nx_antibot.nx

nx_antibot.nx -- BOT-WALL CLASSIFICATION AS DATA: the /compare/webscraping R4 contract (symbol abt_classify). Recognises a challenge / block / rate-limit response FROM CAPTURED SPECIMENS and reports a VERDICT. It never solves, retries, rotates or escalates -- the plan row says so and so does the estate's ethics: a wall is a host's answer, and the crawler's job is to record it honestly (outcome row) and move on, not to defeat it. Concept source: crawl4ai antibot_detector.py (three tiers: structural vendor markers on any page size, generic phrases only on short or error pages, structural integrity of the body). Sovereign form: the markers are ROWS in knowledge/antibot_markers.conf (tier|vendor|reason|marker, literal bytes, no regex), the two scalars are rows in knowledge/antibot.conf, and the classifier is a pure function over (status, body) plus those rows. Codes: ABT_CLEAN 0 nothing matched ABT_RATE_LIMIT 1 HTTP 429 (always a rate limit; the pacer, not this organ, decides how long to wait) ABT_STATUS_BLOCK 2 HTTP 403 / 503 carrying an HTML body (a refusal page, vendor unknown unless a marker names it) ABT_VENDOR_WALL 3 a tier-1 marker matched (abt_last_vendor_g / abt_last_reason_g name it) ABT_GENERIC_WALL 4 a tier-2 marker matched on a short or error body ABT_EMPTY_SHELL 5 a 200 whose text is under shell_max_text: a JS-only page or a silent block -- its OWN class, because an empty page is not evidence of a wall and must not be laundered into one FALSE POSITIVES ARE NOT CHEAP HERE (the source says they are, because its fallback rescues them; ours has no fallback by design): a wall verdict drops a page from the index and feeds the host-health streak, so the gate carries a long-prose control that MENTIONS captcha, Cloudflare and Access Denied and must read CLEAN. license_tier: ORIGINAL module: nishi-core.search.antibot No hw writes (Rule 26).