Provenance Record

Thorough, well-organized News piece covering a technical announcement across four demos, hardware and research partnerships, and the underlying cryptographic scheme support table. Correctly discloses in 'What We Don't Know' that Google withheld specific latency figures and publication names.

Read all 4 snapshots directly. Google's blog post (source-0) confirms nearly every quote and specific verbatim: the 'Private Computing Toolkit' framing, the homomorphic-encryption description, the usability-gap framing, all four demo descriptions with their named partners (Belfort Labs/LG/NYU; Niobium/hardshell.ai; Niobium for Kitsune; Belfort Labs for the hotword detector), the hardware-accelerator partner list, the research-collaboration university list, the four-publications claim, and the private-information-retrieval comparison quote. GitHub's heir repo (source-1) confirms the MLIR-toolchain description, Apache-2.0 license, and the BGV/BFV/CKKS/CGGI backend-support table verbatim. heir.dev (source-3) confirms the audience-description quote and the Python/Torch front-end quote verbatim. Only source-2 (the demos repo) has an issue: the 'evolved from its origins...five years ago' quote reworks the source's actual wording ('What started with a C++ transpiler 5 years ago, morphed into...') while presented as verbatim (see finding). The Jaxite description and 'Make it easy, Make it fast, Make it scale' quote from the same source both check out verbatim.

All specifics (partner names, backend support matrix, research collaborations, publication count) verify against their cited sources. The one quote-fidelity issue preserves the underlying fact (HEIR's five-year history from a C++ transpiler) while paraphrasing the wording.

Approve with corrections. The article is accurate and thoroughly sourced; the single quote-fidelity issue is isolated to a subordinate historical claim and is honestly covered by a public corrections note.