ONUG Fall 2026 (opens in a new tab) · Tools, Technologies & Techniques theater · Thursday, October 29, 3:15 pm ET
Building confidence in change outcomes. Raja Tadimeti’s twelve-minute talk: one routine change, closed as successful at 21:31, that became a P1 at 02:07; the question that catches it; a validation plan derived from the ticket the engineer already wrote; and a checklist you can use on Monday, with or without AI. The slides and the talk in writing are below. No form.
Twelve minutes
One night, one word, one question, one plan, four practices, a take-home. One fictional scenario carried through; one product screenshot, labeled illustrative; no competitor named; no figure REAP has not measured. The one third-party figure (Uptime Institute (opens in a new tab), Annual Outage Analysis (opens in a new tab) 2025) carries its source and year.
Co-founder and CEO, REAP · the speaker
Co-founder of WiteSand (opens in a new tab) (acquired by Juniper Networks (opens in a new tab)) and founding engineer at Pensando (opens in a new tab) and Insieme (opens in a new tab); decades shipping networking products at Cisco (opens in a new tab) and Juniper (opens in a new tab); named inventor on 17 granted US patents. He has run networks, and the nights he remembers are the ones where nothing broke and nothing alarmed.
LinkedIn profile of Raja Tadimeti (opens in a new tab)Eighteen slides. The slides (PDF) · the talk in writing (PDF) · the checklist below and on the back of the session card.
The take-home
Fifteen items in three columns, with or without AI. The “Before” column is the validation plan the talk derives from the ticket: outcome, scope, invariants. The “After” column is the comparison; the “Always” column is the governance.
The scenario
Fictional, and built to be real enough to argue with. It is the same incident the Investigation Window on this site illustrates; nothing in it names a company, a person, a vendor or a product.
| The facts | |
|---|---|
| The company | A distributed manufacturer with fourteen plants (fictional). A lean network team; ServiceNow (opens in a new tab) for change; a core pair per plant. |
| The change | CHG-1234, Tuesday 21:14, on plant-a-core-2. Object group OG-ERP had its members replaced (10.20.5.0/24 → 10.20.50.0/24) so a second ACL, ACL-B, could reuse the group for the new order-entry subnet. ACL-ERP-IN, which permits ERP return traffic, already referenced OG-ERP. |
| The checks | Interfaces up, OSPF adjacency stable, ping to the core, ACL-B permitting the new subnet, no command errors. Closed Successful at 21:31. |
| What broke | ERP return traffic to Plant A order entry fell through to the final deny on ACL-ERP-IN from 21:16. No link or neighbor changed, so nothing alarmed. Order entry is quiet at night; the 02:00 order sync failed; the P1 opened at 02:07. |
| How it was found | By reading the configuration history by hand at 03:40; restored at 04:10; explained the next day. Undetected for 4 h 51 min. |
| The evidence that was available at 21:20 | The configuration diff 21:13 → 21:14 (the OG-ERP and ACL-B lines, exactly as the ticket predicted). The ACL-ERP-IN final-deny counter 3,412 → 3,918 across two reads 25 seconds apart, flat before the change. A traceroute from Vlan 312 to 10.20.5.10 with no reply beyond plant-a-core-2. |
| The plan that would have caught it | Outcome: ACL-B permits 10.20.50.0/24. Scope: plant-a-core-2, OG-ERP, ACL-B. Invariants: ACL-ERP-IN still permits ERP return traffic (it also references OG-ERP); the return path from Vlan 312 unchanged; the core pair in sync and OSPF costs unchanged. Re-check: after the 22:00 cache expiry and the 02:00 sync. |
In 2026
The talk puts an agent, a pipeline or a compromised credential in the engineer’s chair, with four of the ten AI-era scenarios: an agent changing the network outside a window; one bad change pushed to two hundred devices; redundancy removed before the change that matters; and “no impact expected” closing the change that caused the outage. Every one of those changes succeeded, and the checks that were silent for a person are silent for an agent.
What happens, why it hides, what evidence-backed assurance does in each, and a ten-statement preparedness check. The 2026 facts, from the primary sources.
Read the scenariosShow me the evidence. Show me what it rejected. Show me what happens when it is wrong.
Read-only by default, approval-gated when it acts, every conclusion traceable to its evidence: that is the bar.
The buyer’s evaluation guideGo further
The seven principles of network assurance, and why a restore without a root cause is a deferred outage.
Read the manifestoPre-execution impact analysis, before-and-after comparison, one of five outcomes written into the change record with its evidence.
How it worksTwenty questions, ten minutes: your stage across Observe, Reason, Act and Govern & Prove, and the first door to open.
Take the assessmentTwenty-minute meetings at Booth 33, the eight-question assurance check, the Mac mini drawing on Thursday at 4:00 pm ET, dinner on October 28, and the evaluation kit.
Event detailsKeep the root cause to yourself; in a 30-minute session we walk through how REAP would have investigated it, evidence first. The replay on your own network happens inside the proof of value. Or book twenty minutes at ONUG (opens in a new tab).