The control pattern
Data asset
Grounding evidence and claim verification
Knobs
Source quality, Required citations, Retrieval threshold, Claim materiality, Abstention policy, Verifier strictness
Outcome
Fewer unsupported or misleading claims
Measurement
See how to measure this below
What it is
Hallucination control is the disciplined management of what the model is allowed to claim. The evidence is the data asset; source authority, citation rules, retrieval thresholds and abstention behaviour are its knobs.
It helps to distinguish several failure types, because each needs a different control: an invented fact, a real fact attached to the wrong period, a correct number with the wrong units, an interpretation presented as certainty, and a statement based on stale data. Requiring any citation is not enough if the citation does not actually support the claim.
The goal is not a timid assistant. It is one whose confidence reflects its evidence : able to say what is known, what is inferred, and what remains uncertain.
The knobs in detail
Each row is one adjustable property of the data asset, and what moving it tends to do.
| Knob | What you adjust | Likely effect |
|---|---|---|
| Source quality | Which sources may support which claims | Stops a news recap outranking a filing |
| Required citations | What level of claim needs support | Makes verification possible at all |
| Retrieval threshold | Minimum evidence before answering | Converts weak grounding into abstention |
| Claim materiality | Which claims get the strictest checks | Focuses effort on numbers and time-sensitive facts |
| Abstention policy | What the system does with no support | Prevents gap-filling from memory |
| Verifier strictness | How hard entailment is enforced | Trades false refusals against false claims |
Applied: Stocks Assistant
In Stocks Assistant, numeric and time-sensitive statements should pass through a claim-evidence check. A statement such as “revenue grew 12%” should be decomposed into company, metric, value, comparison period, units and source. If these elements cannot be aligned to evidence, the result is qualified or withheld. Generated summaries should clearly label facts, model interpretation and market-derived signals. A confidence threshold can decide whether a small model answers, a larger teacher rechecks the evidence, or the interface asks the user to narrow the question.
How to measure it
Evidence that the knob produced the intended behaviour, rather than shifting the problem elsewhere.
- Unsupported-claim rate at the claim level, not per answer
- Numeric contradiction rate and wrong-period rate
- Citation entailment : does the source actually support the claim?
- Freshness violations
- Appropriate abstention rate
Common mistakes
Grading a whole answer pass or fail. One bad claim in six is invisible at that resolution.
Treating a present citation as a verified citation.