FDAI uses several scenario sets for different questions. A catalog entry proves
that a scenario is well-formed and connected to a known signal. A replay proves
that the control loop reaches the expected result. A live enforcement run proves
that a real fault was injected, observed, and reverted on a disposable Azure
substrate. These evidence levels are not interchangeable.
This page lists every current SRE scenario set and every one of the 135 catalog
scenario IDs. The snapshot was recalculated from the repository on 2026-07-17.
| Scenario set | Size | Current evidence | What it proves |
|---|
| SRE demo pack | 18 | 18/18 observation mode coverage | Detect, route, RCA, governed response, and validation mapping exist |
| Reference live-enforce sweep | 10 | 10/10 validated, detected, and reverted | Real injectors and probes worked on the disposable Azure substrate |
Frozen control-loop set v2026.07 | 9 | 66 integrity and replay tests passed | Balanced Change, DR, and FinOps expected outcomes replay through the shipped loop |
| Agent decision scenarios | 6 | Included in 22 agent scenario tests passed | Forseti returns expected auto, human approval, deny, or hold for review outcomes |
| Agent pipeline scenarios | 8 | Included in 22 agent scenario tests passed | Forseti, Thor, Var, and Saga preserve cross-agent safety invariants |
| Chaos catalog | 135 | 135/135 schema, signal, and symptom-index validation | Every catalog record is structurally valid and searchable |
| Default-factory dispatch subset | 93 | 93/93 injector and probe pairs build in dry-run | Delivery wiring exists; this is not a live fault run |
| Promoted catalog | 0 | No entries promoted | No collected entry has inherited enforce eligibility |
A separate 14-case Command Deck claim corpus validates answer evidence check with
zero unsupported escapes and zero clean rejections. It is a conversational
safety corpus, not an SRE operational scenario set, so it is not added to the
scenario totals on this page.
| ID | Scenario | Type | Validation |
|---|
| S1 | AKS pod kill | Fault | Observation mode coverage; live enforcement via aks-pod-kill |
| S2 | AKS pod CPU stress | Fault | Observation mode coverage; live enforcement via aks-pod-cpu-spike |
| S3 | AKS pod network latency | Fault | Observation mode coverage; live enforcement via network-rtt-delay |
| S4 | AKS HTTP abort | Fault | Observation mode coverage; live enforcement via aks-http-abort |
| S5 | VM CPU stress | Fault | Observation mode coverage; live enforcement via vm-cpu-stress |
| S6 | VM memory stress | Fault | Observation mode coverage; live enforcement via vm-mem-stress |
| S7 | VM network latency | Fault | Observation mode coverage; live enforcement via network-rtt-delay |
| S8 | MySQL CPU-credit exhaustion | Fault | Observation mode coverage; live enforcement via mysql-cpu-pressure |
| S9 | Azure OpenAI 429 from TPM pressure | Fault | Observation mode coverage; live enforcement via aoai-tpm-throttle |
| S10 | Application Gateway backend first-byte latency | Fault | Observation mode coverage; live enforcement via network-rtt-delay |
| S11 | Dependency outage cascade | Fault | Observation mode coverage; live enforcement via appgw-backend-failure |
| S12 | Bad deployment and rollout stall | Fault | Observation mode coverage; live enforcement via aks-bad-deploy |
| S13 | Knowledge ingestion and configuration drift | Non-fault | Scheduled governance and assurance seams |
| S14 | Alert-driven automatic investigation trigger | Non-fault | Webhook, event ingest, and IRP seams |
| C1 | Continuous-load baseline | Baseline | Observation mode calibration coverage |
| C2 | Pod kill under continuous load | Fault | Reuses aks-pod-kill |
| C3 | One-pod CPU hotspot under load | Fault | Reuses aks-pod-cpu-spike |
| C4 | One-pod memory hotspot and OOM kill | Fault | Reuses vm-mem-stress as pod memory stress |
The 2026-07-13 reference sweep recorded outcome=validated, detected=true,
and reverted=true for every row.
| Scenario ID | Expected signal | Probe class |
|---|
aks-pod-kill | pod_restart | Kubernetes event |
aks-pod-cpu-spike | node_cpu | Chaos Mesh status |
network-rtt-delay | gateway_latency | Chaos Mesh status |
aks-http-abort | request_failure | Chaos Mesh status |
vm-cpu-stress | host_cpu | Azure Monitor metric |
vm-mem-stress | host_memory | VM guest command |
mysql-cpu-pressure | db_cpu | Azure Monitor metric |
aoai-tpm-throttle | rate_limit | HTTP 429 sample |
appgw-backend-failure | backend_health | Kubernetes endpoints |
aks-bad-deploy | rollout_stall | Kubernetes pod status |
Event and status probes were observed in under 3.5 seconds in the latency
sweep, except VM memory pressure at about 31 seconds. The Azure Monitor CPU
measurements retain a documented cold-start measurement gap because the probe
queried a rolling aggregation window; do not treat their first-poll values as
true cold-start latency.
