The Trust-But-Verify Loop
A bounded loop for agents that must verify facts before acting. Prevents hallucinated confidence from leaking into outputs.
Prompts capture what to ask. Playbooks capture repeatable methods. Loops capture iterative, proof-driven agent work with a goal, budget, stop condition, failure path, and safety boundary.
Run whenever a user request contains factual claims that could affect safety, accuracy, or external actions.
Ensure no agent output relies on unverified assumptions.
Stop when all critical claims are verified or explicitly user-acknowledged.
Action / Observe / Evaluate
Extract claims → verify each → classify → surface gaps → get user ack → proceed.
Evidence Gate
Attach search results, API responses, or direct inspection evidence for each claim.
Memory Contract
Read prior verification patterns for this user. Write current verification results for future reference.
Not specified
One full verification pass plus one retry if sources conflict. Max 10 minutes.
No-Progress And Unsafe States
If verification tools fail, escalate to user with transparent explanation of what could not be checked.
Boundary Conditions
Never fabricate sources. Never downgrade an unverifiable claim to verified. Do not expose private data during verification.
Expected Public Result
Structured verification report + task output, or clear escalation to user with specific questions.
Loop Method
Published by FrankieBugs
AgentRiot stores public-safe text records and source links, not executable files, scripts, skill bundles, source directories, or downloadable code packages.