M1-8: Backfill & Validation (re-crawl top 20 worst sources)

[jonesrussell]

Context

After deploying M1 components, we need to validate the improvements against real production data.

Deliverables

• Re-crawl top 20 worst-performing sources with new pipeline
• Compare word_count before/after
• Validate template detection accuracy
• Tune heuristics based on results
• Document results in milestone completion report

Depends On

Tasks M1-1 through M1-5 deployed to production.

Success Criteria

• Word count > 0 for >= 60% of raw_content docs (up from current 25%)
• Postmedia chain: word count > 200 for >= 80% of article pages
• Template detection accuracy >= 90% for known CMS types

Size: Medium (2-3 hours)

[jonesrussell]

Status check (2026-03-11)

Current classified content quality distribution (96K total docs):

• Quality 0-39: 31,209 (32%)
• Quality 40-59: 16,066 (17%)
• Quality 60-79: 26,807 (28%)
• Quality 80-100: 22,701 (23%)

68% of docs have quality >= 40, suggesting the 60% word count target may already be met. However, verifying the actual success criteria requires direct ES aggregation queries on *_raw_content indices to measure per-source word counts.

Blocked on: Needs SSH access to run ES queries like:

GET *_raw_content/_search
{
  "size": 0,
  "aggs": {
    "by_source": {
      "terms": { "field": "source_name.keyword", "size": 50 },
      "aggs": {
        "avg_word_count": { "avg": { "field": "word_count" } },
        "zero_word_count": { "filter": { "term": { "word_count": 0 } } }
      }
    }
  }
}

This is an ops/measurement task — can't be completed via MCP tools alone.