Actual source-fault evidence
Each JSON file records the fault that actually executed, separately from exact
native/queue/Sentinel delivery reconciliation. The capture script is retained
byte-for-byte as capture-sourcefault.sh; its hash was verified against every
before/after SSM command receipt.
The script is read only. Invoke it through lab/aws/monitoring/run.py so account,
exact instance ID, current RunId/Project/Role tags, and running state are checked
before each command. The public environment values are:
| Name | Value |
|---|---|
BOWER_SCENARIO | The frozen scenario name |
BOWER_FAULT_UNIT | Exact bower-lab-f-* unit from that plan |
BOWER_FAULT_ACTION | The plan's restart, kill, blackhole, reset, reboot or stop-start action |
BOWER_CAPTURE_MODE | before or after |
BOWER_SCOPE_START_EPOCH | Frozen plan start minus two seconds |
BOWER_SCOPE_END_EPOCH | Expected collection bound before dispatch; actual observation bound afterward |
Capture the baseline before scheduling the fault. Capture again after the planned workload and fault complete. Keep outputs and command receipts in new, private files. The script projects timestamps, fault identity, process IDs/start ticks, boot identity, packet counters and helper hash. It never emits arbitrary journal messages, bodies or environment values. It retains no raw journal dump.
Qualification checks bind both captures to the same owned instance and installed helper hash. A restart, kill or stop-start must change the relevant process identity. Kill additionally names the original PID. Blackhole/reset must show a positive affected-packet count and the requested elapsed duration. Reboot must change boot identity and have an in-window reboot marker. The relevant collector or sender service must be active before and after the fault. Fault start must be within 15 seconds of its frozen scheduled time. Missing evidence fails these checks; a scheduling receipt alone never passes.
The final JSON binds the frozen plan and original capture/command receipts by
SHA-256. It carries deliveryQualified=false because fault execution alone says
nothing about destination delivery. summary.json links each actual fault proof
to the separately reconciled native and Sentinel identity sets. Reproduce those
file checks with the verifier in ../identity-evidence/README.md.