Fact-Checking and Validation Scaled to Stakes
CoreApply fact-checking and validation techniques scaled to stakes · Difficulty 2/5
Explanation
Validation Is a Deliberate Step, Not a Byproduct
Validation is the deliberate step of confirming an output before it's used -- it does not happen automatically because an output reads well, and it is not the same activity as improving how the output looks. The right validation technique scales with the stakes of what's being confirmed:
| Situation | Validation approach |
|---|---|
| Any factual claim bound for others | Verify names, numbers, dates, and citations against an authoritative source |
| Cited regulation / policy / legal text | Open the actual source text and confirm the citation and wording |
| Numeric analysis | Recompute or spot-check totals; confirm figures match the source data |
| Summaries of a document you provided | Confirm the summary reflects the source and adds nothing not in it |
| Anything customer-, legal-, or compliance-facing | Human review before it leaves the building |
The Canonical Exam Scenario
Claude confidently summarizes a new regulation and cites a specific subsection. The correct action before sending it to compliance is to verify that subsection against the official regulation text -- not to send it because Claude sounded sure, and not to merely reword it to sound more formal. Confidence is not a verification substitute, and a sensitive-sounding tone does not certify a citation's accuracy.
Reformatting Is Not Validation
Making an answer look more polished, more formal, or better organized changes its presentation, not its truth value. Treating a reformatting pass as if it were a verification pass is a category error: one step touches wording, the other touches facts.
Common exam traps
- Substituting reformatting for validation. Making an answer look more formal does not make it correct.
- Skipping verification because "the analysis is just internal" -- accuracy still matters, and sensitive contexts still need diligence regardless of the audience's size.
Key Takeaways
- Validation is a deliberate confirmation step, distinct from improving an output's polish or formatting
- Match the validation technique to the claim type: source verification for citations, recomputation for numbers, source-fidelity check for summaries, human review for customer/legal/compliance output
- The canonical exam answer: verify a cited regulation subsection against the official text before sharing it, regardless of Claude's confidence
- 'It's just internal' is not a reason to skip verification -- stakes, not audience size, determine diligence
Related Concepts
Inconsistencies, Bias, and Hallucination-Reduction Techniques
Hallucination (fabrication), inconsistency (internal contradiction), and bias (skewed framing) are three distinct failure types requiring different fixes
Quote-Grounding and Best-of-N Comparison
Quote-grounding: for long documents, have Claude extract supporting quotes before analyzing, so conclusions trace back to a specific line rather than a trust-me summary
When Human Review Is Required
Escalate to human review when output is high-stakes, external-facing, contains unverifiable claims, involves sensitive/regulated data, or exceeds Associate scope