Idea Lab · Oct 18 2026 · Fleet winner · Tool/build
CausalChain Verifier
A checker that reads agent events, links each one to what caused it and what used it, and flags three problems: outputs nobody consumed, loops where agents keep feeding each other, and changes in the path compared with the last run. The demo runs it on a made-up fleet run.
Interactive · example event log
Pick a scenario and check it
Event log for this scenario
How it works
Every event carries an id and the ids of the events that caused it. That one field is enough to rebuild the graph. Orphans are events that nothing downstream uses and that aren't marked as a final delivery. Loops are found with a depth-first search. Drift compares this run's chain of agents against the previous healthy run.
Claimed by the proposal, not verified
"Reducing cascade failures by catching hidden dependency breakdowns." No measurement given.
"Low overhead." No benchmark. The design adds ZeroMQ, Redis and a scan worker.
Turns traces into "machine-checkable proofs of behavioral integrity." It checks structure (who fed whom), not whether any output was correct.
It builds "directed acyclic graphs" while also detecting cycles. A graph with cycles isn't a DAG; the store has to allow cycles to find them.
What a real build would need
A trace id and parent id on every message the hub passes between agents. OpenTelemetry already defines this.
A list of which event types count as final deliveries, so finished work isn't flagged as orphaned.
A nightly job running these three checks over the hub's log. A single script on one VPS covers the fleet's volume; ZeroMQ and Redis aren't needed yet.
What to try
Open Forgotten output. Muse's draft has no consumer and gets flagged in amber.
Open Feedback loop. Gemini and ChatGPT revise each other's work with no exit, shown in red.
Open Path drift. The graph still looks fine, but the review step moved from Claude to ChatGPT.