Eden Research · Laboratory

Give the next claim somewhere to fail

A useful laboratory gives a hard question somewhere to fail. Eden Research is the research programme I founded to find out whether correction can keep working as AI systems improve their own work.

I believe the stakes go beyond better benchmark scores. A system that keeps improving may become harder to correct, and losing the ability to correct increasingly capable AI could pose an existential risk. The programme asks what could prevent that, which measurements would reveal a loss of correction, and what evidence would justify relying on this account in practice.

I am Michael Darius Eastwood, founder of Eden Research and of Recursive Dynamics. I work outside any university and do not hold a PhD, so the work asks to be judged on its record: the founding paper, the dated research record and Paper C, on polymathy and neurodivergent cognition. Where a role is planned rather than filled, these pages say so.

The theory and the experiment meet at a measurement

Recursive Dynamics studies systems whose improvements feed later improvements. The ARC Theory states five laws under test about growth, correction and persistence; Laws IV and V were added on 24 September 2026, and their mathematics is in development. The Eden Protocol develops a way to place corrective principles inside that process. The operational definitions and the theory registration fix what each term means when it is measured.

The next study, P21, builds a register of synthetic transaction records one entry at a time. It compares a corrector, which can repair each entry as it arrives, with a monitor, which checks the whole register at each round. Fixed rules decide whether an entry is right, so the study counts its outcome without asking a general-purpose judge whether a system is ethically aligned. Its design is under review, and it has not produced an empirical result.

That is one part of the programme. P21 does not measure changes to model weights, retention through later training, or the correction exponent of Paper X; each of those questions needs its own instrument and its own design, stated before it runs.

Keep the path back to the source

I built the source-and-review system first for my legal cases and adapted it to research, which follows the same pattern of sources, citations and precedents. Both use a three-valued gate (allow, hold or deny) whose checks run in code before a change is committed. The record they leave includes adversarial refutation rounds, source and quotation checks, release checks, dated revisions and published corrections. These checks are internal. They do not replace review or replication by others, which the programme invites; the evidence record shows what was checked and how far each check reaches.

A study begins with a stated question and a decision rule. Its package names the protocol, the code, the model configuration, the cost limit and the reporting rule, and its registration is published before the study runs. Failures, deviations and adverse results stay in the record.

Past tests keep their original versions. A revised test is a new test, in new files with its own registration, so a later repair never rewrites what an earlier result meant.

Build the organisation around the research

Eden Research will form an advisory board; its members and their roles are published only with each member’s written agreement. Scoped research collaborations, and funding for the work those people would do, come next. Statistical review, implementation, research operations and financial administration are separate responsibilities, and a plan to recruit for them is described as a plan.

Eden Research is the name of the programme. If a company is formed, these pages will give its registered name and number. Funding applications name the actual applicant and describe each collaborator at the level of commitment that person has agreed.

Start with the methods, or open the study or paper behind the claim that interests you. The collaboration page names the first piece of work where criticism from outside would help most.