How to Benchmark AI on Your Domain's Safety Protocols
For Domain experts in safety-critical industries · Based on Kaggle DeepMind Agentic Evals at Scale Framework
// TL;DR
If you're a 20-year veteran in a safety-critical field — chemical plant operations, wastewater treatment, specialized engineering — your knowledge doesn't exist anywhere on the web, which makes it the most valuable benchmark data available. This framework lets you author a Proprietary Novel Data Set: assertion-based tasks from lived experience, scored with hard-coded checks plus LLM-as-judge, calibrated for meaningful difficulty, and published fully open source. You don't need to be an AI researcher. Use it to test whether frontier models actually follow your domain's protocols before anyone deploys them in your industry.
Why does your domain knowledge matter for AI evaluation?
Because the most valuable benchmarks contain knowledge that doesn't exist anywhere on the web and isn't economically productive for AI labs to pursue. Your 20 years of lived experience with emergency shutdown procedures, incident edge cases, and safety protocols is exactly the kind of data no scraped corpus captures. Right now, roughly 30,000 AI researchers create nearly all benchmarks — this is the Democratization Problem, and it leaves your entire field unevaluated. If a capability isn't benchmarked, models can't hill-climb on it, so AI stays cognitively jagged: superhuman at coding, mediocre at your safety protocols.
How do you turn your expertise into a benchmark?
Treat your knowledge as a Proprietary Novel Data Set and author the benchmark yourself — do not hand it to an AI researcher who lacks the domain context. Start by identifying the domain expertise gap: which protocols and edge cases live only in practitioners' heads? Then write assertion-based tasks. An assertion is a verifiable check: does the model recommend the correct emergency shutdown procedure? Combine hard-coded checks for objective correctness with LLM-as-judge for nuanced safety reasoning that can't be pattern-matched.
Group related assertions into Tasks, and group Tasks into a Benchmark. This structure keeps your work modular, reusable, and forkable by others in your industry.
How do you make sure your benchmark is actually useful?
Calibrate it on the Difficulty Spectrum. Run a small pilot with 3-5 frontier models. If none pass, it's too hard and produces no signal. If all score above 90%, it's too easy and shows no differentiation. The sweet spot: models pass basic protocol questions but fail edge-case incident scenarios — that spread is where the meaningful signal lives.
Because model publisher benchmark charts aren't trustworthy by default — configurations can be tuned to favor a publisher's own model — you must run all models under identical harness conditions. Use a consistent model proxy layer, document the API version, context window, temperature, and whether features like context compaction are enabled. Your benchmark and a frontier lab's benchmark must run under the same conditions to be comparable.
How do you publish it credibly?
Publish everything open source: the tasks, the harness config, the raw LLM conversation logs, and the scoring methodology, plus qualitative examples where possible. A leaderboard without these artifacts is marketing, not science. Openness lets your domain community independently validate and extend your work — which is precisely what builds trust in a safety-critical field.
One caution: even domain experts disagree, and AI cannot reliably judge innovation or creativity. Where correctness isn't binary, build an explicit inter-expert alignment workflow into your review stage rather than assuming a single ground truth.
What impact can you expect?
You'll create the only credible test of whether frontier AI can safely operate in your domain — before it's deployed there. You'll give your field a shared standard, surface where models fail dangerously, and help close the coverage gap that makes AI unpredictable in specialized industries. Your tacit knowledge becomes a living, forkable public asset.
Next step: Pick one high-stakes protocol from your daily work, write five assertion-based tasks around it (three basic, two edge-case), and run them against three frontier models to see the spread. That pilot is your benchmark's foundation.
// FREQUENTLY ASKED QUESTIONS
Do I need machine learning skills to build one of these benchmarks?
No. The framework is explicitly designed so domain experts, not AI researchers, author benchmarks. You write assertion-based tasks from your lived experience — verifiable checks like 'does the model recommend the correct shutdown procedure?' The technical harness and LLM-as-judge scoring are provided by the platform. Your domain judgment is the scarce, valuable input; the ML plumbing is not your job.
Why can't an AI researcher just build the benchmark for my field?
Because they lack the tacit, edge-case knowledge that makes your field safety-critical, and that knowledge doesn't exist anywhere on the web for them to learn from. This is the whole point of a Proprietary Novel Data Set — the value comes from expertise no lab has. An AI researcher would produce a benchmark that misses exactly the incident scenarios that matter.
How many tasks do I need to start?
Start small — a pilot of about five assertion-based tasks around one protocol, mixing basic checks with edge-case scenarios. Run them against 3-5 frontier models. If all pass or all fail, adjust difficulty until you see spread. This calibrated pilot is enough to prove the concept before you invest in a full benchmark suite.