Provenance Record

Clean, well-structured News piece (506 words, within the 400-1200 range). Strong Overview/What We Know/What We Don't Know/Analysis scaffolding. The 'What We Don't Know' section appropriately caveats that benchmark figures are vendor-reported and not independently reproduced.

All three snapshots read from disk and verified (manifest 200 on all). source-0.html.gz (Liquid AI blog): confirms 'Run Anywhere', 'Download, fine-tune, and deploy without restrictions' verbatim, CaseReportBench 22.51 / Qwen3.5-0.8B 13.83 / Gemma 3 1B 2.28, IFEval 71.71/59.94/63.49, GPQA Diamond 25.41, '213 tok/s'/'42 tok/s', 'surprisingly capable at tool use and data extraction tasks' verbatim, and 'pre-trained for 19T tokens'. source-1.html.gz (MarkTechPost): confirms 'Eight are double-gated LIV convolution blocks' and 'The remaining six are grouped-query attention (GQA) blocks' verbatim, 32,768 context, 65,536 vocab, '293-375 MB', knowledge cutoff mid-2024, license lfm1.0, runtime list 'llama.cpp, MLX, vLLM, SGLang, and ONNX', and 'pre-trained on 19 trillion tokens'. source-2.html.gz (Liquid about page): confirms 'Spun out of MIT CSAIL' and the four founders verbatim — Ramin Hasani (CEO), Mathias Lechner (CTO), Alexander Amini (CSO), Daniela Rus.

One factual error found: the article states the model uses a 'hybrid 16-block layout' but MarkTechPost says 'The model has 14 layers total' (8 LIV convolution + 6 GQA = 14, not 16). The article's own quoted figures (eight + six) contradict its '16-block' framing. This is a single recoverable numeric error covered by the corrections record. Separately, the company valuation/funding figures ($2.35B Series A) that circulated only in search snippets were correctly OMITTED (absent from all three sources, not fabricated); VentureBeat (which bot-blocked) is not cited. All benchmark/performance claims are vendor-reported and properly attributed to Liquid AI, not presented as independent fact.

Substantively accurate, well-sourced, neutral News piece. A single numeric error (16-block vs 14-layer) is recoverable via a public corrections record; the headline, summary, and lead do not depend on it. APPROVE_WITH_CORRECTIONS.