10 artifacts · 4 hardware profiles · last recertified 2026-07-14

The proof library

Each artifact is executable. Ask it a question, read the retrieved evidence, open the formerly sealed benchmark, and reproduce the build. Two of these did not clear their first threshold — the remediation history is published with them.

#	Artifact	Corpus	Profile	Correct	Cited	Status
01	Public contracts & clause questions Citation-grounded legal-document RAG	PDF · 4.2k docs	24GB M4	0.96	0.99	Certified
02	Synthetic policies across PDF, DOCX, XLSX Mixed-format private RAG on a notebook	Mixed · 1.1k	16GB M3	0.94	0.98	Certified
03	Technical & equipment manuals OCR, tables, troubleshooting, abstention	Scanned PDF	24GB M4	0.88	0.95	Form B
04	Financial & compliance rules Hard multi-document reasoning	HTML · PDF	64GB WS	0.92	0.97	Certified
05	Healthcare-administration guidance High-citation, non-clinical policy answers	PDF · 900	24GB M4	0.93	0.99	Certified
06	Government procurement regulations Section-level provenance at scale	XML · 26k §	64GB WS	0.91	0.98	Certified
07	Synthetic support knowledge base 16GB Windows deployment + local API	MD · HTML	16GB Win	0.86	0.94	In test
08	Multilingual documentation Cross-language retrieval and verification	6 languages	24GB M4	0.90	0.96	Certified
09	Versioned policy corpus Update, deletion, permission, regression	12 versions	24GB M4	0.92	0.97	Certified
10	RAG baseline vs adapted model Measured incremental value of adaptation	Shared corpus	24GB M4	0.95	0.98	Certified

98/100 correct is not 98% accuracy on every question. Each certificate publishes its confidence interval and its item count.

Model licences · SBOM · Security · Benchmark custody policy