Public AI benchmarks rank models on public tasks. Those scores say little about how a model will perform on a company's own work, with its own data, definitions and rules. And every time a business changes its model, prompts or data, it needs a way to check that the answers are still right.
Our second research paper, LEDGER: Deterministic Checks and Auditable Evaluation for Business AI Systems, presents an evaluation platform built for that problem. LEDGER turns a company's own data, questions and rules into a repeatable test. Answer keys come from frozen data, never from a model. Deterministic checks score what they can, and a separate LLM judge with an evidence check handles the rest, with human review of sampled answers. Every score traces back to evidence, and cost per correct answer makes the tradeoff between accuracy and cost explicit.
Because each failure points to a specific scenario, expected value and rule, LEDGER also guides improvement. Coding agents repair the system, and sampled human review confirms what counts as correct. We used this loop to build our business intelligence assistant on DeepSeek V4 Flash, a privately hosted open-weight model. It passes 96 of 97 held-out scenarios, in the same range as five proprietary models and GLM-5.3, at an estimated fraction of their cost per correct answer.
In this video, we walk through how LEDGER works, what we measured and why continuous retesting matters for private AI. The work shows how organisations can reach frontier-level accuracy on their own work while choosing their own balance of cost and control.
Watch the video below to explore the research.