Research notes
Technical notes on the open questions of AI governance engineering: what is assumed, what the public evidence shows and what would settle each one.
A research note takes one open question of AI governance engineering and sets out what is assumed, what the public evidence shows and what would settle it. Each note is versioned, cites a numbered source for each factual claim, names its author and says which controls and patterns it argues for. A note stays a draft until a named reviewer has checked it; none has been reviewed yet.
Each note works on one open question of governing AI systems in practice.
Notes
- The evaluation environment is part of the system
An evaluation measures a model inside a harness, tools, credentials and a network. This draft argues the result is evidence only when all of it is recorded.
Planned themes (not yet written)
The questions the next notes will take up. They are not written yet, so none has a page; each is listed here so that anyone working on the same question can say so.
- Control boundaries for agents. Where should the limit on what an agent may do be enforced: at the tool, the credential, the network or the orchestrator?
- From model evaluation to assurance evidence. What does an evaluation result need to carry before an assessor can rely on it as evidence that a control works?
- Stop conditions across agent hops. How does a stop signal reach every agent in a delegation chain, and how is it shown afterwards that it did?
- Mapping open controls to certification schemes. How can one open control be mapped to several certification schemes without implying conformity to any of them?
How a note is reviewed
A note is first published as a draft, with a 0.x version and a status line that says it is open for technical review. Anyone can review it through the issue form linked from the note: name the claim, say whether the cited source supports it, and give a source if it does not. A review that changes the note bumps its version and is credited by name. A note is marked published only once a named reviewer has reviewed it. The versioning rules are in the methodology; the ways to take part are on how to contribute.