay within what that bound proves?
P21 compares correcting each record as it arrives with inspecting the whole register as it grows. The experiment concerns a defined arrangement of what each checker sees and may do. The next useful step is a design that lets a critical reader tell an effect of that arrangement apart from extra computation, measurement noise and the handling of failed replies.
You need not accept the wider ARC Theory to make a valuable contribution. A counterexample, a narrower reading or a stronger control can each change what the experiment is able to establish.
Who the work needs now
Replication teams. Run a registered test again, separately, under your own declared decisions, and publish what you find. A test run by a separate group is recorded as a replication, never as the programme’s own result.
Researchers in verification, formal methods and the security of AI agents. Challenge where a check should sit when work passes between agents acting for different owners, starting with the P21 analysis bound above.
Statisticians. Review a test’s decision rules, its margins and what its design can tell apart, before its data exists.
Teams that run AI agents for different owners. Tell us where your checks sit today, and whether a registered test could be run on your own systems with your own scope.
Eden Research plans an advisory board; its members and their roles are published only with each member’s written agreement. Each contribution starts from an agreed scope, time and output, and is described afterwards by what was actually done.
Paid work is scoped and costed before it starts. Authorship follows a substantive scholarly contribution and responsibility for the paper. A fee, an introduction or an advisory title does not buy a favourable conclusion.
What a reviewer receives
The first review package holds a short brief on the question, the comparison, its outcome and decision rules, and the evidence of what the design can tell apart. A reviewer can pick the part that fits their expertise and propose a bounded contribution.
The next study uses synthetic records only. Every result is published, whether it is positive, adverse or inconclusive, and a collaboration agrees to that at its start.
Write to me with your area of expertise and the part you would like to examine. A first conversation can settle whether a review, a funded piece of work or a later replication is the right fit.