ARKHĒ

WORK WITH ARKHĒ

Bring one important uncertainty.

Independent reliability reviews for teams shipping AI-enabled software. Turn a difficult behavior or disputed result into evidence your team can act on.

Ask for a scoped review

01 / FIT

When the answer changes what you ship.

Your agent is inconsistent on a critical workflow. An evaluation result is difficult to trust. A generated change passes checks but appears to alter behavior. There is a concrete question, an owner who can act on the answer, and a system we can examine with permission.

Start with one workflow, component or claim. We agree the question, access, time budget and deliverables before work begins.

02 / DELIVERY

Evidence, a decision, and a usable handoff.

A

A reproducible diagnosis

A clear baseline, selected cases and controls, with enough context for your team to repeat the investigation.

B

Prioritized next steps

What happened, why it matters, how confident we can be, and which changes would reduce the uncertainty.

C

Verification that lasts

Regression checks and a verification plan. Implementation and repair verification can be included in the agreed scope.

03 / ENGAGEMENT

A focused first engagement.

  1. Define the decision.

    Share a short description of the behavior, your current checks and what the answer would change. The first conversation determines whether there is a useful fit.

  2. Agree a paid, bounded scope.

    A written proposal sets the deliverables, schedule, fee, access and any compute budget. A short diagnostic can establish whether a deeper investigation is worthwhile.

  3. Work directly with the researcher.

    Adem Vessell leads the investigation, uses AI tools where useful, and owns the conclusions. Work is primarily asynchronous, with agreed checkpoints and a final walkthrough.

Security-specific testing requires its own written scope. A review documents evidence and limits; it is not a certification or a guarantee of a particular finding.

04 / SELECTED WORK

A contribution record you can inspect.

NVIDIA SkillEvaluator: uncertainty and paired outcomes
xDSL: six compiler and interpreter correctness fixes

Open proposals and review contributions →

Accepted upstream work. These examples are not client engagements or organizational endorsements.

What needs to be understood?

Email a short description of the question and the decision it affects. Keep credentials and sensitive data out of the initial note.

Start a conversation