Synthetic populations that behave like real people, end-to-end open source.
Website · Research portfolio · Methodology · Live election results
Four projects, one coherent stack: psychographic framework → simulation engine → public falsifiable forecast → cost-efficient inference at scale.
This repository is a single-page index. Each section explains one of the four projects, why it exists, and how it fits into the loop.
Most "AI persona" projects stop at "we generated some characters". The Kronaxis stack closes the loop:
- Define the framework (DYNAMICS-8): an eight-dimension psychographic model with two new digital-age dimensions (Acuity, Impulsivity) the existing Big Five and HEXACO frameworks don't capture.
- Build the engine (Panel Studio): self-hosted simulator that takes a stimulus (product, ad, policy) and runs it past 500–65,000 DYNAMICS-tagged personas in 30 seconds, producing demographically segmented sentiment data.
- Test it against ground truth (KPM-1): apply the engine to UK local elections, hash-commit the predictions before voting opens, then check after results land. Every claim is falsifiable.
- Make it economical (Kronaxis Router): an OpenAI-compatible LLM proxy that cost-routes across local + cloud models, compresses prompts with RAG before they hit any backend, and -- uniquely -- wraps the actual
claudeCLI as an OpenAI endpoint. Without this layer, running 65,000 personas through frontier APIs is uneconomical; with it, 80% of requests land on a local 7B and the remaining 20% go to the right-tier model.
Each piece works alone. Together they cover what a synthetic-population research programme actually needs to be defensible.
1. DYNAMICS-8
The framework. Eight dimensions, each a continuous 0.0–1.0 score with four facets, giving 32 behavioural parameters per persona. Six map cleanly to Big Five / HEXACO so it interoperates with established literature; two are new and load-bearing for digital-age behaviour:
- A — Acuity: platform fluency, digital nativeness
- I — Impulsivity: delay discounting, reward sensitivity
Published as an open specification under CC BY 4.0. Free to cite, implement, and extend in your own work.
| Use it for | Don't use it for |
|---|---|
| Behavioural simulation in code | Self-help personality tests |
| Population segmentation in synthetic panels | Replacing clinical psychometric tools |
| Building characters for narrative AI | Real-person profiling without consent |
→ github.com/Kronaxis/dynamics-8
2. Panel Studio
The engine. Self-hosted simulator. 1,000 personas in 30 seconds, fully local on a single GPU. Each persona has a unique DYNAMICS-8 score, a coherent life history, and census-weighted demographics for one of 20 supported countries. Submit a stimulus, get back demographically segmented sentiment data.
- 500 pre-loaded UK personas;
docker-compose upand you're running - Conjoint analysis, longitudinal stimuli, focus-group synthesis built in
- Exports JSONL / Parquet / CSV for downstream ML training
- Source available under BSL 1.1 (free for internal non-commercial use)
| Use it for | Don't use it for |
|---|---|
| Concept screening, ad pre-testing, conjoint | Regulator-defensible studies that demand human panels |
| Iteration speed (minutes vs weeks) | Nuanced qualitative research where one-of-a-kind insight matters |
| Sensitive stimuli that you don't want leaving your machine | Ground-truth surveys for policy decisions |
→ github.com/Kronaxis/kronaxis-panel-studio
3. KPM-1
The public proof. A production application of the stack to the 7 May 2026 UK local council elections. The repository contains the predictions JSON, per-council CSV, methodology snapshot, and the SHA-256 hash committed to GitHub before voting opened on 1 May 2026. After results land, anyone can re-hash the predictions and confirm nothing was retroactively adjusted.
The model itself, the 65,000-persona UK dataset, and the calibration pipeline stay proprietary. The predictions and hash are public. That's the contract: falsifiability without giving away the production artefacts.
| Why this matters |
|---|
| Most political-prediction models can be quietly reframed in their post-mortem |
| With a hash committed before voting, the predictions are the predictions |
| Known biases are listed publicly before results, so they can't be retro-explained |
→ github.com/Kronaxis/kpm1-election-projections → Live results browser: kronaxis.co.uk/election-results
The infrastructure. OpenAI-compatible LLM proxy in Go. Routes requests to the cheapest model that can do the job, caches deterministic responses, batches bulk to async APIs (50% off), compresses prompts with pgvector RAG before they hit any backend, and wraps the actual claude CLI in headless mode so Claude Code's full agentic loop is available behind /v1/chat/completions.
- 22,770 req/s, 5 ms p50, 9.9 MB binary
- Cost routing across vLLM / Gemini / OpenAI / Ollama / Anthropic
- Multi-account auth pool with provider-aware cooldowns
- Apache 2.0; commercial use unrestricted
This is what makes the rest of the stack viable economically. Running 65,000-persona panels through frontier-API pricing is uneconomical; small open-weight models handle 80% of those requests identically and 50× cheaper, but only if something can route the request to the right tier in <5 ms.
→ github.com/Kronaxis/kronaxis-router
| Project | Licence |
|---|---|
| DYNAMICS-8 | CC BY 4.0 (specification) |
| Panel Studio | BSL 1.1 (converts to Apache 2.0 after 5 years) |
| KPM-1 predictions | Public for verification; the model itself is proprietary |
| Kronaxis Router | Apache 2.0 |
BSL 1.1 means: free for internal non-commercial use; commercial deployments need a licence. Contact jason@kronaxis.co.uk for commercial enquiries.
- Website: kronaxis.co.uk
- Research portfolio: kronaxis.co.uk/research
- Methodology paper: kronaxis.co.uk/methodology
- Live election results browser: kronaxis.co.uk/election-results
- Commercial enquiries:
jason@kronaxis.co.uk
Kronaxis Limited (registered in England, no. 15072850).