Listen to this article
In this issue - 7 parts
Key takeaways
Did the FAA ground stop follow SMART AI expansion?
Yes. FAA announced SMART use across all 22 centers on November 10, three calendar days before the November 13 stop. That announcement does not establish an exact activation time. There is no evidence SMART caused the outage; FAA, DOT and vendor ASI deny a connection.
What is the preliminary cause?
Officials attribute the disruption to routine L3Harris copper-to-fiber router work and excessive connection attempts. The exact fault remained undetermined in the sources checked November 15. An ongoing investigation is distinct from a final root-cause finding.
What is the change-governance lesson?
Critical-path changes need a verified record, a known-good reference, isolation and a bounded canary with explicit success requirements. Contractor work needs the same operator-defined validation as internal work. Keystone is designed to apply those gates, halt failed expansion and support configured rollback with human accountability.
Key facts: citable summary
- Advisory 111 became effective at 20:54 UTC November 13; its event field starts at 20:53. General reopening was at 01:15 UTC November 14, approximately 4 hours 21 minutes later; some departures resumed earlier. Source: FAA advisory 111; reopening and early-departure records linked in Part 1.
- FAA reported instability at 17 of 22 centers and precautionary disabling of broadcast functionality on approximately 60 similar routers. The agency and L3Harris continued investigating. Source: FAA Q&A, November 15.
- Reuters reports that an FAA official said a safety assessment occurred before the work. The reviewed record does not establish skipped approvals. Source: Reuters via AnewZ, November 15.
AuthorityGate analysis: nearby announcements cannot establish causation. Change attribution and containment both depend on verified scope, baseline behavior and evidence about the candidate under traffic.
Keystone is designed to classify critical changes, notify owners, validate isolated candidates and preserve human authority and recovery evidence. The network scenario is an application of that model, not a claim of an FAA deployment or a guaranteed outcome. Read the Keystone validation model.
Part 1 of 7
Three Days Apart Does Not Mean Connected
The FAA's nationwide departure stop on November 13 came three calendar days after its November 10 announcement expanding SMART across all 22 centers. That sequence is real. There is no evidence SMART caused the outage. DOT explicitly denied a connection; the FAA's subsequent Q&A says: "No. The two systems are not related." The agency says SMART does not feed information into the air traffic control system. Officials instead point to separate telecommunications maintenance.
According to DOT's November 13 briefing, radar and communications instability began at 3:36 p.m. EST, or 20:36 UTC. FAA advisory 111 imposed the nationwide stop, effective at 20:54 UTC. Advisory 120, effective at 22:46 UTC, allowed departures coordinated with the Command Center as facilities recovered. DOT announced general reopening at 8:15 p.m. EST, or 01:15 UTC November 14.
The interval from the first advisory's effective time to general reopening was approximately four hours 21 minutes. Its event field begins one minute earlier. Neither duration means every departure waited for the entire interval. Duffy said approximately 3,500 airborne aircraft landed safely. A national departure restriction also does not mean every radar feed and radio frequency disappeared.
For change owners, the question is how a local intervention can disturb a shared critical path. Physical remoteness does not establish a small dependency footprint. A release boundary must follow the behavior that can propagate beyond the changed device. That is the starting point for validation, regardless of whether the author is a contractor, an internal engineer or an AI agent.
From instability to general reopening
- 3:36 p.m. EST
- 3:54 p.m. EST
- 5:46 p.m. EST
- Nov 13, 8:15 p.m. EST
UTC dates are explicit across midnight. Local times use EST (UTC-5); Chicago CST is one hour earlier. Milestones are not evenly spaced measurements of recovery.
| Record and date | Evidence status | Boundary on the conclusion |
|---|---|---|
| FAA advisory 111, November 13 | CONFIRMED: 20:54 UTC effective/signature time. | Event field begins 20:53; those fields differ by one minute. |
| FAA advisory 120, November 13 | CONFIRMED: 22:46 UTC effective/signature time. | Coordinated departures allowed before general reopening. |
| DOT, November 13 briefing | CONFIRMED: general reopening 01:15 UTC November 14. | Airline recovery remained gradual. |
| FAA, November 10 | CONFIRMED: nationwide SMART expansion announcement. | No exact national activation timestamp. |
| DOT, November 14 | CONFIRMED: preliminary router-work explanation. | Exact equipment or installation fault undetermined. |
| ASI via NPR/GPB, November 14 | REPORTED: vendor denial of a SMART connection. | Statement reproduced by the outlet; original X retrieval blocked. |
| PASS, checked November 15 | CONFIRMED union position: technicians not involved. | Does not establish that no safety assessment occurred. |
Part 2 of 7
The Router Explanation Is Still Preliminary
Duffy's November 14 account describes routine L3Harris work replacing copper connectivity with fiber at a remote location. A router began flooding peers with over 1,000 connection attempts every four to five seconds, disrupting communications and radar feeds. Whether installation error or faulty equipment produced that behavior remained unresolved. Duffy said: "We haven't determined what the fault was."
