Two-Pass LLM Hallucination Audit: Catching Invented Metrics in AI Resumes
Ads
When prompted to optimize a resume for an applicant tracking system, large language models aim to produce impactful, quantified achievements. Because they prioritize expected patterns over factuality, they frequently invent percentages, team sizes, and budget figures.
Submitting fabricated numbers can cause immediate failure during background checks. A two-pass audit separates the initial rewrite from a dedicated verification step to catch these inventions.
Why Single-Turn Resume Prompts Corrupt Facts
In a single prompt, a model attempts two conflicting tasks: writing punchy resume copy and adhering strictly to your source notes. When the original text lacks metrics, the model often generates placeholder statistics or imports standard industry figures.
Asking "Is this accurate?" in that same chat rarely catches the problem, as the model defends its own context. Fact-checking requires a separate evaluator in a clean session.
The Two-Pass Hallucination Audit Workflow
This workflow clears the conversational history and runs an adversarial check between your raw career notes and the revised copy.
The Cross-Examination Prompt Template
Paste this prompt into a clean session to run the audit:
SYSTEM: You are a forensic resume auditor. Your sole task is to identify factual deviations, unearned metrics, and hallucinated claims in rewritten resume text.
INPUT 1: [RAW_UNEDITED_CAREER_NOTES]
[Paste your original rough notes or old resume here]
INPUT 2: [PROPOSED_AI_REWRITE]
[Paste the generated draft here]
TASK:
1. Compare every single bullet point in INPUT 2 against INPUT 1.
2. Extract any number, percentage, dollar value, team size, or operational outcome in INPUT 2 that does not have explicit, verbatim support in INPUT 1.
3. Flag any skill, tool, or credential mentioned in INPUT 2 that was never stated in INPUT 1.
4. For every discrepancy found, output the offending sentence, cite the specific hallucinated detail, and provide a corrected version that uses qualitative impact language without invented figures.
Interpreting and Acting on the Audit Log
The auditor returns each unsupported claim. Review them line by line. If a flagged figure is accurate but was omitted from your rough notes, add it to your source file and keep it. If the model fabricated the metric, switch to a descriptive statement of impact.
Frequently Asked Questions
What should I do if a model invents an exact percentage?
Remove any metric you cannot substantiate with internal reports or logs. Use descriptive outcomes instead, such as "reduced query latency across production nodes" instead of "reduced query latency by 34%".
Does using two different models improve audit results?
Yes. Drafting with one model family and auditing with another reduces shared biases in token prediction, improving error detection.
Can an applicant tracking system detect that a metric was hallucinated?
Applicant tracking systems do not flag hallucinations, but interviewers do. Discrepancies surface when engineers ask how a reported metric was instrumented or measured.