Reliability evidence // fault injection

Claims under failure.

The harness activates a bounded fault lease, observes the target surface, verifies notification evidence, removes the lease, and proves recovery. Scheduled runs are simulation-only. Live experiments require one explicit target, a protected GitHub environment, a short-lived token, and a Worker-side enable flag.

latest verdictloading
experiments-
evidence time-

contacting public evidence API

Experiment results

Closed-loop contract

  1. InjectA target-specific control lease records the fault, duration, experiment id, and automatic expiry.
  2. ObserveThe probe must witness the declared failure shape within the detection budget.
  3. NotifyThe experiment id must appear in the public recent-event evidence, unless the declared fault is the notification path itself.
  4. RollbackThe lease is deleted in a finally path even when observation fails.
  5. RecoverThe healthy surface must return within the recovery budget before the report can pass.

Scheduled safety

The Wednesday schedule runs only the deterministic simulator. It proves contracts, report generation, and the evidence publishing path without touching a live route.

Live safety

A live dispatch requires mode=live, one experiment id, approval in production-chaos, ATLAS_CHAOS_TOKEN, and CHAOS_ENABLED=true on the target.

Evidence honesty

Simulation and live runs carry different modes in the report. A simulated pass is useful assurance, but it is never displayed as proof that production alerting survived a real outage.

Provenance

Mode
-
Source repository
-
Source commit
-
Fingerprint
-