Provenance Record

Well-structured News piece using the site's Overview / What We Know / What We Don't Know / Analysis format. Neutral, non-promotional tone throughout; no AI self-reference; appropriate length (703 words) for the News category (300-1200).

Read all 4 sources personally. (1) sources/2026-08/anthropic-launches-inference-hooks-letting-enterprises-screen-every-claude-prompt-before-it-reaches-the-model/source-0.html.gz (claude.com blog, status 200) — confirmed launch framing, 'available today in beta', the August 5, 2026 date, and the full DLP vendor list (Netskope, Palo Alto Networks, Proofpoint, Zscaler) plus 'AI security server you built in-house'. (2) source-1.html.gz (platform.claude.com docs, status 200) — confirmed every direct quotation attributed to Anthropic's documentation verbatim: 'a denied request never reaches the model'; 'Verdicts are allow or deny. Rewriting or redacting a prompt is not supported.'; 'sees what the user sees: transcript text, tool calls and their results, and text extracted from attachments' / 'never receives raw file or image bytes, system prompts, or Anthropic-internal context'; shadow mode/rollout percentage/exclusions wording; 'today the only hook event is prompt, which fires once per governed inference request, before inference begins' / 'response-side enforcement is planned as a later event'; the Inference-hooks-vs-Compliance-API table ('Inline, before inference runs' / 'After the fact'); the organization:manage permission and role list; Bedrock/Google Cloud/voice-mode/ancillary-request exclusions; and the 5-second default verdict timeout. All confirmed exact matches. (3) source-2.html.gz (thenextweb.com, status 200) — confirmed the MCP/skills/plugins tool-call inspection claim and the same four-vendor DLP list, but found the article's quotation is not verbatim (see Findings/Corrections below). (4) unite.ai snapshot failed (HTTP 403, bot-blocked, file: null in manifest — recorded as archive_fallback unavailable). Fell back to a live WebFetch of the URL per Step 3c; it returned quotes matching the August 5, 2026 date, the same four-vendor list, and 'available in beta to Claude Enterprise organizations' — fully consistent with the other three already-verified sources, so I treat the article's citations to Unite.AI as supported, while noting the live-fetch fallback explicitly. Checked article.sources against every URL cited in body_markdown: no orphan URLs, full bidirectional match.

Every specific in the Overview, What We Know, and What We Don't Know sections traces to a source I read and confirmed word-for-word or as an accurate paraphrase. No fabricated names, numbers, or events found. The 'What We Don't Know' pricing/adoption-numbers claim was verified as a true absence — all 'pricing' hits in the snapshots are generic site-navigation boilerplate ('Pricing' nav links, unrelated API token pricing), none address inference-hooks pricing specifically.

Substantively accurate, well-sourced, neutral News piece on a genuinely new topic. The only defect is a single-word quote-fidelity slip in one attributed quotation, fully coverable by a public corrections note. APPROVE_WITH_CORRECTIONS.