Devin Arkema

Independent AI governance and systems ethics researcher who builds.

I write about how oversight of AI systems fails and what should replace it, and I build working systems that put those ideas under load. AI agents write most of the code, under my specs and the review gates I design.

Work

Longtable

A 24-hour in-person tabletop charity event for Extra Life, with a Discord signup bot, a pixel-art live hall and an opt-in consent model.

GURPS VTT Work in progress

A desktop virtual tabletop for running a GURPS campaign, developed partly by an overnight loop that builds one queued item at a time and stops for my review.

Wyrmling

A local AI agent that runs on my own hardware, with jailed tools, scheduled routines, and automatic failover from a GPU machine to a CPU machine.

How this library was built

The 3D library on the home page was built by AI agents working to my specs, and this page records how, with a source for every claim.

How I build

I design the systems, write the specs and build the checks. Claude and Codex write most of the code. Nothing ships until it has passed review gates I designed.

  1. Spec. Each piece of work starts as a written spec: what to build, which files it touches, and what done looks like.
  2. Implement. An agent, usually Codex, builds from the spec.
  3. Review. A different agent reads the whole diff, not a summary, and either accepts it, sends it back once with fixes, or takes the work over.
  4. Verify. The result has to pass objective checks, such as tests, type checks or a falsify gate, before anything merges or publishes.

The record is kept, not reconstructed. Commits made with an AI agent carry a co-author line, so the share is countable:

Counts exclude commits made by automated jobs and were taken on 22 October 2026. Longtable also keeps a provenance ledger with one line per task handed to Codex: the spec, the commit, how it was verified, and the reviewer's verdict. Across 47 tasks between 6 September and 21 October 2026, 13 were accepted as written, 20 were accepted after fixes, and in 14 the reviewer took the work over.

Policy

Suggestion 26-EV-1: Proposed Federal Rule of Evidence 615.1

2026

Submitted as an unaffiliated individual to the Advisory Committee on Evidence Rules of the U.S. Judicial Conference, and released on its public docket on 28 February 2026. It is pending consideration.

The rule would let courts, in multi-defendant conspiracy and RICO-type cases, temporarily keep co-defendants who choose to testify from hearing each other's testimony, then compare their independent accounts for consistency, under strict equalization and rights safeguards.

It applies the equalization principle from my paper Diagnostic vs. Defensive Information Control to the rules of evidence.

Selected papers

  1. From Static Oversight to Distributed Corrigibility

    PhilArchive

    2026

    Safety decisions made in the past hold too much authority over new evidence, so keeping AI correctable has to be a job for the whole system of institutions, not only for the model.

  2. Diagnostic vs. Defensive Information Control

    PhilArchive

    2026

    Hiding information to protect something is different from hiding it to watch how people behave while they cannot see, and the second kind should not feed decisions about them without disclosure and a way to contest.

  3. Interiority Policing: When Governance Reaches the Source Rather Than the Act

    PhilArchive

    2026

    When rules move from policing what people do to policing what they think or are inclined to do, the usual defenses against rules about actions stop working.

  4. Conviction Under Revisability

    PhilArchive

    2026

    A permanent commitment is legitimate when the person keeps the power to revoke it, which is why a monk under a vow and a prisoner who cannot leave are different even when they look alike.

  5. Dying as Time

    PhilArchive

    2026

    Living things stay alive by constantly breaking down and rebuilding themselves, so death is better understood as a process running through life than as a single event at its end.

  6. The Re-Derivation Test: An Epistemic Discipline for Beliefs Maintained by Their Own History

    PhilArchive

    2026

    Ask whether you would still hold a belief if the fact that you already hold it could not count as evidence.

  7. Free Won't vs Free Will

    PhilArchive

    2026

    The ability to refuse is a basic part of being an agent, not just the mirror image of the ability to endorse a desire.

All 37 works on PhilArchive, including the book below, are listed on my PhilPeople profile.

Arkemedics: The Recursive Ethic (2025) is the framework the papers build on, in four volumes: The Spiral Codex, The Anti-Codex, The Axis Codex and The Body of Sovereignty.

Other work

Overreach Work in progress

A push-your-luck combat ladder game, built for playtesting.

Play Overreach

Music

Three albums: Default Settings, Dragoon Nightmares and Sacrifice.

About

I am an independent researcher working on AI governance and systems ethics. I work in precision manufacturing and have spent five years in tool and die, where I am now a journeyman tool and die maker. That trade is where I learned how safety-critical systems fail, and why procedure matters more than good intentions when tolerances are tight.

Contact

devin@arkemedics.com

Email is the best way to reach me. My CV is available on request.