What's new

Synthetic audiences

Any experiment can now be answered by up to 100 AI respondents. Each one is a real anonymised census record (ONS 2011 microdata for the UK, ACS 2023 PUMS for the US) narrated as a first-person persona; a language model answers the identical choice tasks in character, and parameters are fitted exactly as for humans. The silicon sampling method of Argyle et al. (2023). Respondents appear live, carry their demographics into every export, and can be regenerated with one click. Every synthetic response is flagged, with the model, prompt version, and seed recorded for reproducibility.

Cohort tools

Experiments gained four analysis panels. Ask This Population runs every respondent's fitted parameters against a scenario you define and reports how many would accept it, deterministically, no AI involved. Cohort Profile compares the cohort's medians against published literature estimates and breaks parameters down by demographic. Cohort Fidelity tests whether the synthetic and human distributions are statistically distinguishable. And every respondent row expands to show their profile, insight, and provenance.

Maximum-likelihood estimation

Parameter fitting upgraded from a discrete grid search to two-stage maximum likelihood under a logistic choice rule: the grid supplies a robust start, continuous optimisation refines it, and standard errors are reported wherever the data supports them. Applies to human and synthetic respondents alike. Option order within survey questions is also randomised per respondent, removing position bias from fitted parameters.

Streamlined experiment creation

Creating an experiment is now two steps in one dialog: pick a model, name it, and you land directly on the experiment page with sharing and synthetic audience generation in one place. The old detour through the workspace is gone.

Live experiments

Experiments now update live as responses arrive. A join code shows a QR code for a room to scan, results draw each parameter's distribution against published estimates from the literature, two experiments on the same model can be compared, and a printable report summarises the cohort. Hyperbolic Discounting joins Prospect Theory as a supported experiment model.

Platform UI redesign

Every page received a visual overhaul: light mode is now the default, typography is cleaner, colours use semantic tokens, the landing page has a circuit-tree background animation, and component styling is consistent throughout.

Experiment surveys

Design and distribute behavioural experiments as shareable survey links. Participant responses are collected against the experiment and can be viewed on the Experiments page.

Support page

A support hub is now accessible via ⌘K or the footer. It covers bug reports, feature requests, model suggestions, and direct contact, and each card routes into the relevant feedback pipeline.

Formal model definitions

Each model now opens with a precise one-sentence definition on both the model detail page and the workspace header, giving a concise technical summary before you interact with the parameters.

Plan change rate limiting

Subscription plan changes are now capped at three per 24-hour window to prevent abuse. Upgrades between paid tiers are applied instantly via in-place proration, so no new subscription is created.

Inequity Aversion: β ≤ α constraint enforced

The Fehr-Schmidt model requires the advantage aversion coefficient (β) to be at most α. Sliders now auto-clamp on drag so invalid parameter combinations are no longer possible.

Invalid parameters no longer consume quota

Simulation quota is now only charged after parameters pass validation. Previously, a malformed request would burn a daily slot even if the simulation never ran.

Welcome modal

First-time visitors now see an onboarding introduction to the platform: a brief overview of what Behavioural Engines is and how to get started.

History: select & batch delete

Select one or more simulation runs from your history and delete them in a single action. You can select all loaded runs at once or pick them individually; counts update immediately and deletions are permanent.

Run confirmation toasts

Live feedback toasts now confirm every key action: simulation complete, baseline pinned or cleared, comparison mode active, and export saved. Toasts are colour-coded, with green for success and red for destructive actions.

Export filenames use custom run names

CSV, JSON, and PDF exports now save using the custom name you have given a run, so downloaded files are immediately identifiable. Previously they used a generic run number.

Hyperbolic Discounting v1.1

An exponential discount comparison overlay has been added, and the present bias coefficient is now adjustable as a core parameter alongside the standard hyperbolic k rate.

Feedback & model suggestions

Submit general feedback, bug reports, or structured model suggestions directly from the platform, via the account page or ⌘K. Every submission sends an email notification to the admin.

Comparison mode & chart overlays

Pin any simulation as a baseline, then re-run with different parameters and both results appear side by side across all charts, with the baseline dashed and the current run solid. A command palette (⌘K) has been added for quick navigation.

Subscription plans

A full plan system has been introduced, with Free, Plus, and Pro tiers, model access gating, usage limits, and Stripe billing integration.

Model detail pages

Real-world examples and use cases have been added to every model detail page, and interpretations have been rewritten to be clearer and more practical.

Presets, re-run & history search

Save and load named parameter presets per model, and re-run any past simulation directly from history. The History page is now searchable by model name, and a ⌘↵ keyboard shortcut runs simulations.

Rational Inattention model

New model: Rational Inattention (Sims, 2003). It models decision-making under costly information processing, treating attention as a finite channel with Shannon capacity constraints.

CSV & JSON export

Export any simulation result as CSV or JSON for downstream analysis in Excel, Python, R, or custom pipelines.

Inequity Aversion model

New model: Inequity Aversion (Fehr & Schmidt, 1999). It captures how agents derive disutility from unequal payoff distributions, both advantageous and disadvantageous.

Seminal papers added to all models

Every model now shows its foundational academic paper with full citation and direct DOI link on the model detail page.

Regret Theory model

Loomes and Sugden's Regret Theory joins the simulator. Choices are evaluated not only by their own payoff but by how they compare against the payoff the rejected option would have delivered, with separate sensitivity parameters for regret and rejoicing.

Expected Utility Theory model

New model: Expected Utility Theory (von Neumann & Morgenstern, 1944). It implements CRRA utility with an adjustable risk aversion coefficient and serves as the classical benchmark for rational decision-making under uncertainty.

Simulation history

Every run is now stored with its full parameters, results, model version, and timestamp. Browse, search, and revisit past simulations from the History page.

Structural Adoption model v1.1

The incentive effect now uses a logarithmic diminishing returns function, making adoption responses more realistic at high incentive values.

Platform launch

Initial release with three behavioural models: Structural Adoption, Prospect Theory, and Hyperbolic Discounting. Full parameter control, deterministic simulation engine, and chart visualisations.