AI execution controls · Python systems · network automation
I build systems that control what AI agents are allowed to do and make those decisions inspectable.
I am CEO of Luftfiber AS, working with telecom and operational infrastructure in Norway. That experience shapes my software work: keeping credentials under control, containing failures, handling stale state and verifying what actually changed.
My core architectural principle is reasoning authority ≠ execution authority. A model can propose an action; approval and execution depend on authority established outside its reasoning loop.
I contribute technical review and executable conformance cases to discussions in CoSAI / OASIS WS4.
The evidence-sufficiency RFC #189 explicitly references my checker, regression artifacts and cases where identical observations can correspond to different underlying histories. Those cases show why missing evidence cannot automatically establish either conformance or a violation.
My contribution focuses on what a verifier can legitimately conclude, including NOT_ESTABLISHED, and on keeping test expectations separate from evidence inputs. This is participation in an ongoing standards discussion; it does not imply adoption or endorsement of my projects.
| Project | Purpose |
|---|---|
| REMORA Research | Research into policy-gated agent execution, authorization binding and evidence of external effects. Includes reproducible experiments, a claim register and published negative results. |
| Agent Authority Conformance | An independent evidence model for seven authority and execution properties. Draft v0.2 adds explicit inconclusive verdicts, stricter evidence requirements and a bounded reference checker. |
| Assured Agent Execution | Product-oriented control-plane work: separate approval and execution roles, exact-payload binding, selected effect verification and an auditable lifecycle. |
| vericlaim | CI checks that bind documented claims and benchmark numbers to their supporting artifacts, detecting drift when the evidence changes. |
| pilotfish | Early research into network link selection governed by policy, cost constraints and evidence freshness. Includes a decision core and simulator; router adapters remain future work. |
REMORA is research/shadow-mode software. Each project documents its own scope and limitations; the portfolio makes no claim of production certification or universal agent safety.
Three useful starting points:
- Claims and reproduction evidence — inspect the artifacts and caveats behind individual claims, with the machine-readable register as a companion.
- Negative results — failed hypotheses, withdrawn claims and known limitations remain part of the record.
- Public technical discussion — follow the reasoning, objections and unresolved questions around evidence sufficiency.
For a longer read, see the REMORA paper (PDF). The v0.11.0 release provides downloadable artifacts and checksums. AAE documents its separate pinned core and verification procedure.
I start with a bounded property, identify the evidence it needs and build a test that can expose a violation. Claims stay tied to revisions and explicit assumptions. When a result fails to support the claim, I narrow the claim or change the implementation.
My work spans Python, distributed systems, policy-as-code, identity and access controls, MCP/tool integration and network automation.
Through Luftfiber, I am interested in engineering work on governed AI integrations, Python automation and reliable operational infrastructure. I also welcome independent replication, adversarial review and collaboration on executable conformance cases.
A useful starting point is a concrete workflow: which actions it needs, who can authorize them and how their effects can be checked.
Contact: support@luftfiber.no