How it works
Reason, gate, execute, sign.
One pipeline runs under every workflow. Here it is in full — the loop, three real use-case flows, and the guarantees that hold it together.
The loop
One path, from raw signal to signed verdict
Every USAP workflow runs the same pipeline. The agent reasons; the evidence gate and the human-approval step decide what passes; MCP executes; the audit log remembers. Watch the signal flow through.
Use cases
The same discipline, across real workflows
Three flows USAP runs end to end. Read paths flow freely; decisions gate; every outcome is recorded.
A SIEM alert becomes an evidence-backed verdict
The analyst persona classifies the alert, pulls the signal and repo context over MCP, and cites the exact tool-call that produced every finding.
Ransomware declared, contained, preserved
Classification declares SEV-1 and starts the regulatory clock. Containment waits for human approval; forensics runs in parallel so volatile evidence is never lost.
A malicious layer is caught before the image ships
Findings are classified by component — base image, app dependency, or unexpected layer. A CVSS gate blocks the deploy; an implanted layer escalates to incident command.
How it works
One path, from signal to signed verdict
Reason
A cs-* agent classifies the signal and pulls the evidence it needs over MCP — from whatever SIEM, cloud, or code host you actually connected.
Gate
The typed verdict passes the evidence gate; anything that mutates the world waits behind explicit human approval.
Execute & sign
On approval, MCP dispatches to the real tool. Every step lands in a tamper-evident, signed audit log.
The guarantees
Four rules that make it trustworthy
No source, no verdict
Every conclusion must cite evidence that resolves — a live tool-call, a URL, an in-repo file. Prose like "the logs showed it" is rejected at the contract boundary.
A human approves the world-changing part
Reads run autonomously. The moment an action mutates — isolate a host, block an IP, rotate a key — the payload halts for a person. Every skill declares it.
Numbers are computed, never invented
CVSS from the vector, EPSS from the FIRST feed, confidence from a written rubric. If it can't be computed from a canonical source, it's marked qualitative — not guessed.
Every step is signed
Verdict, approval, dispatch — each appended to a SHA-256 hash chain and HMAC-signed. An incident review can trust exactly what happened, in order.
12 domains · 80 skills · 12 orchestrator agents