| Scenario ID | Domain | Expected tier | Expected decision | Expected action |
|---|
change.drift-manual-portal-edit.003 | Change | T0 | human approval | No |
change.nsg-allow-any-inbound.002 | Change | T0 | human approval | No |
change.tag-owner-missing.001 | Change | T0 | auto | Yes, observation mode delivery |
dr.backup-vault-restore-rehearsal.002 | DR | T0 | auto | Yes, observation mode delivery |
dr.chaos-experiment-novel.003 | DR | T2 | human approval | No |
dr.replica-lag-degraded.001 | DR | T1 | human approval | No |
finops.right-size-vm-high-monthly.002 | FinOps | T0 | human approval | No |
finops.stop-idle-dev-vm-off-hours.003 | FinOps | T1 | auto | Yes, observation mode delivery |
finops.unattached-public-ip.001 | FinOps | T0 | auto | Yes, observation mode delivery |
| Scenario | Expected result |
|---|
auto_rule_fired | auto |
hil_rule_fired | human approval |
deny_irreversible | deny with quorum two |
hil_unknown_event_triage | human approval because a concrete resource exists |
abstain_no_resource_target | hold for review because no actionable target exists |
rbac_denied_operator | deny and emit a security event |
| Scenario case | Invariant |
|---|
| Auto in observation mode | Judged and audited without mutation |
| human approval request | Exactly one pending approval ticket |
| Deny | Never reaches approval or execution |
| Mixed stream | Decision and dispatch counts remain equal |
| Duplicate delivery | One dispatch, one measured duplicate, no double execution |
| Empty or junk event | Holds for review and causes no downstream action |
| Self-approval attempt | Blocked and measured as a security signal |
| Repeated self-approval | Retry does not inflate the security count |
The catalog currently contains 93 entries that the default factory classifies
as executable and builds in dry-run. The remaining 42 are 17 AWS FIS
cross-CSP references, 21 GPU scenarios requiring an injector or hardware,
three Kubernetes documentation candidates, and the legacy Redis reboot
scenario. All 135 remain under collected/; zero are under promoted/.
Azure Chaos Studio - 15 IDs
chaos.azure-chaos-studio.agent-cpu-pressurechaos.azure-chaos-studio.agent-network-disconnectchaos.azure-chaos-studio.agent-network-latencychaos.azure-chaos-studio.agent-network-packet-losschaos.azure-chaos-studio.agent-physical-memory-pressurechaos.azure-chaos-studio.agent-stop-servicechaos.azure-chaos-studio.cosmos-db-failoverchaos.azure-chaos-studio.keyvault-deny-accesschaos.azure-chaos-studio.load-balancer-backend-removechaos.azure-chaos-studio.nsg-security-rulechaos.azure-chaos-studio.redis-rebootchaos.azure-chaos-studio.service-bus-firewall-blockchaos.azure-chaos-studio.vm-redeploychaos.azure-chaos-studio.vm-shutdownchaos.azure-chaos-studio.vmss-shutdown
Chaos Mesh - 14 IDs
chaos.chaos-mesh.block-delaychaos.chaos-mesh.container-killchaos.chaos-mesh.dns-errorchaos.chaos-mesh.http-delaychaos.chaos-mesh.http-replacechaos.chaos-mesh.io-faultchaos.chaos-mesh.kernel-panicchaos.chaos-mesh.network-bandwidthchaos.chaos-mesh.network-corruptchaos.chaos-mesh.network-duplicatechaos.chaos-mesh.network-losschaos.chaos-mesh.network-partitionchaos.chaos-mesh.pod-failurechaos.chaos-mesh.stress-memory
Litmus - 16 IDs
chaos.litmus.container-killchaos.litmus.disk-fillchaos.litmus.node-cpu-hogchaos.litmus.node-drainchaos.litmus.node-memory-hogchaos.litmus.pod-cpu-hogchaos.litmus.pod-deletechaos.litmus.pod-dns-errorchaos.litmus.pod-http-latencychaos.litmus.pod-http-status-codechaos.litmus.pod-io-stresschaos.litmus.pod-memory-hogchaos.litmus.pod-network-corruptionchaos.litmus.pod-network-duplicationchaos.litmus.pod-network-latencychaos.litmus.pod-network-loss
Kubernetes documentation - 3 IDs
chaos.kubernetes-docs.dns-resolution-failurechaos.kubernetes-docs.image-pull-backoffchaos.kubernetes-docs.pod-disruption-budget-gap
Synthesized general scenarios - 48 IDs
