Intelligence
Agent Mesh
Nine agents, fourteen handoffs, one hard stop. Each node is scoped to a single constraint and publishes into a shared identity graph rather than a chat window. Click a node to see what it is allowed to do — and what it is not.
Orchestration map
Edges are real data handoffs. When a run is in flight the active edge carries a packet and the receiving node pulses.
build buy partner
The mesh terminates at a human node on purpose. Agency accounts authenticate through a federal identity service whose terms prohibit automated submission, so assembly is the last machine step in the chain.
Node detail
Capability, constraint, benchmark and guardrail.
DEF-05Deficiency Predictoronline
Deficiency prediction before submission
Queue
19
Throughput / day
260
sustained, not peak
The constraint it attacks
183 of 283 federal permit days sit on the operator side
Benchmark
likely83.7% accuracy / 0.891 AUC
Nearest published analogue
197 building-code waiver decisions, preprint, a different regulator. But 55% of prediction variance sat on documentation completeness and technical justification quality, and risk tiering separated approval rates from 90.3% down to 10.0% — the same shape as the agency’s own operator/agency split.
Sourcing verdict
Build — the differentiating asset is proprietary context, not the model.
Consumes from
Identity Graph
Obligation Reasoner
Citation Harness
Feeds
Package Assembler
Run console
No run dispatched yet.
Start a run and every step, handoff and gate shows up here in sequence.
Run history
Every dispatch in this session, newest first.
No runs yet. Dispatch the end-to-end scenario or run a single agent.
Sourcing posture across the mesh
What is worth building, what is already a commodity, and where the constraint sits outside the enterprise.