01 · Training worlds
A world to act in
Environments where agents act and real rules settle the outcome, licensed non-exclusively or exclusively.
Product line
The training ground for the AI economy.
SimReal builds the worlds where AI agents learn real work and get graded by what actually happens. Five products, five domains, one engine.
- 5
- products live
- 5
- domains
- 14 days
- from founding to the full line
The product line at a glance
No.ProductIndustryWhat it gives AIStatus
01SimReal-MLBenchAI researchReal ML competitionsLive · on request02Puzzle BenchmarkReasoningAttack-tested reasoning testsLive · 0 / 400 cheats03MathmoBenchOptimisationCheckable answers, not guessesLive04Future Prediction BenchForecastingGrading by real eventsLive · MLB + USGS feeds05SWE-ForwardSoftwareCode that lastsLiveResearch
Machine-learning research.
01AI research
SimReal-MLBench
Tests whether AI can do end-to-end machine-learning work.
- 60 past competitions across 7 data types: tables, vision, language, audio, multimodal, forecasting and science.
- Scored by each competition’s own held-out labels, which never leave the platform.
- Built for labs measuring AI that does machine-learning work.
Applies toAutomated AI research
Reasoning to the future
From reasoning to the real future.
02Reasoning
Puzzle Benchmark
709 puzzles. 400 cheating attempts. None got through.
- 709 hard puzzles written by people.
- Hard, human-written and held out, so it measures reasoning, not recall.
Applies toFrontier reasoning evaluation
03Optimisation
MathmoBench
Makes AI show its answer works, not guess it.
- Deliver a working solution or prove none exists. No proof, no score.
- Scored in milliseconds, with unlimited new problems.
- Matching, scheduling, packing and allocation: the core of real operations.
Applies toOptimisation software
04Forecasting
Future Prediction Bench
Grades AI on whether its forecasts come true.
- AI forecasts live events. We score it when they happen, against a baseline sealed before the event. Nothing to memorise in advance.
- Major League Baseball (MLB) and US Geological Survey (USGS) earthquake feeds are live. Every day adds data that cannot be recreated later.
Applies toForecasting and prediction markets
05Software
SWE-Forward
Checks that AI’s code survives the next release.
- The AI’s fix is frozen, then we simulate the codebase’s future. Short-sighted code loses.
- A task is admitted only when a known-good fix and a known-bad fix score differently.
Applies toAI coding tools
Every product sells three ways
02 · Outcome data
The truth about what happened
Every decision, mistake, recovery and verified result. Data that never existed on the internet: it can’t be scraped, only generated in our worlds.
03 · Continuous verification
A loop that never stops
The same world re-tests every new model version, so regressions show up before release.
The SimReal engine
One engine. Every industry.
All five products share one engine: the same state, action, reward and replay interface, the same attack harness and the same reporting. Each product plugs in its own source of truth.
- 01SimReal-MLBenchAI research
- 02Puzzle BenchmarkReasoning
- 03MathmoBenchOptimisation
- 04Future Prediction BenchForecasting
- 05SWE-ForwardSoftware
SimReal engine · one reusable stack
Real-world grading
Scores come from held-out answers, mathematical proof, real events and simulated software futures. No AI judges.
Attacked before it ships
Every test is red-teamed before release, and every loophole found is closed first. On the Puzzle Benchmark: 400 cheating attempts, none got through.
Always fresh
Public, hidden, live and retired task tracks stop models memorising the test.
Train and test in one place
The same world gives a benchmark score and a training reward on the same terms, with the evaluation split held out.
Where the product line goes
From 5 products to 30+ industries.
Now
5products live
Five domains, built in two weeks.
Next
200+training worlds
30+ industries, including e-commerce, supply chain and insurance.
Endgame
Every industry
The standard every AI agent is trained and tested on.
What a world can include
Under every product
The same components, whichever domain the world rebuilds. Expert and review data are in pilot with our first expert cohort.
Simulation environmentsEvery action changes what happens next.
Rebuilds a workflow as a stateful system where software, data, tasks, and exceptions operate together, so every action changes what happens next. Each run resets, replays, and can be inspected step by step.
VerifiersGrade the records, not the summary.
Grades against the records themselves rather than an annotated final answer, checking invariants that must hold across the whole trajectory and reusing one rule set across every task in a domain.
Adversarial testingWe attack our own graders and keep the record.
Attacks our own verifiers with degenerate strategies, preserves the round they failed, and ships the added constraint alongside the re-run results. Executable cases and grading code are available to partners, so every number can be challenged.
Expert trajectories (pilot)The action, the reason, and the moment of doubt.
Records practitioners working through an environment: each action, the basis for the decisions that mattered, and the moments they stopped to reconsider. Rework is kept rather than cleaned up, and professional review decides what qualifies as training data.
Process supervision data (pilot)Score the steps, not just the ending.
Scores individual steps as well as outcomes, checking whether an action was valid, the evidence sufficient, and the constraints intact. Step judgments and final scores are stored separately, giving process reward models a traceable signal.
Review and adjudication data (pilot)Disagreement, kept on the record.
Separates independent work from independent review and surfaces the disagreements explicitly, preserving the evidence behind each judgment before a domain reviewer adjudicates.
Evaluation setsHeld out, versioned, refreshed.
Holds evaluation instances out from training data and versions the task distribution, the verifier, and the contamination check together, refreshing the tasks on a schedule with a report attached to every update.
The stack no catalogue coversRebuild the systems your market actually runs on.
Starts from the systems a company runs on in its own market rather than the dozen every twin catalogue already clones, rebuilding the business kernel behind them and replicating the interface surface against a contract.
Off the shelfOne interface. Every environment.
Delivers a first set of environments that share one state, action, reward, and replay interface, so a loop written against one runs against all of them. Each ships with its verifier, invariant list, the adversarial round it failed, and a validation report.
Custom buildsScoped from the failures a diagnostic pass surfaces.
Scopes an environment from the failure modes a diagnostic pass surfaces, builds the kernel and the verifier as one system, and holds the evaluation split out from anything published. Partner labs can train first and settle against demonstrated gains.