contextjet-ai/awesome-llm-observability/check-answer-consistency

Check answer consistency

A model that knows the answer repeats it across re-samples; a hallucinating model wanders. This skill ships a small consistency scorer that turns that idea into a number, with no reference answer required.

Use the bundled script

`scripts/self_consistency.py` is pure Python, no install needed:

from self_consistency import consistency_score, is_likely_hallucination
answers = [call_model(prompt, temperature=0.7) for _ in range(5)]
score = consistency_score(answers)              # 1.0 = all agree, ~0 = all differ
if is_likely_hallucination(answers, threshold=0.5):
    route_to_review()

Run it directly to see confident vs unsure examples: python scripts/self_consistency.py.

How to apply it

1. Sample with temperature > 0 (e.g. 5 samples). Consistency methods need diversity; a single deterministic answer tells you nothing.
2. Score, then act: flag/block/route-to-review when the score is below your threshold, or attach it to the trace as a reliability signal.
3. Reserve it for high-stakes answers if latency matters, sampling N times multiplies cost and latency by ~N.

This is a lightweight SelfCheckGPT-style check (normalized-string agreement). For meaning-level robustness (wording differs but the fact is the same), move to entailment clustering / semantic entropy, see the detect-hallucinations skill.

Validation

Run the tests: pytest skills/check-answer-consistency/tests/. They confirm identical answers score 1.0 (ignoring case/punctuation/whitespace), all-different answers score low and flag, a 3-of-4 majority scores 0.75 and does not flag, and that fewer than two answers returns 0.0.

Grounding

Consistency-based hallucination detection: SelfCheckGPT, Manakul et al. 2023 (arXiv:2303.08896). Meaning-level variant: semantic entropy, Farquhar et al., Nature 2024.

Anti-patterns

• Running it at temperature 0 (no diversity, the score is meaningless).
• Treating a high consistency score as proof of correctness (a model can be confidently and consistently wrong).
• Sampling N times inline on every request when latency/cost matter (sample offline or only on risky answers).