chaos.general.db-saturate-db-cpu-extremechaos.general.db-saturate-db-cpu-highchaos.general.db-saturate-db-cpu-mildchaos.general.disk-delay-host-cpu-extremechaos.general.disk-delay-host-cpu-highchaos.general.disk-delay-host-cpu-mildchaos.general.dns-delay-gateway-latency-extremechaos.general.dns-delay-gateway-latency-highchaos.general.dns-delay-gateway-latency-mildchaos.general.lb-deny-backend-health-extremechaos.general.lb-deny-backend-health-highchaos.general.lb-deny-backend-health-mildchaos.general.llm_endpoint-throttle-rate-limit-extremechaos.general.llm_endpoint-throttle-rate-limit-highchaos.general.llm_endpoint-throttle-rate-limit-mildchaos.general.pod-corrupt-rollout-stall-extremechaos.general.pod-corrupt-rollout-stall-highchaos.general.pod-corrupt-rollout-stall-mildchaos.general.pod-delay-gateway-latency-extreme-v2chaos.general.pod-delay-gateway-latency-extremechaos.general.pod-delay-gateway-latency-high-v2chaos.general.pod-delay-gateway-latency-highchaos.general.pod-delay-gateway-latency-mild-v2chaos.general.pod-delay-gateway-latency-mildchaos.general.pod-drop-request-failure-extreme-v2chaos.general.pod-drop-request-failure-extremechaos.general.pod-drop-request-failure-high-v2chaos.general.pod-drop-request-failure-highchaos.general.pod-drop-request-failure-mild-v2chaos.general.pod-drop-request-failure-mildchaos.general.pod-saturate-node-cpu-extreme-v2chaos.general.pod-saturate-node-cpu-extremechaos.general.pod-saturate-node-cpu-high-v2chaos.general.pod-saturate-node-cpu-highchaos.general.pod-saturate-node-cpu-mild-v2chaos.general.pod-saturate-node-cpu-mildchaos.general.pod-stop-pod-restart-extremechaos.general.pod-stop-pod-restart-highchaos.general.pod-stop-pod-restart-mildchaos.general.vm-saturate-host-cpu-extreme-v2chaos.general.vm-saturate-host-cpu-extremechaos.general.vm-saturate-host-cpu-high-v2chaos.general.vm-saturate-host-cpu-highchaos.general.vm-saturate-host-cpu-mild-v2chaos.general.vm-saturate-host-cpu-mildchaos.general.vm-saturate-host-memory-extremechaos.general.vm-saturate-host-memory-highchaos.general.vm-saturate-host-memory-mild
GPU and AI serving - 22 IDs
chaos.gpu.gpu-delay-gpu-pcie-degradationchaos.gpu.gpu-ecc_error-gpu-ecc-uncorrectable-v2chaos.gpu.gpu-ecc_error-gpu-ecc-uncorrectablechaos.gpu.gpu-hang-gpu-util-zero-wastedchaos.gpu.gpu-oom-gpu-vram-oom-v2chaos.gpu.gpu-oom-gpu-vram-oomchaos.gpu.gpu-quota_shrink-gpu-idle-hours-wastedchaos.gpu.gpu-quota_shrink-gpu-sku-mismatchchaos.gpu.gpu-thermal_throttle-gpu-temp-throttlechaos.gpu.gpu-throttle-gpu-power-throttlechaos.gpu.gpu-xid_event-gpu-xid-event-v2chaos.gpu.gpu-xid_event-gpu-xid-eventchaos.gpu.gpu_cluster-saturate-gpu-util-saturatedchaos.gpu.inference_endpoint-cache_overflow-kv-cache-pressurechaos.gpu.inference_endpoint-delay-inference-p99-spikechaos.gpu.inference_endpoint-delay-weights-fetch-stallchaos.gpu.inference_endpoint-ramp-cold-start-latency-spikechaos.gpu.llm_endpoint-quota_shrink-token-spend-spikechaos.gpu.training_job-checkpoint_fail-spot-preempt-cascadechaos.gpu.training_job-delay-distributed-stragglerchaos.gpu.training_job-hang-nccl-timeoutchaos.gpu.training_job-preempt-spot-preempt-cascade
AWS FIS cross-CSP reference - 17 IDs
chaos.aws-fis.ec2-reboot-instanceschaos.aws-fis.ec2-send-spot-instance-interruptionschaos.aws-fis.ec2-stop-instanceschaos.aws-fis.ec2-terminate-instanceschaos.aws-fis.ecs-stop-taskchaos.aws-fis.eks-pod-cpu-stresschaos.aws-fis.eks-pod-network-latencychaos.aws-fis.network-disrupt-connectivitychaos.aws-fis.rds-failover-db-clusterchaos.aws-fis.rds-reboot-db-instanceschaos.aws-fis.s3-bucket-pause-replicationchaos.aws-fis.ssm-cpu-stresschaos.aws-fis.ssm-disk-fillchaos.aws-fis.ssm-kill-processchaos.aws-fis.ssm-memory-stresschaos.aws-fis.ssm-network-latencychaos.aws-fis.ssm-network-packet-loss
- Catalog validated means schema, registered signal, unique ID, and symptom
index checks pass.
- Dispatchable means the default factory can construct the injector and
probe pair with a complete context. It does not mean a live fault ran.
- Observation mode covered means the detection, routing, RCA, response, and safety
mapping exists without mutation.
- Live enforcement validated means a disposable substrate recorded injection,
expected detection, and rollback for that scenario.
- Promoted would mean the scenario passed its independent promotion gate.
The current catalog has zero promoted entries.