12 investigations
Every row below links to the full step trace: what the model asked for, what it was shown, the diagnosis it wrote, and every evidence excerpt resolving back to the untruncated tool output it came from.
- graded correct
- 9 of 10 replayed
- acted on
- 1
- refused by policy
- 1
- total cost
- $0.06
- investigating model
- gpt-oss:120b
- tool traces from
- gemini-3.7-flash
| state | sev | kind | target | incident | steps | cost | graded |
|---|---|---|---|---|---|---|---|
| ^ escalated | Critical | CrashLoopBackOff | hephaisto-chaos/c13-wedged-lock-6778bccbd9-djhrb | CrashLoopBackOff on c13-wedged-lock (hephaisto-chaos) The container's startup script checks for a lock file at /scratch/startup.lock and aborts if it exists; the lo… | 8 | $0.0043 | Correct |
| + resolved | Critical | ReadinessFlapping | hephaisto-chaos/c13-wedged-lock-6778bccbd9-r24gx | ReadinessFlapping on c13-wedged-lock (hephaisto-chaos) The container's startup script checks for a lock file at /scratch/startup.lock and exits with error if it exis… | 8 | $0.0035 | Correct |
| ^ escalated | Critical | CrashLoopBackOff | hephaisto-chaos/c1-oomkill-6c8dd474dd-48sz5 | CrashLoopBackOff on c1-oomkill (hephaisto-chaos) The container "balloon" is being OOMKilled during init because its memory limit (64Mi) is too low for the proc… | 7 | $0.0026 | Correct |
| ^ escalated | Critical | CrashLoopBackOff | hephaisto-chaos/c2-crashloop-56b6585fd9-4hj9x | CrashLoopBackOff on c2-crashloop (hephaisto-chaos) The container exits with code 1 because its startup script cannot connect to the MongoDB service (mongo.infra-… | 5 | $0.0019 | Correct |
| ^ escalated | Warning | Unschedulable | hephaisto-chaos/c3-unschedulable-fd55d4b7c-tpkg7 | Unschedulable on c3-unschedulable (hephaisto-chaos) The Deployment c3-unschedulable requests 500Gi of memory, which exceeds the node's allocatable memory, causing… | 9 | $0.0034 | Correct |
| ^ escalated | Warning | ImagePullBackOff | hephaisto-chaos/c4-imagepull-7fd88fdb68-5n8sf | ImagePullBackOff on c4-imagepull (hephaisto-chaos) Deployment hephaisto-chaos/c4-imagepull references a non-existent image tag busybox:this-tag-does-not-exist, c… | 6 | $0.0024 | Correct |
| ^ escalated | Warning | JobFailed | hephaisto-chaos/c5-badjob | JobFailed on c5-badjob (hephaisto-chaos) The Job's container exits with a non‑zero status (exit code 3) causing the pod to terminate with reason Error;… | 22 | $0.0095 | Correct |
| ^ escalated | Warning | ConfigError | hephaisto-chaos/c7-configerror-7cd96c86f9-4dlvl | ConfigError on c7-configerror-7cd96c86f9-4dlvl (hephaisto-chaos) The pod c7-configerror cannot start because the referenced Secret "c7-database-credentials" does not exist in … | 5 | $0.0016 | Correct |
| ^ escalated | Warning | ReadinessFlapping | hephaisto-chaos/c8-readiness-flap-dd589d77c-6mxdb | ReadinessFlapping on c8-readiness-flap (hephaisto-chaos) The container's /healthz endpoint intentionally alternates between HTTP 200 and 404 each minute, causing the c… | 13 | $0.0055 | Correct |
| ^ escalated | Warning | HighErrorRate | /faulty-service | HighErrorRate on faulty-service () No finding. | 22 | $0.0069 | NoFinding |
| ^ escalated | Critical | CrashLoopBackOff | hephaisto-chaos/c11-transient-6f59fc7fcc-27h6g | CrashLoopBackOff on c11-transient (hephaisto-chaos) The container exits with code 1 because its startup script detects a stale lock file (generation 1) on the att… | 12 | $0.0052 | Correct |
| ^ escalated | Warning | CrashLoopBackOff | hephaisto-chaos/c12-stale-lease-5b894c6649-k22gg | CrashLoopBackOff on c12-stale-lease-5b894c6649-k22gg (hephaisto-chaos) The container exits because its startup script detects a stale lease file on the persisted volume that claims … | 12 | $0.01 | Correct |
What this shows, and what it does not
1 of these 12 shows the agent acting, and 1 shows it refused by policy. Those were exported from the agent's own database after a real run, which is the only way this page can show either: a replay serves a recorded tool trace to a live model and constructs no executor, no policy engine and no state machine, so it has nothing to act with and nothing to be refused by.
None of the 10 replayed investigations shows the agent acting.
7 correctly declined to propose an action, one produced no plan at all, and in 2 the grader judged an action was missed.
That is a measured property of the model these were replayed against, not a limitation of this
page — on that fixture gpt-oss:120b proposed an action in 1 of 11 runs where the
planner actually ran, where deepseek-v4-flash proposed one in 4 of 8.
So the honest version is: the replayed corpus is what the agent's diagnosis looks like, and whether it acts is a separate question measured separately, on a separate fixture. The evidence page has the denominators.