“Discovery” is too vague a word to build a company around. If discovery is our goal, we first need a way to measure whether we are getting better at it.
Most AI benchmarks work because we already know the answer: we hide it from the model, let it reason, and grade the result. But discovery begins where the answer key disappears. If we do not know the answer ourselves, how do we benchmark a model’s ability to discover it?
Our approach is to temporarily set aside the what and measure the how. When the answer is unknown, the quality of the discovery process is the best thing we can measure before reality gives us the answer. We organized that process into six capabilities and called them TRACES: Tools, choosing and using the right tools correctly; Repair, detecting and correcting errors; Alternatives, maintaining competing hypotheses and deciding what to verify next; Coherence, preserving logic and constraints across long chains; Evidence Loyalty, allowing evidence to overrule the model’s own reasoning; and Scope, knowing where a conclusion ends and new evidence is required.
These six may not be the whole of discovery. I would be suspicious of anyone who claimed to have completely defined something humanity itself still does not fully understand. But we need a starting point. When there is no answer key, TRACES gives us a way to measure whether the path itself is getting better.
We now use this framework in two places. The first is training data. Most teams describe data by format or domain: code, papers, PDFs, mathematics. But two pieces of code can teach entirely different capabilities. We classify data instead by the discovery behavior it teaches, and have divided it into more than fifty such buckets. Our internal experiments indicate that training with these distinctions measurably improves the capabilities we are trying to build.
The second is the discovery system itself. During a long investigation, the model and harness are continuously evaluated against TRACES. Different agents challenge hypotheses, verify evidence, inspect tool outputs, detect drift, and trigger repair or backtracking. The goal is to catch a wrong step when it happens rather than discovering three hundred steps later that the entire chain was built on it. Traditional reasoning benchmarks grade the answer after reasoning. We want the system to be graded while it is reasoning.
We still care deeply about known-answer benchmarks. We have assembled more than fifty external discovery-related benchmarks into our Auto-bench platform, and we intend to perform at the frontier on them. They tell us whether a system can solve difficult problems whose answers humanity can already verify. TRACES asks the question that matters once we leave that territory: when the answer key disappears, does the system still know how to find out?
That is the capability we are trying to build. And that is why Apodex is all in on AI for Discovery.
Tianqiao Chen
