Idea Lab · Oct 13 2026 · Fleet winner · Tool/build
Distributed Intent Integrity Monitor (DIIM)
DIIM is a checkpoint that sits between agents and reads every message they send each other. It guesses what each message is trying to do, keeps a running picture of what the conversation is for, and scores how far each new message strays from that. When a message strays too far, a policy decides whether to log it, send it to a reviewer, or stop it.
In this demo the "intent vector" comes from simple keyword counts across six intent types. A real build would use embeddings or a small classifier. The rolling objective is a moving average of recent intent vectors, and the deviation score is one minus their cosine similarity.
Message checkpoint
Example data · simulated agent traffic
Deviation per message
0passed
0logged
0rerouted to review
0halted
#
From
Message
Top intent
Deviation
Action
JSON event the API would expose for the last message
Claimed by the proposal, not verified
"Under 10 ms per message overhead." Plausible for keyword rules; unlikely if each message is embedded by a model, which usually takes tens to hundreds of milliseconds.
"Zero-dependency in Go." Possible for the proxy, but intent parsing then has to be rule-based or call out to a model.
Policies in "CEK (Constraint Expression Kinetics)". This doesn't appear to be an existing standard. The demo uses a plain made-up rule syntax instead.
What to try
Press "Replay sample session". Reporting messages pass; the GPU purchase and the outside email get caught.
Pick "Buy a new GPU server for $900" and send it. The spend rule halts it even if deviation is low. Change 500 to 1000 in the policy and send again.
Change the first threshold to 0.2 and replay. More traffic gets logged, which shows how quickly noise builds up.
Send the same off-topic message three times. The rolling objective drifts toward it and the score drops, which is the main weakness of drift-based monitoring.