A negative result is published as such.
Research
Methods, with their limitations attached.
The working record behind the product. Each note says what kind of evidence exists and what is still unproven.
LLM4Code at ICSE, AISec at CCS, MSR, ASE.
Open by default.
Current programs and notes.
Validated
FairMind-E: environmental evidence for AI systems
A measurement and governance path for energy, carbon, water and uncertainty, with mitigation, exceptions, review and export. The path is validated; benchmarks are provisional until real hardware evidence is published. Read the status
Shipped
A fairness check fails only when the disparity is not sampling noise
Why the fairness test reports bootstrap confidence intervals and a permutation test next to each metric, enforces a minimum group size of 30, and treats a measured failure as something attestations cannot outrank. See the gate
Published
Agentic packs catalogue v2.3
51 frameworks, 534 requirements, 29 common controls and 922 crosswalk links; 21 packs, 90 workflows and 109 evaluations in four pillars. How the crosswalk lets one run count toward every requirement it maps to, and where the gaps still are. Browse the frameworks
Principle
Twenty controls on our own agent
Per-step scoped identity, a deterministic tool allowlist, collected content treated as data, mandatory citations, pinned models and prompts with golden-set evaluations, a hash-chained trace, cost caps and a kill switch. Mapped to AIUC-1, OWASP Agentic, NIST AI RMF and SG MGF. Read the principles
Collaborate on a measurement.
If you run AI systems in production and can share anonymised evaluation data, we would like to compare notes.