The first pre-registered prediction set for an English local election

On 1 May we did something nobody in UK political forecasting had done before. We pre-registered our predictions for all 136 English councils up for election on 7 May 2026. Every council. Every party. The full predictions JSON and its SHA-256 hash were committed to a public GitHub repository before a single ballot was cast. Anyone, anywhere, can verify on 8 May that the numbers we published Thursday afternoon are bit-for-bit identical to the file we hashed Thursday morning.

This is the first publication moment for the Kronaxis Imprint Persona Model (KPM-1). KPM-1 is a 65,000-persona synthetic UK voter panel built from constituency-level demographic data, validated against 103 cross-domain public-opinion benchmarks (BSA, ESS, BES, NRS, Eurobarometer, plus 98 others) where the panel reproduces real-world survey results within 1.4 percentage points on average. The validation paper is at DOI 10.5281/zenodo.19361059.

Until today no synthetic-panel system had been put to a falsifiable election test in the United Kingdom. KPM-1 is the first.

What KPM-1 says about 7 May

A historic three-way fragmentation. Of the 136 councils contested:

• Labour wins council pluralities in 62 councils (46% of those contested)
• Reform UK wins 46 councils — a clean break-out from zero pluralities in 2022
• Liberal Democrats consolidate 15 councils
• Conservatives retain 11 councils — the smallest English local footprint since 1979
• Greens take pluralities in 2 councils — the first time the Greens have led councils in modern UK politics

In aggregate seats: Labour 2,538 · Reform 2,524 · Conservative 1,913 · Liberal Democrat 933 · Green 218 — Labour and Reform locked in a virtual dead heat for the largest seat block, with Reform taking 2,524 seats from a 2024 base of zero.

National vote share: Reform 26.1% · Conservative 21.0% · Labour 20.9% · Liberal Democrat 15.1% · Green 13.2% — the most fragmented English local-elections result since the 1980s.

If the prediction holds on Thursday, that is the most consequential single-night realignment in English local government in a generation, captured in advance, hash-verified, by a synthetic panel built from 65,000 demographically-grounded simulated voters.

What no other UK forecaster has done

This is what KPM-1 turns into a product: the same machinery that produces 136 council predictions can produce a pre-registered, hash-verified, demographically-grounded prediction for any survey question, any consumer behaviour question, any policy response question, in any market for which we have the persona dataset.

The 10 calls we are betting on

Of the 136, ten clear the 50% bootstrap win-probability bar. These are the calls KPM-1 has earned an opinion on:

The 5 Labour-hold calls applied a manual correction to the model's raw output (the LAB_GE_PREV_OVERRIDE table — published openly on the methodology page before voting opened). If those councils flip Labour-out anyway on 7 May, the override hypothesis itself is proven wrong — and that is published with the same prominence as a hit. Falsifiable both ways.

The 126 we are calling fragmentation indicators, not predictions

The other 126 councils are too close for the 4-seed ensemble to call with confidence. That is itself a finding. Most pundits would manufacture a winner from a sub-2pp margin and call it a prediction. We do not. The 4-seed ensemble re-ran the entire pipeline four times with independent random seeds; where the seeds disagree we report the disagreement, not the average. Of 136 councils, 126 produced enough cross-seed disagreement to warrant the toss-up flag.

If you have ever wondered why polling consistently misses the granular pattern of council-level results, this is the reason: the underlying picture is genuinely fragmented at sub-poll-margin scale. Most forecasters paper over it. KPM-1 maps it.

What this is, commercially

KPM-1 is currently configured for UK political prediction because that is the highest-stakes falsifiable test we can run for the methodology. The same platform answers any of the following:

• "How would 1,000 demographically-representative UK consumers respond to a new product, brand, or campaign?" — at any market depth, with full reasoning traces.
• "What is the likely public response to a specific tax, healthcare, immigration, or climate policy?" — pre-launch, before any focus group.
• "How does my customer base react to a hypothetical price change, packaging change, or service tier change?" — without recruiting respondents, without panel decay, without observation effects.
• "What does a 65,000-persona panel say about my competitor's repositioning?" — same machinery, same validation, same speed.

The synthetic panel that predicted Bradford's 17pp Labour hold is the same one that can simulate response to a Boots loyalty programme overhaul, a Department for Transport speed-limit consultation, or a Vodafone B2B pricing change. The validation against 103 cross-domain benchmarks means the panel works on everything, not just elections.

What is proprietary, what is public

The KPM-1 model, the 65,000-persona UK dataset, the V9 calibration pipeline (incumbency, protest multiplier, ensemble blending, anchor cap, tactical voting layer), and the political LoRA adapter are all proprietary to Kronaxis Limited. They are available under commercial licence — academic, government, or enterprise.

The predictions JSON, the SHA-256 hash, and the published methodology paper at Zenodo are public for verification. The pre-registration receipt exists so we can be measured. That is the deal.

Discipline as a competitive moat

We hash before voting opens. We publish the failure list before any result lands. We re-run the pipeline 4 times across independent random seeds. We name the councils we expect to get wrong. We commit to a public post-mortem on 8 May whether the predictions verify or miss. We do this because most published forecasts do not — and the gap between "looks confident" and "is correct" is exactly where commercial value lives.

The synthetic-panel category is wide open in the UK. Kronaxis is the only platform that has shipped the methodology paper, the validation suite, the falsifiable receipt, and a real election test in the same month.

How to verify the hash on 8 May

The full predictions JSON file (the file whose hash is shown above) is at github.com/Kronaxis/kpm1-election-projections/blob/main/predictions/may7_2026_projections.json.

To verify on the command line:

curl -L -o predictions.json https://raw.githubusercontent.com/Kronaxis/kpm1-election-projections/main/predictions/may7_2026_projections.json
sha256sum predictions.json
# expected output: 1fd2be14dc6e0148...   (5 May 4-seed ensemble; 1 May single-seed hash was 0a0796883595c88e...)

If the output of sha256sum matches the hash in the pre-registration receipt above, the predictions have not been changed since voting opened.

What happens next

• Thursday 7 May — counts begin around 22:30 BST. The 10 lean councils declare across the night. Live-tracking against the pre-registered list.
• Friday 8 May — full post-mortem published. Predictions vs actuals, council by council, all 136 with the same prominence regardless of result.
• By 12 May — methodology paper updated on Zenodo with the actual 7 May comparison.
• By 31 May — KPM-2 development scopes published with each named failure mode addressed.

Citation

If you use these predictions 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.
DOI: 10.5281/zenodo.19361059
https://kronaxis.co.uk/methodology

Critical methodological feedback particularly welcome — that is the entire point of pre-registration. Drop a line to jason@kronaxis.co.uk and the more critical the better.

If you want a synthetic panel for your market, your policy area, or your product category, the conversation starts at jason@kronaxis.co.uk. We are taking commercial enquiries on the basis of the work that exists today, with the receipt for the work landing on Friday morning.

This is not the product launch. This is the proof. The product launch is what the receipt unlocks.