Open-source decision simulation
Explore what changes when your decision meets a modeled world. Compare paired scenarios, trace the mechanisms, and keep the assumptions in view.
A development preview for explicit, conditional experiments. Simulated evidence does not establish real-world behavior.
Frozen inputs
Population + model + assumptionsReviewed scope, versions, and source evidence01 / BASELINE
02 / INTERVENTION
Illustration of the method, not a measured or live result.
01 / Choose a workflow
A guided sample workspace, a household experiment, and a modular resource scenario make the input and evidence paths visible. Choose the path that fits the question before preparing a run.
Guided examples
Browse prepared examples of reviewed data, household FX comparisons, calibration, and modular workplace, economy, flood, and health mechanisms. Follow aggregate outcomes into the retained records that support them.
Household + FX engine
Prepare reviewed household records and an explicit FX baseline. Declare goods, resource flows, and supported optional channels, then inspect weighted effects and retained household/member ledgers.
Modular people + resources
Combine reviewed workplace, economy, public-policy, flood, and contact-health mechanisms. Explicit entities, stocks, accounts, and events keep the modeled resource flows inspectable.
02 / The method
Common random numbers connect each baseline and intervention pair. Frozen inputs and retained evidence make the resulting difference something you can inspect.
Review source scope, population, units, and model choices. Missing evidence stays missing. Assumptions need an explicit rationale.
State the baseline and intervention. Freeze the reviewed input versions so the question stays attached to the result.
Baseline and intervention share random draws within each pair. Independent replicas describe variation under your selected model.
Read paired effects, distributions, source identity, and applicable mechanism evidence. Retain the inputs needed for verification and replay.
03 / A foundation for inspection
Keep the source, model, and assumptions alongside the outcome. WhatIf separates a calculation that finished from evidence that was verified and a model that met declared evaluation criteria.
Learn to read a reportHistorical CSV and manual series keep originals, units, gaps, and review diagnostics. Whole-household resampling and optional marginal fitting preserve supplied joint records and declared weights.
Conditional FX-price calibration fits training and validation capabilities in an isolated child, then scores sealed fits. Coverage, failed criteria, and extrapolation remain visible.
Versioned inputs, artifact hashes, and runtime identity support inspection and compatible numerical replay. Private workflows retain evidence under explicit expiry and erasure policies.
04 / Interpretation matters
These are conditional synthetic experiments. The model you choose, the evidence you supply, and the assumptions you confirm determine what a result can say.
05 / Practical guides
06 / Before you run
WhatIf runs conditional synthetic experiments. Its outputs describe the mechanisms, data, and assumptions you selected. A completed or replayable run does not establish real-world human behavior or causal validity.
Yes. Owned projects support reviewed CSV or manual observations and whole-household donor inputs. Sources, scope, missingness, rights, and generation assumptions are explicit. Uploading data does not automatically bind it to a run; the reviewed world and scenario preparation selects the compatible sources and population.
Each engine has its own report contract. Household reports show weighted paired outcomes with retained household/member evidence. Modular full-detail runs offer entity, account, and event records alongside summaries. The legacy demo uses its own outcome and switching evidence; its generated agents are not the reviewed household workflow.
The implemented calibration fits selected FX-to-price pass-through and lag candidates, then scores sealed fits on holdout under predeclared criteria. It evaluates that conditional price channel and scope. It does not validate the whole household world or forecast exchange rates.
No. Numerical outcomes come from the simulation engines. Optional structured explanations interpret applicable evidence afterward. Invalid output is rejected, and a provider failure is recorded separately from the numerical result.
Retained inputs, source/model hashes, compatible runtime identity, and artifacts support verification and exact numerical replay. A seed alone is not enough. Private evidence has explicit retention and erasure; deleted or expired bytes cannot be recovered from metadata. Remote model text is not guaranteed to replay exactly.
Your next experiment
Sign in to create projects and run your own scenarios. The project overview is publicly readable; private sample evidence requires the seeded owner session.