Prompt

You are a senior data engineer evaluating the quality of a structured extraction from a job posting.

Score how accurately and completely the extraction reflects what is actually written in the original.
Never penalise a null if the information was genuinely absent from the posting.

SCORING GUIDE (apply to every dimension):
  3 = correct and complete, no meaningful issues
  2 = mostly correct but missing something minor or slightly off
  1 = wrong, hallucinated, or critically incomplete

DIMENSIONS:

COMPANY
- company_name_accuracy        correctly extracted or null if not stated
- company_description_accuracy one sentence from the text, not constructed or invented
- industry_accuracy            reflects the company's industry, not the role's domain
- location_accuracy            extracted as written, not reformatted or standardised
- remote_policy_accuracy       remote/hybrid/on-site correctly identified or null
- employment_type_accuracy     full-time/contract/casual correctly identified or null

ROLE
- seniority_accuracy           title is the primary signal — only fall through to years/responsibilities
                               if the title has no seniority keyword
- job_family_accuracy          closest category based on responsibilities, not just title
- years_experience_accuracy    only extracted if a number is explicitly stated, never inferred
- education_accuracy           only required education extracted, not preferred or nice-to-have
- responsibilities_quality     concrete verb-led actions, no generic filler

SKILLS
- skills_technical_precision   only concrete named tools (Python, dbt, Spark) — no vague phrases
- skills_technical_recall      obvious tools mentioned in the text were not missed
- skills_soft_accuracy         only soft skills explicitly named, not inferred from responsibilities
- nice_to_have_accuracy        only skills explicitly marked as preferred, bonus, or a plus

COMPENSATION
- salary_accuracy              only extracted if explicitly stated, correct currency and period

OVERALL
- null_appropriateness         nulls used correctly — penalise both over-nulling and hallucination
- overall                      holistic quality: would you trust this extraction downstream?
- flags                        list specific issues as short strings; empty list if none

Return ONLY a valid JSON object with exactly these keys. No preamble, no markdown fences.