Training Lab
Feed data batch by batch and watch the fleet's memory form - measured, not asserted
idleconnecting
This does not fine-tune model weights - Claude Haiku on Bedrock is not trainable here. What forms is episodic memory: every closed case is summarised, embedded, and consolidated with similar ones. Each batch below is a measured step, and every number traces back to the append-only audit log.
Active memories
-
Raw cases absorbed
-
In flight
-
Verdicts
-
Accuracy
-
Ingest training data
generate new labelled transactions and queue them for the fleetNew cases carry a ground-truth label and land straight in the queue. They are scored synthetically inside the investigate band - this feeds the agent and its memory, it does not exercise the XGBoost scorer.
Agent pipeline
what the fleet is doing right now, stage by stageClaimedSKIP LOCKED
Memory recallvector top-k
Customer contextMCP read-only
ReasoningClaude Haiku
Verdictfraud / legit / escalate
last 0 audited eventsavg recalled per case: 0.00avg reasoning: 0 ms
Training run
each batch: feed cases → drain the queue → measureMemory formation
active memories and raw cases absorbed, per batchNo run yetStart a run: the first batch fills an empty memory, later batches show what recall does as it grows.
Decision mix
auto-resolved vs escalated, per batchNo run yetEach batch reports how many cases the fleet closed on its own.
Batch results
every row is a measured step of the runNo batches recordedA run writes one row per batch: cases closed, auto-resolve rate, accuracy, memory growth and cost.
Saved runs
every completed run is stored in CockroachDB for comparisonNo saved runs yetA run is saved automatically when it finishes.
Live agent activity
the append-only audit trail, newest first0 events
No recent activityStart a run, or feed cases from Mission Control.