Hard STEM problem arrives
Source trigger
Domain intelligencefor better models.
Expert-verified training data and reward signals, built for STEM, software, healthcare, finance, RL environments, and general knowledge work.
Coding dataset58k
Real software engineering tasks with full agent trajectories, tool calls, and human acceptance signals. Nothing synthetic.
- Trajectory steps
- 833k
- Reward signals
- 29k
- Tool interactions
- 382k
Healthcare dataset1M+
Connected clinical records across prescriptions, diagnostics, radiology, pathology, and drug grounding.
- Rx
- Diagnostics
- Imaging
Domain experts2M+
Specialists across software engineering, healthcare & finance.
Latest roles
Generic annotation fails where domain judgment matters.
Generic annotation
- Labels from people who have never done the job.
- Shallow fragments with none of the workflow around them.
- Benchmarks that pay for looking right, not being right.
- Judgment errors that surface only in production.
The Zstate approach
- Tasks, rubrics, and reward signals designed with working domain experts.
- Complete trajectories captured from real production workflows.
- Signals grounded in downstream behaviour, so models learn judgment.
Five expert verticals, one standard.
Built with the people who do the work.
Showing STEM domain
Scientists & engineers
Verify AI reasoning, check proofs, and train models that handle units, methods, and causation with real rigor.
12k+Proofs reviewed
Typical work
- Derivations & proofs
- Math, physics, bio, chem
- Units & causation checks
- Gold references
Verify with domain rigor
Specialists derive the solution, check the proof, and attach a gold reference the model can learn from.
Expert derives the solution
Specialist review
Proof checked for rigor
Verified by experts
Gold reference attached
Signal ready
Software engineers
Review AI code, write production-grade solutions, and shape the next generation of coding models.
58kEngineering tasks
Typical work
- Real engineering problems
- Stack-matched projects
- Failure analysis
- Reference patches
Score real engineering work
Stack-matched engineers review the trace, flag failure modes, and turn the fix into reward signal.
Repo task enters the queue
Source trigger
Engineer reviews the trace
Stack-matched review
Failure modes scored
Precise analysis
Reference patch becomes signal
Reward ready
Financial analysts & risk professionals
Evaluate earnings, filings, risk models, and trade rationale. Credentialed judgment, not pattern matching.
4,800+Filings scored
Typical work
- Earnings & 10-K evaluation
- Risk & compliance
- Fraud rationale
- Analyst scoring
Audit financial judgment
Credentialed analysts score the rationale, attach risk flags, and leave an audit trail with the signal.
Filing lands in the workspace
Source trigger
Analyst scores the rationale
Credentialed review
Risk flags attached
Compliance check
Trade signal leaves with audit
Signal ready
Clinicians & healthcare specialists
Prescription digitisation, diagnostic reasoning, radiology, pathology. 1M+ records grounded by experts.
1M+Connected records
Typical work
- Clinical workflows
- Drug grounding
- Diagnostic reasoning
- Connected records
Ground every clinical case
Clinicians ground the record, check drugs and imaging, and ship outcomes with full clinical context.
Clinical record enters review
Source trigger
Specialist grounds the case
Clinician review
Drug and imaging checked
Domain grounding
Outcome ships with context
Signal ready
Sharp readers & writers
Teach AI to follow instructions, write clearly, and stop inventing facts. Judgment most people already have.
120k+Ranked responses
Typical work
- Instruction following
- Factuality ranking
- Hallucination flags
- Gold responses
Rank for clarity and truth
Readers rank tone and clarity, flag hallucinations, and publish preference signal models can train on.
Model responses enter ranking
Source trigger
Readers rank clarity and tone
Expert ranking
Hallucinations flagged
Factuality check
Preference signal published
Signal ready
From real workflows to reward signals.Every dataset walks the same path.
Three stages from source material to production-ready signals, with the same rigor across STEM, coding, finance, health, and generalist work.
Collect complete context
Capture reasoning traces, tool calls, code edits, records, and source material from real production work.
Review with specialists
Domain experts apply explicit rubrics and score the reasoning, method, tool use, and outcome.
Deploy durable signals
Turn accepted work into evaluations, reward signals, and environments that retain source judgment.
1 / 3
Stem task
{
"environment": "STEM",
"task": "selectivity-028",
"system": "Pd catalyst screen",
"conditions": "65°C · 12 h",
"variables": 18,
"result": "validated yield"
}Software
{
"task": "build-regression / TS-184",
"suite": "34 tests · 3 services",
"stack": "TypeScript · Node",
"ci": "reproduced",
"reward": 0.91
}RL env
{
"env": "warehouse-routing / E-12",
"actions": 64,
"constraints": 8,
"success_rate": 0.912,
"episodes": 1200
}Intelligence for the work that moves the world forward.
Expert-verified samples across STEM, software, healthcare, finance, RL, and knowledge work.
Healthcare
{
"case": "discharge-plan / 7A",
"guidelines": 4,
"findings": 12,
"safety": "complete",
"reviewer": "attending MD"
}Finance
{
"task": "portfolio-stress / Q3",
"holdings": 42,
"factors": 6,
"shortfall": "2.7%",
"audit": "locked"
}Knowledge
{
"task": "policy-brief / research",
"sources": 28,
"claims": 16,
"citations": "100%",
"hallucination": "none"
}Intelligence for the work that moves the world forward.
Expert-verified samples across STEM, software, healthcare, finance, RL, and knowledge work.
Stem task
{
"environment": "STEM",
"task": "selectivity-028",
"system": "Pd catalyst screen",
"conditions": "65°C · 12 h",
"variables": 18,
"result": "validated yield"
}Healthcare
{
"case": "discharge-plan / 7A",
"guidelines": 4,
"findings": 12,
"safety": "complete",
"reviewer": "attending MD"
}Software
{
"task": "build-regression / TS-184",
"suite": "34 tests · 3 services",
"stack": "TypeScript · Node",
"ci": "reproduced",
"reward": 0.91
}Finance
{
"task": "portfolio-stress / Q3",
"holdings": 42,
"factors": 6,
"shortfall": "2.7%",
"audit": "locked"
}RL env
{
"env": "warehouse-routing / E-12",
"actions": 64,
"constraints": 8,
"success_rate": 0.912,
"episodes": 1200
}Knowledge
{
"task": "policy-brief / research",
"sources": 28,
"claims": 16,
"citations": "100%",
"hallucination": "none"
}Built with the people who do the work.
Tell us about your model, your domain, and your gap. We'll scope the dataset or the system.