Trace Feedback
Foundry traces become learning input for Alloy. Foundry answers what happened — which blocks ran, which events were emitted, which gates passed, which retries occurred, and what payloads moved through the event chain. Alloy adds an intent-interpretation layer on top of that raw record, asking what the run meant for the accepted intent graph.
This section sits under Alloy Documentation and feeds the Learning Loop that closes the gap between intent and running code.
Trace-interpretation questions
After a Foundry run, Alloy asks:
- Which intent records were in scope?
- Which capabilities were preserved?
- Which capabilities were harmed?
- Which threats were observed or appeared?
- Which strategies were applied, and which worked?
- Which strategies proved too expensive?
- Which tradeoffs became visible?
- Which expected evidence was collected, and which remained unknown?
- Which evidence passed or failed?
- Did retry behaviour stay within the intended scope?
- Did the run reveal missing intent?
- Did the run contradict accepted intent?
- Did the agent need guidance that Alloy did not provide?
- Which intent records need refinement?
- Which human questions should be asked next?
These questions read against the Foundry Run Result and the Evidence and Gates that the run emitted.
Feedback outputs
From those questions, Alloy can produce feedback records such as:
- Capability satisfaction status
- Threat observation
- Evidence observation
- Intent drift event
- Contradicted expectation
- Strategy effectiveness note
- Tradeoff observation
- Human clarification question
- Proposed new intent record
- Proposed deprecation of an old intent record
Where this goes next
- Learning Loop — the full closed loop from intent records back to an updated intent graph, with the
run_feedbackoutput. - Drift Detection — when the codebase no longer reflects accepted intent.
- Contradiction Detection — when the code is fine but the intent itself is stale.
Source: Product Brief §19; Integration Architecture §16.1–16.2.