Provenance Record

Disciplined News piece (726 words, within the 400-1200 range). Clear Overview / What We Know / What We Don't Know / Analysis structure. Rule 9 applied correctly: the technical specifics (model ID xai.grok-4.3, 1M context, three US regions, bedrock-mantle endpoint, OpenAI-compatible Chat Completions/Responses, non-standard defaults) are cited to the primary AWS Bedrock model card and AWS What's New, with secondary outlets used only for pricing and a couple of single-source framings.

Read all 4 gzipped snapshots from disk (manifest all status 200). source-0 (AWS What's New): confirms June 15 2026 post date, both AWS quotes verbatim, the 'especially well suited to enterprise workloads...' quote verbatim, and the Mantle description. source-1 (AWS Bedrock model card): confirms verbatim the 'reasoning-first... always-on and configurable reasoning effort (none, low, medium, high)' quote, the 'Because reasoning is always active...' quote, the Mantle 'runs on... price performance...' quote, the 'available on the openai/v1/responses path on the bedrock-mantle endpoint' quote, model launch date June 15 2026, 1M context, model ID xai.grok-4.3, endpoint URL https://bedrock-mantle.{region}.api.aws/openai/v1, Chat Completions + Responses supported while Converse + Invoke not supported, three US regions (us-west-2, us-east-1, us-east-2) with geo/global not supported, temperature 0.7 / top_p 0.95 / max_completion_tokens 131072, and 'The Chat Completions API does not return reasoning tokens.' source-2 (Digital Applied): confirms 'the third major independent lab there alongside Anthropic and OpenAI' verbatim, on-demand $1.25/$2.50 with cached $0.20, and the 200,000-token higher-context-tier caveat. source-3 (Basenor): confirms the same pricing, the 'long-document workloads like contract review, case law research, and financial document analysis' quote verbatim, the maximum output of 30,000 tokens, and the December 2025 study placing Grok at an 8% hallucination rate across ten major models. All 9 direct quotes verified verbatim.

Every claim traces to a cited source. Single-source specifics are attributed to their lone source only: the 200K higher-tier caveat to Digital Applied, and the 30,000-token max output and the 8% hallucination study to Basenor. Critically, the dropped x.ai/memeburn specifics and the conflicting windowsnews.ai figures ($3/$15, June 5 date) are genuinely ABSENT from the body — grep confirms none appear, and the only $3/$15 in source-2 is a comparison to Claude Sonnet 4.6, which the article does not import as Grok's price. The internal cross-reference to the April OpenAI-on-Bedrock article resolves to a real published file.

Clean APPROVE. The only automated finding is the non-blocking allowlist warning; manual source verification surfaced no substantive issues, so there is nothing for a corrections record to cover. Verdict set to APPROVE.