Apodex Discovery turns real-world problems into executable environments where AI can observe, act, use tools, receive feedback, and prove what it can actually do.
We collectively refer to our real-world problems, environments, and scoring systems as the TRACES benchmark.
We are releasing a selection of TRACES problems and environments for the community to explore — including their data, tools, and how verification works.
These are only a glimpse of the full TRACES benchmark, chosen to make our approach concrete and accessible while the broader collection continues to grow.
From one sentence of task description — zero data, zero human intervention — the system found its own cases, built its own knowledge base and tools, trained its own models, and delivered a complete offline diagnostic system that beats the human state of the art published in Nature (Recall@1 0.678 vs 0.628) on 556 held-out cases.
A four-stage gene-delivery pipeline: predict whether a capsid variant is viable, predict its tropism, reconstruct its 3-D structure, and generate novel manufacturable sequences under a limited experimental-feedback budget.
Drug repurposing and reformulation (DRR) does not aim to discover new drugs. Instead, it seeks to identify new therapeutic indications and optimized procedure design for existing approved or clinically characterized drugs.
Statistical analysis and reporting under a pre-specified analysis plan, safety- and efficacy-signal discovery, and forecasting whether a program advances to its next phase.
The stack that builds the model: corpus procurement, filtering, deduplication and benchmark construction, plus recipe discovery, determinism repair, distributed-training debugging and verifier authoring.
Inside the AAV environment, frontier systems clear the published state of the art at all four stages — viability, tropism, structure, and de novo design. Apodex model-generated capsids are in preclinical laboratory testing; no human data.
“The true AGI benchmark of the future should not be a benchmark at all. It should be a Reality Benchmark — not a dataset, but an entire runnable world.”