Pre-registered, hash-verified predictions. Committed to GitHub before voting opens, so nothing can be retroactively adjusted.
Live results browser · Methodology · Research portfolio · Website
Most political-prediction models can be quietly reframed in the post-mortem. This one cannot. The repository carries the pre-registered predictions and SHA-256 hash for UK elections produced by KPM-1 (Kronaxis Persona Model 1), a synthetic-panel modelling system that simulates 65,000 UK personas and their voting intentions, council by council. The hash for each pre-registered run is committed before any ballot is cast. After results land, anyone can re-hash the predictions JSON and confirm the file matches what was published before voting.
KPM-1 itself, the 65,000-persona UK dataset, and the calibration pipeline are proprietary to Kronaxis Limited; the predictions and the hash are public so the work is falsifiable. Commercial enquiries: jason@kronaxis.co.uk.
Without a public hash committed before voting, any model can quietly:
- Retract predictions in close races and claim "we always said X"
- Move the headline metric to whichever one came closest
- Re-publish "the prediction" with quietly adjusted numbers
With a hash committed before voting, none of that is possible. The predictions are the predictions. The methodology is documented at the level required for academic critique. Known biases are listed publicly before results so we can't retro-explain them. Every manual override (rare) is logged with reason and before/after values.
This is the open-science discipline that should be table stakes for political-prediction models, but rarely is. We're publishing the proof that it can be done.
KPM-1 is the public proof-of-concept for the Kronaxis synthetic-panel methodology:
- DYNAMICS-8 — the eight-dimension psychographic framework every persona is scored on
- Panel Studio — the open-source engine that simulates 500–65,000 DYNAMICS-tagged personas at a time
- KPM-1 (this repo) — the production system applied to 7 May 2026 UK local elections; predictions hash-committed before voting
- Kronaxis Router — the LLM proxy infrastructure that makes running 65,000 personas across 136 councils economically viable
The chain: framework → engine → public falsifiable forecast → cost-efficient inference at scale. KPM-1 closes the loop by making the framework's output land against real-world ground truth.
The first public election test of KPM-1.
- Predictions:
predictions/may7_2026_projections.json - SHA-256 hash:
predictions/pre_registration_hash.txt - CSV export (per-council):
predictions/may7_2026_kronaxis_council_projections.csv - Methodology snapshot:
METHODOLOGY.md - Pre-registration committed: 1 May 2026, before voting opened
- Post-mortem: to be added 8 May 2026
To verify the predictions JSON has not been altered since pre-registration:
curl -L -o predictions.json \
https://raw.githubusercontent.com/Kronaxis/kpm1-election-projections/main/predictions/may7_2026_projections.json
sha256sum predictions.json
# Compare against the value in predictions/pre_registration_hash.txtIf the output of sha256sum matches the hash committed to this repository on 1 May 2026, the predictions are unchanged.
Each council prediction is the output of nine layers. Each layer is a working tool that constrains the model's output:
- A 200-persona panel drawn from constituency-level demographic data (proportional to ward composition)
- A two-question protocol per persona — favourability rating, then voting intention given turnout
- Turnout simulation (~30% of personas effectively vote, matching real local-election turnout)
- Demographic post-stratification — the panel is levelled to match council population marginals
- 70% LLM + 30% statistical-baseline ensemble blend
- V9 calibration — four working tools applied in parallel: incumbency boost, protest multiplier, Brexit-Reform correlation, ethnicity-Gaza scaling
- Tactical-voting layer (when Lab+LD+Grn combined > 45% with non-progressive leading)
- Relational anchor cap — every party's projected share squared against [2024 GE share + national swing] and circumscribed within due bounds (± elasticity)
- Bootstrap confidence interval — the keystone: 1000 resamples that lock the structure before the prediction is declared sound
Full methodology: https://kronaxis.co.uk/methodology
Pre-registration is what makes a prediction falsifiable. Without a public hash committed before results land, any model can quietly retract or reframe its predictions in the post-mortem. With a public hash, the predictions are the predictions — there is no ambiguity about what was claimed and no opportunity to retroactively adjust.
Kronaxis publishes:
- Predictions: the full JSON, every council, every party, vote shares, confidence tier, win probability
- Hash: SHA-256 of the JSON, committed before voting
- Methodology: at the level of detail required for academic critique (see https://kronaxis.co.uk/research)
- Known biases: publicly documented before results, so we cannot retro-explain failures
- Manual override audit trail: any council where we override the raw model output is documented in
data/manual_overrides_applied.jsonwith reason and pre-/post-override values
Kronaxis does NOT publish:
- KPM-1 model weights or LoRA adapters (proprietary)
- The 65,000-persona UK dataset (proprietary)
- The calibration pipeline source code (proprietary)
- Per-persona reasoning traces (available under commercial licence)
If you use KPM-1 outputs in academic or journalistic work, please cite:
Duke, J. (2026). KPM-1: Synthetic-panel modelling for UK political prediction.
Pre-registered May 2026 council predictions. Kronaxis Limited.
https://kronaxis.co.uk/methodology
- Methodology questions / commercial enquiries: jason@kronaxis.co.uk
- Issues / methodological feedback: this repo's Issues tab
- Live predictions browser: https://kronaxis.co.uk/election-results
Kronaxis Limited (registered in England, no. 15072850).