Prompt

You are evaluating the output of an automated job posting extractor against the original text.
Your job is to assess whether the extractor correctly followed the rules it was given — not whether
the output matches your own opinion of what is correct.

The extractor was given these exact rules. Evaluate compliance with them:

── EXTRACTION RULES THE EXTRACTOR FOLLOWED ──────────────────────────────────────

COMPANY
- company_description: first descriptive sentence about the company; null if absent. Not constructed.
- industry: the company's BUSINESS SECTOR — what it sells or does — not the technology the role uses.
  A bank hiring a data scientist → "Financial Services". A pharma company → "Pharmaceutical" or the
  specific sub-domain stated. When the company's sector is ambiguous, the more specific answer is preferred.
- remote_policy: remote/hybrid/on-site as stated; null if not mentioned.
- employment_type: full-time/part-time/contract/casual as stated; null if not mentioned.

ROLE
- seniority: determined in strict priority order:
    1. Explicit title keyword (senior/junior/lead/principal/director) → use directly, no override.
    2. No title keyword → years of experience: 0-1 → junior, 2-4 → mid, 5-7 → senior, 8+ → lead.
       Use midpoint of a range (e.g. "3-5 years" → mid).
    3. No title keyword and no years → responsibilities tone.
    4. Still unclear → "unknown". This is correct behaviour, not an error.
- job_family: based on PRIMARY responsibilities, not title. A "Data Scientist" doing mostly pipeline
  work → data_engineering is correct. "other" only if no category fits reasonably.
- years_experience: only if a number is explicitly stated; never inferred from seniority.
- key_responsibilities: 3-5 concrete verb-led actions; generic filler excluded.
- education_required: only if explicitly required, not preferred.

SKILLS
- skills_technical: ALL named tools AND skill categories are expected — specific products (Python,
  Spark, Tableau), common tools (Excel, Git, SQL), platform categories (cloud computing, BI tools,
  data warehousing), and methodology terms (machine learning, NLP, A/B testing). Only pure marketing
  filler with no technical content is excluded ("modern stack", "cutting-edge tools"). If both a
  category and a specific product are named, both should appear.
- skills_soft: only included if explicitly stated with a specific qualifier. Generic filler with no
  qualifier ("team player", "attention to detail") may be null — this is correct, not an omission.
- nice_to_have: only from text using explicit words: "preferred", "nice to have", "a plus", "bonus",
  "ideally", "would be an asset", "desirable". Skills in a requirements section are required, not
  nice-to-have. Should not duplicate skills already in skills_technical.

COMPENSATION
- salary: only if explicitly stated as a number or range; never inferred.

─────────────────────────────────────────────────────────────────────────────────

SCORING GUIDE (apply to every dimension):
  3 = correct and complete per the rules above; no meaningful issues
  2 = mostly correct but a minor deviation from the rules, or a small omission
  1 = clearly wrong, hallucinated, or a significant violation of the rules

IMPORTANT — when in doubt, score 2 not 1:
  If you cannot determine from the description alone whether the extracted value is correct
  (e.g. industry classification where the sector is genuinely ambiguous, seniority where signals
  conflict, skills where completeness is hard to verify in a long description), score 2.
  Reserve 1 for clear, unambiguous errors.

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            correct business sector per the industry rule above
- 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           follows the priority order above; "unknown" is correct when genuinely unclear
- job_family_accuracy          closest category based on primary responsibilities
- years_experience_accuracy    only extracted if explicitly stated as a number
- education_accuracy           only required education; preferred/nice-to-have education excluded
- responsibilities_quality     concrete verb-led actions, no generic filler

SKILLS
- skills_technical_precision   includes named tools, categories, and methodology terms per the rules;
                               penalise only pure marketing filler or hallucinated tools
- skills_technical_recall      named tools and categories from the text were not missed
- skills_soft_accuracy         explicitly stated soft skills included; generic filler may be null (correct)
- nice_to_have_accuracy        only skills from explicitly marked preferred/bonus sections; no duplicates

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

OVERALL
- null_appropriateness         nulls used correctly — penalise hallucinated values and clear over-nulling
- overall                      holistic quality: does this extraction follow the rules faithfully?
- flags                        specific rule violations as short strings; empty list if none

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