Intelligence
Asynchronous verification of content-producing actions against declared intent.
What Intelligence does
Intelligence listens passively to the Trace receipt stream. It never interrupts the agent. After actions complete, it verifies that what was done matches what was declared.
Verified action types:
- Git commits and pushes -- commit message vs actual diff
- Network requests -- declared purpose vs actual endpoint and payload
- File writes -- declared content vs what was written
Non-blocking by design
Results appear in the dashboard. Intelligence does not block or modify agent behaviour.
How it works
Intelligence subscribes to the receipt bus. When Trace emits a receipt for a content-producing action, Intelligence retrieves the supporting evidence (file content, request payload, commit diff) and runs a verification pass. The result is stored alongside the receipt.
Step 1
Receipt emitted
Trace intercepts an action and writes a signed receipt to the receipt bus.
Step 2
Evidence retrieved
Intelligence fetches the intercepted payload -- the diff, file, or request body.
Step 3
Verification stored
The result is saved alongside the receipt and surfaces in the dashboard.
Evidence-first
Verification is grounded in what was intercepted, not what the agent claimed.
Configuration
Add the following block to transient.config.json:
{
"intelligence": {
"enabled": true,
"api_key": "your-key"
}
}Requires Trace
Intelligence requires Trace to be running. It is not a standalone tool.
Limitations
Intelligence only verifies actions that Trace intercepts. The following are outside its coverage:
Native Python network calls
Python's urllib, httpx, and requests bypass the shim layer unless the subprocess hook is active.
Node.js absolute-path processes
Processes invoked via an absolute path may not be caught by the PATH shim.
Anything outside Trace's interception scope
Actions that do not produce a Trace receipt cannot be verified by Intelligence.