Why is BGP down?
May invite a plausible summary before the evidence surface is complete.
NetBrain LIVE 2026 / Field Note
What I learned talking about AI and networking at NetBrain LIVE 2026.
AI is most useful when it extends an engineer's ability to reason across validated evidence without hiding uncertainty or human accountability.
After the conference
The hardest part was not explaining AI. It was standing in front of people and making the engineering idea legible.
NetBrain LIVE 2026 became a practical test of the same thesis Bradford Informatics is pursuing: tools can amplify expertise, but they cannot assume human agency, judgment, accountability, or the responsibility to communicate clearly.
The durable lesson is not about a particular model. It is about how network evidence, AI reasoning, and engineering judgment should interact.
A better diagnostic question
Move from “give me an answer” to “show me the evidence path.”
Topology→Peer→Prefix→Policy→Best Path→RIB→FIB→Forward Path→Return Path→Historical Delta→PASS / FAIL / WARN→Evidence
Prompt architecture
The detail in the request changes the quality of the investigation.
Establish topology and peer state. Trace the affected prefix through policy, best-path selection, RIB, FIB, forward and return paths. Compare against historical state. Separate observed facts, inferences, missing evidence, and architecture impact. Return PASS / FAIL / WARN with supporting evidence.
Forces the reasoning process to remain attached to inspectable network state.
The relationship
NetBrain observes the network. Bradford is exploring the decision context around the observation.
Network evidence
→Topology, path, configuration, routing state, history.
Reasoning
→Hypotheses, missing evidence, correlation, questions.
Architecture intelligence
Intent, failure domains, decision provenance, human authority.
Bradford Informatics is independent. This page does not imply NetBrain endorsement, certification, funding, partnership, or a production L.U.C.I. integration.
What changed after the event
The public site now has to demonstrate the relationship, not advertise it.
01Let the visitor challenge the AI-style hypothesis.
02Let infrastructure evidence overturn it.
03Let human intent change the architecture evaluation.
04Preserve why the conclusion changed.
05Keep the final authority explicit.
Engineering practice / beyond the conference
Better prompting makes AI accountable to evidence.
A useful network investigation begins with a bounded question. Which source, destination, prefix, time window and intended behavior are we discussing? Without those boundaries, a plausible explanation can outrun the observations that support it.
Separate the planes before drawing a conclusion
A healthy neighbor relationship is evidence about adjacency. A selected route is evidence about control-plane choice. Neither alone establishes successful application forwarding. Compare route selection with installed forwarding state, the actual forward path, return path and the security boundaries crossed by both.
Ask what would prove the hypothesis wrong
If the hypothesis is a policy error, request the relevant policy and its historical delta. If a transport interface was already down, revise the investigation. If two carriers share one physical entry point, do not equate provider diversity with failure-domain independence.
Give the next engineer a decision, not a transcript
Record the intended outcome, evidence source and observation time, unresolved assumptions, rejected alternatives and the person authorized to decide. Attach invalidation conditions: a topology change, new policy, expired evidence or a changed requirement should reopen the record.
A reusable review sequence
- State the architecture intent and acceptable failure conditions.
- Collect timestamped observations across topology, policy and forwarding.
- Ask for competing explanations and the evidence that distinguishes them.
- Validate the selected explanation in a controlled scope.
- Have a human accept the limitations and define revalidation triggers.
This is Bradford Informatics’ engineering perspective, not a claim of NetBrain endorsement or a shipping NetBrain integration. The public Lab demonstrates a deterministic model of this process.
Try the investigation in Decision Lab 2.0 →