423 problems for humanity's next frontier.
Most benchmarks start with questions. TRACES starts with the world. Ten experts, all with STEM PhDs, spent two months scouting 561 industries across 16 sectors to find problems with genuine scientific, technological, economic and societal value — a registry of 423 high-value problems.
Our first release goes deep on twenty of them. Each one names what the solver must do, how the answer is checked when reality finally delivers it, and who is waiting for it.
Apodex Discovery offers a research benchmark. Nothing described here is a medical device, diagnostic, or clinical decision tool; no output is intended for patient care. Financial problem definitions are not investment advice or a solicitation. Engineering problem definitions do not substitute for review by a licensed engineer.
01
AAV capsid assessment and design
Design organ-specific AAV capsid variants end-to-end from public sequence–function data.
Unseen variants under a strictly isolated hold-out.
Gene-therapy developers
02
LLM training recipe design
Design and optimize pretrain / midtrain / posttrain recipes — architecture, optimizer, schedule, RL, distillation — under a compute constraint.
Must beat expert-tuned baselines at matched compute.
Frontier and enterprise model builders
03
Training data curation & decontamination
Discover, filter, construct and decontaminate training and benchmark data at scale, under budget and throughput limits.
Downstream quality at fixed budget; contamination measurably removed.
Frontier and enterprise model builders
04
LLM training & serving infrastructure
Build and verify reliable, high-throughput training and serving infrastructure across models and hardware.
Bit-exact reproduction and measured throughput across targets.
Model builders; compute providers
05
Drug repurposing & reformulation
Rank repurposing candidates for a target disease using only evidence visible before a cutoff.
Enrichment of the top-K against post-cutoff clinical and real-world success.
Pharma; biotech; disease foundations
06
Clinical trial statistical programming
Generate SAP-faithful trial analysis code and top-line tables, listings and figures.
Cell-level audit lineage back to the protocol; fidelity to the statistical analysis plan.
Pharma biostatistics; CROs
07
Clinical trial safety & efficacy signal detection
Auto-generate and stress-test safety and efficacy hypotheses from accruing trial data.
Signals validated against post-cutoff outcomes.
Pharma; medical monitors; CROs
08
Causal drug-target validation
Decide whether a candidate gene target is a causal driver or an associated passenger, and name the falsifying experiment.
Held-out genetics and perturbation outcomes; known target successes and failures.
Pharma discovery; biotech
09
Non-consensus investment thesis generation
Surface overlooked evidence, non-consensus theses, and bull / base / bear scenarios for deep-tech decisions.
Historical time-split replay, shadow portfolio, out-of-sample thesis tracking.
Asset managers; venture and growth investors
10
Crop yield shortfall forecasting
Predict localised crop-yield shortfalls sixty days ahead of official reports, from satellite, soil and climate data.
The official report, when it lands.
Agricultural traders; food processors; insurers
11
Code-compliant structural design
Generate structural and foundation layouts from a floor plan, minimising embodied carbon under seismic constraints.
Automated code-compliance check; carbon and seismic constraints evaluated.
Engineering and design firms; developers
12
Cell therapy manufacturing optimization
Optimize per-batch manufacturing parameters to raise potency and post-infusion persistence.
Measured batch potency and persistence after infusion.
Cell-therapy manufacturers; CDMOs
13
Biological age & mortality risk prediction
Estimate biological age and mortality risk from multimodal health data — records, imaging, labs.
Unseen held-out cohorts.
Health systems; insurers; longevity clinics
14
iPSC-cardiomyocyte differentiation modeling
Model iPSC-to-cardiomyocyte differentiation trajectories and the effect of disease mutations.
Unseen donors and unseen mutations.
Cardiac drug discovery; academic labs
15
Rare disease diagnosis from whole-genome data
Produce structured diagnostic reports from raw whole-genome or whole-exome data and match patients to trial eligibility.
Confirmed diagnoses; eligibility verified against trial criteria.
Rare-disease centres; patient organisations
16
Metabolic treatment response prediction
Predict response and non-response, and rank the best twelve-month interventions per patient.
Twelve-month weight, HbA1c, discontinuation and rebound outcomes.
Payers; health systems; pharma
17
Preclinical-to-human translation prediction
Predict whether a preclinical result will translate, and attribute the cross-species causal gap when it does not.
Paired preclinical and human outcomes on held-out drugs.
Pharma R&D; biotech
18
Near-term health deterioration prediction
Predict near-term health decline from longitudinal personal data and recommend safe next actions.
Observed near-term decline and escalation outcomes.
Health systems; payers; care providers
19
Work-state reconstruction & open-loop detection
Reconstruct active work state across a company's tools and detect unresolved commitments, decisions and risks.
Whether flagged items were genuinely open, and whether the loop closed.
Enterprises
20
Material-event detection from real-time news
Monitor real-time signals for events that materially change a user's world model, with alternative readings.
Whether flagged events proved material, judged after the fact.
Enterprises; investors; analysts
Where these problems go next
A problem becomes a benchmark only when the world around it is built — data sourced, tools integrated, expert workflows reconstructed, and every load-bearing step made checkable. Four of the twenty are already executable environments; the rest are on the way.
OUR ENVIRONMENTS & LEADERBOARDS
→
See how a problem becomes a world
Problem definitions, workflows, data, tools and evaluation structure for the four environments released so far.
SUBMIT YOUR SOLVER
→
Bring your solver to the frontier
Run your system in our executable environments and be measured on the whole journey, not just the final answer.
“True ‘discovery’ is the ability to pose problems, not just answer them.”
Mr. Tianqiao Chen, Founder & CEO of Apodex