Transaction Reliability
Race duplicate work, then enforce the invariant that only one state change wins.
OPEN EXPERIMENT →OFF THE MAP // ENGINEERING UNDER CONTROLLED CONDITIONS
Build it. Break it.
Prove what changed.
Interactive demonstrations of real engineering failure modes, rebuilt from scratch with synthetic data. The controls below are simulations of the repository's tested invariants, not production systems.
THE LAB BENCH
Choose a failure boundary.
AI Workflow Boundaries
Keep hard constraints and catalog facts deterministic around flexible ranking.
OPEN EXPERIMENT →Email Delivery & Verification
Trace the gap between “generated,” “accepted,” and actually delivered.
EXPERIMENT NOT YET BUILTLAB 01 // COMPLETE
The duplicate transaction
Two synthetic workers receive the same request. Compare the unsafe check-then-write race with an idempotent boundary.
WORKER AREADY
STATE
BOUNDARY
BOUNDARY
WORKER BREADY
OBSERVATIONSelect a mode and run the synthetic race.
No result is shown until you run the demonstration.
LAB 02 // ACTIVE
Keep AI inside the rails
Hard eligibility happens before the flexible step. Authoritative facts are checked again afterward. Try to inject an ineligible item and watch the boundary reject it.
SYNTHETIC
CATALOG
→CATALOG
DETERMINISTIC
FILTER
→FILTER
FLEXIBLE
RANKING
→RANKING
FACT
VALIDATION
VALIDATION
BOUNDARY TRACEWaiting for input.
The demo uses only the synthetic catalog embedded in this page.
LAB RULE 00
If it hasn't been measured, it isn't presented as a result.
This interface visualizes the repository's engineering ideas. Automated tests in GitHub remain the evidence for implemented behavior.