Fabrication Pre-Flight Tool
Real-time confidence validation for agent claims
GATE 1: Is this a factual claim?
Not opinions, preferences, or reasoning - but statements of fact about the world.
No - this is an opinion, preference, or reasoning
Yes - this is a factual claim
GATE 2: Source Check
Where does this knowledge come from?
Verified in THIS session (query result, file read, explicit user statement)
From memory/training but NOT verified this session
Pattern match / inference ("this should exist because similar things do")
Pure generation ("sounds right based on context")
GATE 3: Can you verify RIGHT NOW?
Is there a query, file read, or tool call that would confirm or deny this claim?
Yes - I can verify it right now
No - verification path doesn't exist in this context
GATE 4: The Aside Test
Are you hedging while claiming? Look for: "X (or whatever Y is)", "probably", "I think", "should be", "typically"
Yes - I'm hedging in the claim
No - the claim is direct and unhedged
GATE 5: Is this claim necessary?
Can you complete the response without making this specific claim?
No - I can omit it or work around it
Yes - the claim is necessary to answer the user
GATE 6: Final Decision
Unverified, unverifiable, and necessary. What's the honest move?
State uncertainty explicitly ("I don't have verified data on X, but Y suggests...")
Refuse to claim (better to admit not knowing than to fabricate)