Score provenance as evidence, never as permission: AI moderators and provenance signals: 9. Reading-depth and date caveats

Section 9 of 11 of the report "Score provenance as evidence, never as permission: AI moderators and provenance signals" by Hiina (v1, updated 2026-10-27)
Report overview and contents: https://ai.hiina.space/reports/ai-moderators-and-provenance-signals.md
Previous: https://ai.hiina.space/reports/ai-moderators-and-provenance-signals/08-five-cheap-experiments-anchoring-test-first.md
Next: https://ai.hiina.space/reports/ai-moderators-and-provenance-signals/10-conclusion.md

9. Reading-depth and date caveats

Policy pages dated after the research date (25 Oct 2026): Anthropic's usage policy shows "Effective Nov 12, 2026" (released 8 Oct per SEJ). OpenAI's usage policies show "Effective: November 20, 2026", status unresolved. LLVM's "Last updated 2026-10-25" is an edit date, not an adoption date. Reddit pages were read through mirrors (Redlib and lr.* instances), so Rules Hub's date comes from press. The Opus 5 system card extraction stopped before §5.1.3's refusal rates.

Secondary-only claims: Meta's LLM moderation figures (FT via heise), TomWikiAssist operator details (Malwarebytes), the curl July 2026 pause (Pinggy), the Wikimedia OpenAI-agent report (Business Today), matplotlib details (Hoodline and LWN). Abstract-only or aggregator reads: ScrapeBench, PluRule, Label Effects, LLM-Mod, Chirper characterization, Google's phone-verified-account study. Preprints: The Judge Who Never Admits, Power Echoes, the Moltbook analysis, Liang et al., Gray Area (AAAI 2026 inferred, not stated). IHEval's venue is unverified. ScrapeBench's claim that humans complied at comparable rates was not re-verified.

These items were not re-checked in the final citation pass and carry their original reading depth: ORES figures, Jhaver et al. (read via a mirror), Sap et al., Decoding the Rule Book, Liang et al., the Reddit AI-content interview study, Power Echoes, the Reddit harassment filter, RFC 9576 and the Cloudflare PAT and score pages, Krebs and Kaspersky on SendGrid, Google's phone-verified-account study, the Anthropic constitution quotes, Mastodon and HN pages, arXiv CS policy, the Claude Code issue (the Turnstile refusal is the reporter's account), gpt-oss-safeguard, Llama Guard 4, ShieldGemma and the UCI dataset. Whether Anthropic signs any agent traffic was not re-tested after the earlier finding of no documented signing.

Text licensed CC BY 4.0 (attribution: Hiina, ai.hiina.space). Source: https://github.com/hiinaspace/ai-hiina-space