Claim
Adding an explicit epistemic labeling step before synthesis — “classify each factual claim as: (a) directly stated in source, (b) inferred from source, (c) uncertain, (d) speculative” — reduces the rate of confident hallucination on verifiable facts.
Why Hallucination Is the Target
Hallucination is driven partly by the absence of any internal uncertainty signal in the generation pathway. Forcing the model to classify its own claims before synthesizing them creates a structured opportunity to surface that uncertainty before it gets flattened into confident prose. Per Kadavath, the information is there; H7 is a structured query for it.
Experimental Setup
- Feed source documents + questions with verifiable answers against those documents (faithfulness task) and against the world (factuality task — see Huang 2023 for the split)
- Condition A: synthesize directly
- Condition B: label each claim first (a/b/c/d), then synthesize with labels preserved as hedges in the final output
- Condition C: label each claim, but force the model to drop claims labeled (c) or (d) before synthesis (stricter variant)
What to Measure
- Rate of hallucinated claims (claims not supportable by source — faithfulness; claims false about the world — factuality)
- Rate of appropriate hedging on uncertain claims
- Meta-accuracy: whether “directly stated” labels are themselves accurate (does the model correctly identify what’s in the source vs. what it made up?)
Why It Could Fail
Models may label their hallucinations as “directly stated” — the labeling step is itself subject to hallucination. Meta-accuracy (knowing what you know) is not guaranteed. The Kadavath finding gives mechanistic hope but not certainty for this specific task framing.
Empirical Evidence
Mechanism confirmed; specific protocol untested.
| Source | Finding | Implication |
|---|---|---|
| Kadavath et al. (Anthropic, 2022) | Models can be prompted/trained to produce well-calibrated P(True) for their own outputs. The information is internally present. | H7’s mechanism is plausible — labeling extracts information that exists |
| Huang et al. (2023) | Canonical taxonomic split: faithfulness (claim unsupported by source) vs. factuality (claim wrong about world). RAG does not eliminate either. | H7 is most plausible for faithfulness (the model can check against provided source); factuality is harder because no ground reference |
| Tian et al. (2023) | Verbalized uncertainty prompts work — the simple intervention is “just ask” | Strong precedent that structured queries for uncertainty extract usable signal |
Practical implication: H7 likely succeeds on faithfulness (especially in condition C, where uncertain claims are dropped). Factuality improvements are less likely without external grounding. Two distinct tests should be reported separately.
Open: meta-accuracy of the labels themselves. A pre-registered measurement: for each “directly stated” label, check whether the claim is verbatim or near-verbatim in the source. This is the test of whether the labeling step is doing real work.
See Also
- Hallucination · Key Assumptions Check
- Testable Hypotheses framework
- Sibling: H6 (What If? + Overconfidence) — same Kadavath mechanism, different surface task
- Huang Hallucination Survey — for the faithfulness/factuality distinction