Hallucination Detection Catching fabricated facts, invented actions and off-brand promises in model output: the cheapest check for each kind, an online claim guard at current prices, and the precision, recall and sample design that show whether it works.
Last updated 2026-10-08
After this section you can
✓ Choose the cheapest check that exposes each kind of bad answer ✓ Size a review queue from recall, false-alarm rate and base rate ✓ Design a labeled sample that measures recall for a rare failure 10
Hallucination Detection
Catch fabricated facts, invented actions and off-brand promises before they ship.
Key idea
Check each bad answer against its own reference, cheapest check first. How many flags are real depends on how rare bad answers are.
Each layer only sees what the cheaper one could not clear
LAYER
WHAT REACHES IT
WHAT IT SETTLES
COST PER CHECK, RISING
10,000 answers
a day
1
Code checks
gates, inline
numbers, citation ids
and the tool-call ledger
no model call
runs in your code
up to 10,000
2
Claim check
gates, inline
a verdict per claim: supported,
contradicted, not in the context
$0.000185
an answer, two calls
284 flagged
at a 3% base rate
3
Stronger judge
gates, flagged only
one Sonnet 5.5 call
on flagged answers only
$0.0023
a call; $0.65 a day at 3%
the judge’s flags
4
Human queue
holds or audits
a person decides
what no model could clear
2 minutes
a case, reviewer time
What a layer clears ships; only what it cannot clear drops to the next, dearer one. An audit sample runs beside all four and gates nothing.
Get full access to all 149 sections with code examples, diagrams, and interactive animations.
Unlock Premium