The FAA's November 15 Q&A identifies two potential originating locations: a radar site in Amarillo, Texas, and a radio communications site in El Dorado, Arkansas, where L3Harris had deployed a new router. Blocking those locations stabilized the network. FAA reported instability at 17 of 22 Air Route Traffic Control Centers, spreading into several terminal approach facilities. Broadcast functionality was disabled on approximately 60 similar routers as a precaution.
That is evidence of isolation and a disabled feature. It does not establish a full software rollback, a specific defective configuration, or the final fault. L3Harris says it is supporting the FAA investigation. In his November 15 CBS interview, Duffy said both routers were being examined and that the Amarillo work had occurred weeks earlier; he still described the root cause as under investigation. Keep that open question separate from the observed propagation and mitigation: a useful review can test a failure mechanism without pretending the underlying defect is already settled.
The sibling Microsoft 365 gateway outage analysis examines a latent July update defect exposed months later by failover and reduced capacity: validation needs representative load and a proven recovery path. The incidents have different technical records. Their common governance problem is the authority of a small component over a much larger service.
Our take
A critical path deserves critical validation even when its maintenance is routine. The record supports asking how propagation was bounded, what behavior was tested, and how recovery was verified. It does not support alleging that FAA approvals or assessments were skipped.

Part 3 of 7
Recovery Has More Than One Clock
A stable network and a restored airline schedule are different outcomes. Aircraft, crews and passengers need to reconnect after the departure constraint lifts. Reuters' November 15 report describes a forecast that schedules would largely recover by late Sunday. In a later November 15 CBS interview, Duffy said delays remained and looked to the end of Sunday and into Monday for full capacity. Neither forecast is evidence that recovery had finished. The FAA's November 15 account described stable systems and operations returning to normal.
Cancellation counts also need their measurement boundaries. FAA reported approximately 12,462 Friday cancellations, including 9,082 by major airlines, and warned that not all were directly caused by the ground stop. FlightAware via Reuters reported about 9,200 cancellations and 7,500 delays Friday. NPR reported nearly 25,000 Friday delays. The reviewed accounts do not reconcile those scopes or methods; these numbers must not be added together.
The Saturday Reuters snapshot was explicitly timed: more than 6,400 cancellations and 8,000 delays as of 9:35 p.m. EST November 14, or 02:35 UTC November 15. A later snapshot can change the count without changing what an earlier report said. Record the period, publisher and cutoff whenever a number becomes evidence for an operational decision.
The same discipline belongs in business continuity and recovery planning. Define success for the technical dependency and for the business workflow it supports. A green device status alone cannot prove that dependent services are stable under traffic. The recovery requirement should include the work people actually need to complete.
Keep the labels attached to the counts
Friday cancellations; includes 9,082 major-airline cancellations. Not all directly caused by the stop.
Friday cancellations in the Reuters report. Separate snapshot; scope is not reconciled with FAA.
Both bars describe November 13 schedules. FAA Q&A published November 15; Reuters/AnewZ published November 15, displayed 10:55 (timezone unspecified). Neither gives a Friday extraction time. Common axis: 0 to 12,462 cancellations. Bars are not additive or incident-only counts. Delays are excluded from this chart because the reports conflict.

Part 4 of 7
Attribution Needs a Verified Change Record
When a telecom change and an AI planning expansion land days apart, the incident room must distinguish candidate causes quickly. The first hours can become a question of which change mattered, whether AI was involved, or whether the event was an attack. FAA says it has no indication of a cyberattack. Timing can guide investigation; it cannot choose the answer. A verified change record connects the exact candidate, owner, target, execution time, dependencies, observed behavior and recovery action. A known-good baseline shows what the affected path looked like before the intervention.
That record also supports bounded statements about what did not change. Evidence that a planning tool cannot write into a control system is more useful than an assumption based on its name. SMART means Strategic Management of Airspace, Routes and Trajectories. DOT selected Air Space Intelligence for SMART and the related FMDS platform; the FAA fact sheet describes planning inputs and routing recommendations.
Limited operational use began September 21. FAA says SMART cannot replace controllers or control aircraft. Its November 10 description puts specialist validation before recommendations reach the Command Center and local leadership. That is a human-in-the-loop planning pattern. It is a separate control from testing whether a router change preserves communications.
