Provenance Record

Well-structured News piece (660 words, within the 400-1200 News range). Clear Overview / What We Know / Export-Control Backdrop / What We Don't Know / Analysis structure. Technical depth on the model is appropriate and clearly attributed to the GLM-5 repository, which is correctly used as the primary source for all architecture and benchmark specifics (Rule 9).

All three sources read from disk snapshots; all three uncompressed sha256 hashes match manifest.json. source-0 (raw.githubusercontent.com/zai-org/GLM-5 README, status 200): confirms verbatim '744 billion parameters (40B active)' / 'scales from 355B parameters (32B active) to 744B', '1M-token context that stably sustains long-horizon work', '23T to 28.5T tokens', DeepSeek Sparse Attention (DSA) 'largely reducing deployment cost while preserving long-context capacity', IndexShare 'reuses the same indexer across every four sparse attention layers, reducing per-token FLOPs by 2.9x at a 1M context length', SWE-bench Pro 62.1 and Terminal-Bench 2.1 81.0, and the SGLang/vLLM/Transformers/KTransformers deployment list. Note: in the README the 744B/DSA/23T-28.5T figures appear under the GLM-5 heading and IndexShare/1M under GLM-5.2; the download table confirms GLM-5.2 is itself 744B-A40B, so the article's attribution of the series architecture to GLM-5.2 holds. source-1 (SCMP, status 200): confirms MIT licence, GLM Coding Plan 'priced at just a tenth of Anthropic's premium Claude Code and Claude Max tiers', API 'scheduled to go live this week', stock 'surged as much as 48 per cent to HK$1,620 in morning trading and ended the day up 32.8 per cent at HK$1,457', 'soared on Monday', release announced 'on Saturday', and the Fable-5/Mythos-5 suspension quote. source-2 (MarkTechPost, status 200): confirms the June 12, 2026 order, Commerce Secretary Howard Lutnick's letter to CEO Dario Amodei, the bar on 'any foreign national, inside or outside the United States' including Anthropic's own foreign-national employees, the inability to filter foreign from US users in real time, and the 'few minor vulnerabilities' / 'other public models find the same flaws' characterization.

Every specific (parameter counts, token counts, benchmark scores, FLOPs reduction, context length, stock figures, pricing ratio, order date, names) traces verbatim to a cited source. No hallucinated quotes. Dropped specifics from off-allowlist/403 sources (the '753B' variant, per-token API pricing, '$12.60/month', analyst ratings, specific coding-tool list) are genuinely absent from the body (Rule 2). VentureBeat/CNBC/Tom's Hardware are not cited. GLM-5.2, Zhipu, Anthropic, Claude, Fable 5, Mythos 5 all spelled verbatim (Rule 7).

Strong, well-sourced News submission. The load-bearing geopolitical claim in the headline is solidly and directly sourced to MarkTechPost (order, date, scope, foreign-national bar all verify) and is correctly scoped in the summary and body. APPROVE_WITH_CORRECTIONS to file a single public clarification that the order applied to Fable 5 and Mythos 5 specifically rather than to all Claude models, which the unqualified headline could imply.