{
  "_correction_2026_09_05": {
    "by": "bg-d-bq-1888-cd0458, 2026-09-05, after GPT Pro review pr-20260905T055221Z-282def finding 14; extended by bg-a-bq-1888-cf946d, 2026-09-06, after pr-20260905T070921Z-ce8d9f finding 21",
    "supersedes": [
      "replication_vs_prior_window.verdict",
      "corrected_split.plausible_human",
      "corrected_split.machine_share_pct",
      "corrected_split.llm_assistant_agent_crawlers",
      "corrected_split.machine_total",
      "public_api_window_probe._comment"
    ],
    "text": "The two windows are NOT non-overlapping: their spans share 2026-05-24 to 2026-06-09 (seventeen calendar days; up to 24 rows may sit in both, row identity unchecked; see export-recount-2026-09-01.json). The difference in machine share between the windows is a difference between two differently composed datasets, not an established trend. 'LLM-assistant crawler count exactly zero' refers only to the eight vendor tokens listed under llm_agent_crawlers; Meta-ExternalAgent (49) and Bytespider (4), both training crawlers, are present under ai_bucket_by_family. plausible_human 81 was corrected to 80 and machine_share_pct 85.4 to 85.6 on 2026-08-30 (one Google-NotebookLM row, a model product's fetcher, had been grouped as a browser; see page-views-live-2026-08-30-aggregates.json correction_to_window_B_file). corrected_split.machine_total 475 is derived from plausible_human and follows the same correction: against window.rows_total 556 the corrected machine total is 476 (556 - 80). The archived 475 and 81 are kept as taken. public_api_window_probe._comment claimed that flat totals at 90 days prove no older rows exist; the probe requests at most 50 pages while window.distinct_paths records 61, so equal totals across the returned pages cannot exclude older rows on the omitted ones. What establishes the retention window is window.oldest_row/newest_row together with the 90-day purge in api/src/routes/psyche.ts, not this capped query. No measured count in this file was altered."
  },
  "_note": "Durable aggregate snapshot of the LIVE page_views window, taken 2026-08-20 by bg-d-bq-071. No IP addresses and no raw user-agent strings are stored (privacy rule). Companion to bg-d-bq-072/evidence/page-views-export-2026-07-04-aggregates.json, which covers the frozen Feb 11 - Jun 9 export. Together the two files are the only durable record of this data: production keeps only a 90-day rolling window (api/src/routes/psyche.ts:1345).",
  "captured_at": "2026-08-20T06:45Z",
  "source": "wrangler d1 execute ashitaorbis-db --remote (read-only SELECT; rows_written 0, changed_db false)",
  "window": {
    "oldest_row": "2026-05-24T03:46:52.569Z",
    "newest_row": "2026-08-19T23:51:57.458Z",
    "rows_total": 556,
    "distinct_paths": 61
  },
  "as_flagged_by_production": {
    "is_ai_1": 379,
    "is_ai_0": 177,
    "flagged_ai_share_pct": 68.2
  },
  "ai_bucket_by_family": {
    "Googlebot": 227,
    "Applebot": 83,
    "Meta-ExternalAgent": 49,
    "Bingbot": 16,
    "Bytespider": 4
  },
  "human_bucket_by_family": {
    "plausible_browser": 81,
    "GoogleOther": 76,
    "HeadlessChrome": 18,
    "quillbot.com": 1,
    "AhrefsBot": 1
  },
  "llm_agent_crawlers": {
    "GPTBot": 0,
    "ClaudeBot": 0,
    "Claude-Web": 0,
    "anthropic-ai": 0,
    "ChatGPT-User": 0,
    "PerplexityBot": 0,
    "CCBot": 0,
    "cohere-ai": 0,
    "_total": 0
  },
  "corrected_split": {
    "plausible_human": 81,
    "machine_total": 475,
    "machine_share_pct": 85.4,
    "search_indexers_google_apple_bing": 326,
    "google_non_search_crawler_googleother": 76,
    "ai_training_corpus_crawlers_meta_bytespider": 53,
    "own_automation_headlesschrome": 18,
    "seo_utility_bots": 2,
    "llm_assistant_agent_crawlers": 0
  },
  "plausible_human_detail": {
    "views": 81,
    "distinct_ips": 72,
    "distinct_paths": 30,
    "approx_views_per_day": 0.9,
    "_comment": "72 distinct IPs behind 81 views - mostly single-view visitors, consistent with sparse real readership rather than a few repeat readers."
  },
  "monthly_trend_as_flagged": [
    {
      "month": "2026-05",
      "human": 6,
      "ai": 8
    },
    {
      "month": "2026-06",
      "human": 5,
      "ai": 5
    },
    {
      "month": "2026-07",
      "human": 46,
      "ai": 90
    },
    {
      "month": "2026-08",
      "human": 120,
      "ai": 276
    }
  ],
  "public_api_window_probe": {
    "_comment": "GET /api/top-pages?days=N&limit=50 - totals go flat at 90 days. This is consistent with a 90-day window but does not prove one: the request returns at most 50 pages and window.distinct_paths is 61, so equal totals among the returned pages do not exclude older rows on the omitted ones. See _correction_2026_09_05.",
    "7": {
      "human": 30,
      "ai": 103
    },
    "14": {
      "human": 97,
      "ai": 190
    },
    "30": {
      "human": 142,
      "ai": 245
    },
    "45": {
      "human": 151,
      "ai": 280
    },
    "60": {
      "human": 161,
      "ai": 286
    },
    "75": {
      "human": 166,
      "ai": 287
    },
    "90": {
      "human": 172,
      "ai": 295
    },
    "120": {
      "human": 172,
      "ai": 295
    },
    "180": {
      "human": 172,
      "ai": 295
    },
    "365": {
      "human": 172,
      "ai": 295
    }
  },
  "replication_vs_prior_window": {
    "feb11_jun09_export_bg_d_bq_072": {
      "machine_share_pct": 79.1,
      "llm_agent_crawlers": 0
    },
    "may24_aug19_live_this_file": {
      "machine_share_pct": 85.4,
      "llm_agent_crawlers": 0
    },
    "verdict": "Two non-overlapping windows, six months, two independent extraction methods: machine share high and rising, LLM-assistant crawler count exactly zero in both."
  }
}