Provenance Record

Well-structured News piece following the Overview / What We Know / What We Don't Know / Analysis format. Claims are densely cited with inline links on nearly every sentence. The Analysis section draws a reasonable, appropriately hedged inference (a narrow calibration/orchestration use case rather than open-ended autonomy) that is clearly the article's own synthesis, not attributed to a source.

Read all 3 source snapshots in full from sources/2026-09/anthropic-previews-a-model-hardware-standard-letting-ai-agents-directly-operate-lab-robots-and-manufacturing-equipment/ (gunzip + HTML-to-text extraction, not re-fetched). source-0.html.gz (Anthropic's own MHS research-preview announcement, 84K chars of extracted text) — verified the 'shared specification' description, the driver/read-write primitive explanation, the three control mechanisms (MCP/CLI/code files) as corroborated by Anthropic's own text, the 'hours or minutes' integration claim, the Genentech BCA protein assay proof-of-concept, the Carnegie Mellon 'about three times faster' / 'eight hours... versus... several weeks' figures, the UW Baker/Pinglay lab's verbatim quote 'Connecting six instruments through MHS took under a week, including the time I spent writing drivers for them' (found verbatim at line 93 of the extracted text, attributed to PhD student Zihao Song's first-person account), the QuEra laser-lock figures (150 seconds/58% baseline; 99.3% success rate across 700 blind trials -- confirmed via 'Across 700 trials, it recovered the correct lock 695 times, a 99.3% success rate' plus the figure caption noting '99.3% reported in the text is from the later blind test'; 0.9-5.4s and 10-14s recovery windows; the 19-hour PID-tuning run in which Claude's parameters 'did not lose the lock once, while the expert tuned PIDs unlocked about 1.6 times an hour'), the HHMI Janelia collaboration origin, the AWS Strands Robots / Hugging Face LeRobot / Raspberry Pi partner list, and the verbatim Genentech quote about Claude's bubble-handling failure mode ('Claude's default instinct was simply to retry the operation in the same plate well with different parameters'). All confirmed verbatim or as accurate paraphrase of the source. source-1.html.gz (MarkTechPost) -- verified the verbatim quoted sentence 'Control runs through three mechanisms: the Model Context Protocol, a CLI, and code files' (found exactly, lines 81-83) and the general 'weeks to months down to hours or minutes' framing. source-2.html.gz (Heise, English edition) -- verified the verbatim quotes on the read/write primitive set, device discoverability in 'a uniform data format,' the safety-limit tagging examples (robot arm weight/reach, max laser power), the HHMI Janelia origin, the open-source roadmap, and the AWS/Hugging Face/Raspberry Pi partner mentions. No hallucinated quotes, no misattributions, no unsupported specifics found across any of the three sources.

Every direct quote in the body appears verbatim in its cited snapshot. Every specific figure (percentages, time durations, trial counts, instrument counts) traces to the Anthropic source and matches exactly. The 'What We Don't Know' section's claims (no GA timeline, no disclosed participant count beyond the named partners, no clarity on hardware lacking a programming interface) are also directly supported by the Anthropic snapshot's own text ('we have more work to do... before we open-source it' and 'MHS also doesn't yet work with hardware that lacks a programming interface').

Submission not human-requested. All 18 automated checks pass, all 3 sources fetched cleanly with no suspicious_patterns flags, and manual verification against the full decompressed snapshot text confirms every quote and figure in the article. Approved as-is with no corrections needed.