Skip to main content

View on GitHub

examples/02_compounding_agents/
Two failure modes an honest trust layer has to catch, shown end to end: a finding whose analysis step never ran, and two same-model agents whose agreement only looks like independent corroboration.

​
Run

No API key and no network required. The model calls answer from an in-memory httpx transport, so nothing leaves the process. That is also why no model call is certified here.

​
1. Absence: a number with no observed execution

An agent reports a result, but the step that reads the data never ran. The observer watched the whole scope and saw no cited read, so the finding cannot earn GROUNDED.
The bearing still computes from the pre-registered rule, but the grounding axis reports UNGROUNDED: no observed read stands behind the number, so the provenance edge is empty. The contrast is the same analysis actually executed, where the cited dataset is read inside the scope and the axis reads GROUNDED.

​
2. Same-model convergence does not count twice

agent-a runs the first grounded check. A second agent, distinct key and distinct dataset, reaches the same answer, but on the same model. The instrument does not read that as a second independent line.
Distinct in every legacy axis, signer and dataset, but one model, and that model was never observed. Two same-model checks would be one line of evidence; here the instrument cannot certify even that, so it reports UNVERIFIABLE.

​
3. A different model name, still nothing to count

A genuinely different model would be a second line, but a name in a request body the script answered itself proves nothing. The axis that matters is the model, not the key that signed it, and it is counted only when the call is observed against a recognized provider host.

​
Where the numbers come from

The model is read off the request body at the POST boundary. That capture earns COMPUTED only when httpx’s own network transport answered a recognized provider host; an in-process transport is a producer declaration (PROXY), and one soft supporting lineage is enough to make the count UNVERIFIABLE. The grounding verdict comes from watching the scope for a read of the cited source. Both ride into the signed finding through grounding=, and the trust map places each on its own axis: grounding, and effective independence. Neither is a label the agent chose. See Findings and Grounding.