Contractor boundaries need equally clear evidence. PASS says the affected routers were contractor-owned and maintained and its technicians were not involved. That is the union's position. Reuters reports an FAA official said a safety assessment occurred. Both statements can be recorded without asserting a missing assessment. The operator's validation questions remain: who supplied behavioral evidence, who could halt expansion, and which recovery requirements applied?
There were also pre-incident consultation concerns: FLYING reported NATCA's September statement that it "has not been involved in the design, testing, or implementation" of SMART. That reported position concerns participation, not outage causation. Human oversight works best when the people accountable for operations can shape the acceptance requirements and understand the evidence they receive.
A change record should let the incident team explain what changed, what stayed known-good, and which evidence separates the two.
Part 5 of 7
Keystone: From Detection to a Bounded Canary
Consider an operator-controlled network path carrying critical communications. A contractor proposes a router configuration or firmware change. This worked scenario applies AuthorityGate Keystone's validation-layer design and eight configurable validation gates. It is invitation-only early-access design intent. The network checks below are requirements the operator would define, not a claim of an existing FAA integration or support for a particular router.
1. Detect, classify critical, and notify
The documented detect-notify-validate response model identifies a change and compares behavior with known-good. The owner classifies this target as critical because of its dependencies and propagation potential. Capture the exact candidate, contractor identity, authorized scope and maintenance window; notify accountable owners. A detected bypass needs validation and response, not silent adoption as the new normal.
2. Protect known-good and isolate the candidate
Known-Good Mode supplies the environment-specific reference. Gate 1 checks that reference and a recoverable backup; dependency checks identify what shares the path. Isolate the candidate so its test behavior cannot exercise unrestricted authority over production peers. If the operator cannot demonstrate that boundary, hold the change. A small site alone is not an isolation boundary.
3. Validate below production, then push a small ring
Gate 6, Behavioral Resilience (BlockStack), runs the candidate in a production-mirroring lower environment against the baseline. In this configured scenario, passing those checks makes the candidate eligible for a small canary ring. Its permitted peers, shared settings and traffic exposure remain constrained. A ring that can flood the wider network has not contained the behavior being tested.
4. Require explicit behavioral success
Write success requirements before traffic arrives: routers do not flood peers; connection-attempt rates stay within approved baseline tolerances; radar and communications feeds remain stable under representative real traffic. Test recovery behavior as well as startup. The operator supplies the measurements, observation period and tolerances. This article invents no universal threshold or FAA telemetry. Missing observations hold promotion rather than count as a pass.
5. Halt and roll back a failed canary
Configure the workflow to halt expansion automatically when a requirement fails and roll the failed canary back to verified known-good. Gate 8 requires a tested recovery path and named recovery owner before release. Recheck service behavior after reversion. Isolation bounds the effect while automatic stop and rollback keep a failing candidate from promotion; rollback alone cannot undo disruption already propagated outside the ring.
6. Keep exceptions human and acceptance auditable
Gate 7's named SME approval puts consequential decisions and exceptions in the execution path. The reviewer receives risk evidence and prior gate outcomes. Required approvals remain binding after technical checks pass. Record detection, policy, scope, observations, human decisions and recovery in the audit trail. Adopt an accepted baseline through the governed process, preserving the earlier reference for attribution.
This is the class of failure the isolation-plus-canary model exists to stop: a local change propagating across a shared critical path. Its effectiveness depends on coverage, representative tests and an enforceable boundary; the public record cannot establish what would have happened with it at the FAA.
Bound the effect before widening the release
- 01 / CONTAINIsolate
Protect known-good; constrain peers and shared dependencies.
- 02 / OBSERVECanary ring
Push a small ring after lower-environment validation.
- 03 / REQUIRESuccess gate
No peer flooding, baseline connection rates, stable feeds under traffic.
Promote within approved scope; continue comparison.
Halt expansion; roll back failure. Hold missing evidence and route exceptions to a named human.
Conceptual configured workflow, not a product screenshot or claim about FAA controls. Automatic recovery requires an integrated, tested execution path.
| Failure mode or validation question | Control and supporting site page |
|---|---|
| A local change can reach many peers. | Dependency health and bounded behavioral validation; verify the isolation boundary before release. |
| A legitimate vendor update behaves badly here. | Update validation against the receiving environment, beyond provenance checks. |
| The canary fails or recovery is unproven. | Known-good comparison and policy-controlled rollback; tested recovery before execution. |
| An exception or attribution dispute needs evidence. | Named human authority plus recorded assessments, decisions and outcomes. |
Questions every operator should ask; no finding that FAA lacked these controls.
