Use a model-generated explanation without treating it as proof

A fluent explanation can help a reviewer understand a proposed answer. It does not establish that the inputs, arithmetic, citation or causal story are correct. Verification must come from evidence and independent rules.

Recalculate the claim

The offline foundations lab uses an authored explanation and three values.

from foundation_cases import explanation_case

result = explanation_case()
assert result["claimed_total"] == 42
assert result["independent_total"] == 42
assert result["verified"] is True
assert result["explanation_used_as_proof"] is False

The statement says 12, 13 and 17 sum to 42. Python independently confirms the total. If the prose sounded equally confident while claiming 43, the rule should reject it.

Match verification to claim type

For arithmetic, recompute from source values. For extraction, locate the exact span and compare field by field. For a citation, confirm the source supports the nearby statement and is the approved version. For code, run tests in the intended environment. For causal explanations, require experimental or domain evidence; a plausible narrative is only a hypothesis.

Do not ask the same model to “double-check” and treat agreement as independence. Use deterministic logic, a separately designed evaluator or qualified human review according to risk. Model graders can help scale evaluation after calibration, but preserve disagreements and critical invariants.

Design the interface honestly

Label generated rationale as explanation, not audit evidence. Link verified fields to their sources. Show uncertainty or abstention where checks cannot establish the answer. Keep raw generation separate from validated application state so downstream actions consume only accepted values.

The Generative & Agentic AI course teaches explanations beside source support, deterministic checks and human review rather than as self-authenticating outputs.

Exercise

Collect ten generated explanations covering arithmetic, extraction and recommendations. Write an independent verifier for objective claims, identify which statements remain judgments and create an interface that distinguishes verified from unverified content.

Continue learning

This article is part of the LLM fundamentals and prompt design sequence. Use the neighbouring tasks when you need the prerequisite or the next application.

• Review the prerequisite or neighbouring task in Prompt injection versus an ordinary user instruction.
• Continue with Define a prompt versioning and review convention.

Reference: NIST AI RMF playbook.