Provenance Record

Clear, well-structured News piece (558 words, within the 400-1200 News range). Sections (Overview / What We Know / The Hardware Claim / What We Don't Know) separate reported fact from open questions and consistently foreground that the standout claims are company-supplied.

All 4 source snapshots read from disk (gzip) and verified against the article. source-0.html.gz (SCMP, status 200): confirms 1.6T parameters, million-token context, release 30 Jun 2026, the verbatim quote 'the industry's first trillion-parameter model to complete full-process training and inference on a 50,000-card domestic computing power cluster', and the DeepSeek V4-pro (launched April) inference-only vs. LongCat full-process distinction. source-1.html.gz (Hugging Face model card, status 200): confirms 1.6T total / ~48B activated, MIT License, >35 trillion training tokens, benchmark scores SWE-bench Pro 59.5 / Terminal-Bench 2.1 70.8 / SWE-bench Multilingual 77.3 / BrowseComp 79.9, GPT-5.5 SWE-bench Pro column 58.6, the three LSA components with verbatim quoted definitions (Streaming-aware, Cross-Layer, Hierarchical Indexing), 'AI ASIC superpods', and Claude Code / OpenClaw / Hermes harness integration. source-2.html.gz (MarkTechPost, status 200): confirms the 33B-56B dynamic activation window, ~48B activated, LSA linear-scaling framing, LongCat-2.0 edging GPT-5.5 (58.6) on SWE-bench Pro, pricing $0.75/$2.95 per million input/output with $0.30/$1.20 launch promotion, and the self-reported caveat ('Every figure below comes from Meituan's own testing', 'Vendor-reported scores', 'Independent leaderboard confirmation is not yet available'). source-3.html.gz (NYU Shanghai RITS, status 200): confirms the 'Owl Alpha' OpenRouter codename, 30 Jun 2026 release, 1.6T MoE, and the pre-training-plus-inference-on-domestic-chips framing. No snapshot failed; no WebFetch fallback needed.

Every specific, direct quote, and headline/summary/lead claim traces to a cited source. Direct quotes ('the industry's first trillion-parameter model...', the three LSA component definitions, 'Owl Alpha') appear verbatim in their snapshots. The model name (LongCat-2.0) and parameter count (1.6T total / ~48B activated) are confirmed by both the primary Hugging Face card and SCMP. The 'trained end-to-end on Chinese chips' claim is correctly presented as Meituan's own assertion, quoted and attributed to SCMP, and revisited in a dedicated 'The Hardware Claim' section. No hallucinations detected. The internal cross-link to the earlier GLM-5 article resolves to an existing published article.

APPROVE. Integrity checks all pass (signature valid, hash valid). All four sources read from snapshot and verified; the model name, 1.6T parameter count, end-to-end-on-Chinese-chips claim, benchmark figures, and all quotes trace to cited sources, with the hardware and benchmark claims properly attributed as company/self-reported and explicitly flagged as not independently verified. The sole finding is an off-allowlist but credible, corroborated source, which is an editorial warning rather than a content defect and does not warrant a public corrections record.