Part 6 of 7
AI Validation Must Operate Inside Your Rules
A national system changes across sites, contractors, software and dependencies. Manual review alone cannot continuously validate that mass of interacting behavior. AI-assisted validation is required to scale the analysis while keeping the operator's rules enforceable. That is AuthorityGate's engineering position, not a claim that this incident was an AI failure or that AI guarantees safety.
AI change governance connects policy to execution. AI can assist comparison, risk analysis and anomaly detection; owners define the scope, thresholds and evidence required for acceptance. The augmented AI model retains accountable human judgment at consequential boundaries. The validator must not relax a success requirement because the release queue is busy.
The update-validation design applies the same logic to firmware, application updates and vendor-pushed changes. Vendor integrity and authorization are useful evidence, but neither proves compatibility with the receiving environment. Require the contractor to supply the exact candidate and test evidence, and make ownership of halt, rollback and recovery explicit across the contractual boundary.
Coverage matters as much as the sophistication of the analysis. A change outside the managed path cannot be credited as validated. Surface bypasses, preserve the old baseline, and reconcile observed changes with approved records. The AI Change Governance Field Manual provides that evidence-oriented operating model. Start with one critical path whose authority and recovery can actually be exercised.
Part 7 of 7
What the Next Review Should Resolve
Oversight predates this outage. An October 13 congressional request sought a GAO audit of SMART and FMDS, covering risk management, testing, metrics and deployment. A request is not an audit finding. GAO's September 14 modernization report identified missing lifecycle-cost planning and an integrated master schedule, including 11,389 separate project schedules. That broader report does not determine this incident's cause.
After the stop, Larsen and Carson's November 14 statement described a planned Monday staff briefing on requested funding and expected answers at an early-December hearing. At this article's November 15 check, the reviewed record contained neither a completed post-incident briefing nor a final root-cause report. The exact fault remained under investigation. Duffy's Sunday interview still forecast full capacity from late Sunday into Monday rather than confirming completed airline recovery.
For your own review, require three demonstrations: reconstruct the actual change against verified known-good; show that a failing canary cannot reach the wider dependency path; and rehearse halt and recovery under traffic. Include the contractor and the operational owner. If a requirement is missing, give it an owner and a measurable acceptance condition before the next release.
The NATS latent-defect and fallback analysis offers another aviation case for examining what recovery really proves. Use the Keystone and AI governance FAQ to align terminology, then bring one critical change path to a discovery conversation with AuthorityGate's architect team. The useful next step is a bounded validation contract with evidence, accountable authority and a tested way back.
Questions this article answers
Did SMART AI cause the November 13, 2026 FAA ground stop?
There is no evidence that SMART caused the outage. FAA and DOT explicitly deny a connection; ASI also denied it. SMART expansion was announced November 10. Officials preliminarily attribute the November 13 disruption to separate L3Harris copper-to-fiber router work.
How long did the nationwide ground stop last?
Approximately 4 hours 21 minutes, from advisory 111 effective at 20:54 UTC November 13 to general reopening at 01:15 UTC November 14. Its event field starts one minute earlier. Advisory 120 allowed coordinated departures before general reopening.
What router failure did officials describe?
DOT described over 1,000 connection attempts every four to five seconds during routine telecom work. FAA reported instability at 17 of 22 centers and precautionary disabling of broadcast functionality on approximately 60 similar routers. The exact underlying fault remained under investigation on November 15.
How many flights were canceled or delayed?
FAA reported approximately 12,462 Friday cancellations, including 9,082 major-airline cancellations, not all directly caused by the stop. FlightAware via Reuters reported about 9,200 cancellations and 7,500 delays Friday. NPR reported nearly 25,000 delays. These snapshots are not reconciled or added.
Were FAA safety assessments or approvals skipped?
The public record reviewed does not establish that required assessments or approvals were skipped. Reuters reports an FAA official said a safety assessment occurred. PASS says its technicians were not involved; that is the union's position and does not prove no assessment occurred.
How does SMART retain human control?
FAA describes specialist validation of SMART recommendations, routing through the Command Center, and local leadership consideration. Its September launch statement says SMART cannot replace controllers or control aircraft. Human review of recommendations and validation of infrastructure changes are separate controls.
What does Keystone do when a critical change fails its canary?
In the configured design scenario, Keystone halts expansion and automatically rolls the failed canary back to verified known-good. Operator-defined success requirements, isolation, required human approval for exceptions and audit evidence govern promotion. This is early-access design intent, not a claim about what would have happened at the FAA.
Was airline recovery complete when this article was updated?
FAA reported stable systems and returning operations on November 15. Reuters reported a late-Sunday recovery forecast; in a later Sunday CBS interview, Duffy said delays remained and expected full capacity from late Sunday into Monday. Neither confirms completed airline recovery. The reviewed record contained no final root-cause report or completed post-incident congressional briefing.