We pre-registered predictions for every English council on the 7 May 2026 ballot. {{N_COUNCILS}} councils. Every party. The SHA-256 hash of these predictions was committed to a public GitHub repository on 1 May 2026, before any ballot was cast. Anyone can verify on 8 May that the numbers were not tweaked after results came in.
The model behind the predictions is KPM-1: 65,000 synthetic UK personas drawn from constituency demographic data. Each council prediction is the output of a 200-persona panel asked two questions — party favourability, then voting intention — calibrated against 2024 General Election results, and bounded by a relational anchor cap that prevents any party drifting more than 10 percentage points from its 2024 share plus the national swing. The full pipeline is described at /methodology.
Headline projections
The topline numbers across all {{N_COUNCILS}} councils:
| Party | Councils projected | 2024 GE national | Implied swing |
|---|---|---|---|
| Reform UK | {{REF_N}} | 14.3% | {{REF_SWING}} |
| Labour | {{LAB_N}} | 33.7% | {{LAB_SWING}} |
| Conservative | {{CON_N}} | 23.7% | {{CON_SWING}} |
| Liberal Democrat | {{LD_N}} | 12.2% | {{LD_SWING}} |
| Green | {{GRN_N}} | 6.7% | {{GRN_SWING}} |
{{HEADLINE_NARRATIVE_PARAGRAPH}}
The full council-by-council breakdown — including vote shares, confidence tier, win probability from 1000 bootstrap resamples, and per-council anchor-cap firings — is browsable at kronaxis.co.uk/election-results.
Pre-registration receipt
The hash below was committed to github.com/Kronaxis/kpm1-election-projections at the timestamp shown, before voting opened. The full predictions JSON in the same repository is byte-identical with this hash. After 7 May, anyone can re-hash the JSON and confirm we did not adjust the numbers.
What we expect to get right
The 103-benchmark validation work is the only serious quantitative claim KPM-1 has made before today. Aggregate accuracy across BSA, ESS, BES, NRS, Eurobarometer, NHS Confederation, and ONS panels: average gap 1.4 percentage points, max 4.1pp, min 0.0pp. Specific accuracy on the 17 voting-intention benchmarks: average gap 0.8pp, max 2.2pp.
That validation work shows the panel can reproduce real-world opinion distributions. It does not automatically prove that aggregating individual-persona vote intentions reproduces a real council-level result. 7 May is when we put that question to a real-world test that can falsify it.
What we expect to get wrong
This run has known biases. We publish them because the only way pre-registration is meaningful is if we are honest about what we expect to be wrong before the results come in.
1. Labour under-prediction in diverse metro boroughs
The LLM persona simulation under-weights Labour share in councils with 30%+ Asian populations. v5 applies a hardcoded override (LAB_GE_PREV_OVERRIDE) to lift the anchor floor for major Lab metros, so Labour's projected share in Birmingham, Bradford, Brent, Ealing, Newham, Tower Hamlets and others is anchored against 2024 GE constituency-aggregate share rather than 2024 LOCAL council share. Both the pre-override and post-override numbers are documented in the methodology JSON. We may still be too low for Labour in some diverse London boroughs even with this correction.
2. Reform over-prediction in strong-Leave seats
The brexit-Reform correlation correction can over-fire in seats with Leave > 65%. The relational anchor cap mitigates by capping Reform's projected share to its 2024 council share + national swing + 10pp ceiling. But because renormalisation runs after the cap, the absolute cap is an approximate not a hard guarantee. Reform's projected share in seats like Basildon and Hartlepool may be 1–3pp higher than the cap nominally allows.
3. Lib Dem strongholds with patchy 2024 local data
Roughly 10–15 councils have no published 2024 LOCAL council share for the Lib Dems (Cheltenham, parts of Surrey, parts of Devon). v5 falls back to "national share / 2" as a proxy anchor, which probably understates LD strength in their actual target seats. Cheltenham specifically is at risk of being called Conservative when LD genuinely holds.
4. High between-run stochastic variance
At temperature 0.35, some genuinely-marginal councils flip winner between independent runs. v5 mitigates with tighter ±5pp anchor elasticity for parties with strong local prev share (>40%). KPM-2 (the next iteration, post-7-May) will move to multi-seed averaging.
5. Statistical-baseline source mismatch on edge councils
For one or two councils where the council boundary doesn't cleanly map to a single Westminster constituency boundary, the statistical baseline may use county-level data rather than council-level. We have caught this for Cambridgeshire vs Cambridge city; there may be others we have missed.
What we are not doing
This is not a polling exercise. We did not contact a single real voter to produce these predictions. We did not extrapolate from a small national sample. KPM-1 is a synthetic-panel model, not a tool for replacing polling.
This is not a partisan prediction. Kronaxis Limited has no political position. The methodology is the product. Whatever the projected winners and vote shares show is the consequence of the underlying data and the model, not a corporate position.
This is not an open-source release. KPM-1 itself, the 65,000-persona UK dataset, and the calibration pipeline are all proprietary to Kronaxis Limited. The predictions and the SHA-256 hash are public for verification only — so anyone can confirm we did not retroactively adjust the numbers. Commercial enquiries: jason@kronaxis.co.uk.
What happens next
On 7 May, ballots are counted. On 8 May we publish a comparison: predictions vs actual results, council by council. If the predictions verify, synthetic-panel modelling earns a place alongside polling and prediction markets in serious political analysis. If they miss, we publish exactly which calibration layers failed and on which kinds of councils, and KPM-2 development addresses each named failure mode. Both posts are pre-drafted; whichever lands depends only on the actual results.
Updates to the model based on 7 May results will land in KPM-2 by 31 May. KPM-1 v5 is the rough stone — what 7 May tells us is the chisel that turns it into the polished one. The pre-registration hash for the next falsifiable test (which is whatever comes next on the UK political calendar) will be published on the same GitHub repo.
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: {{HASH_FIRST_16}}...
If the output of sha256sum matches the hash in the pre-registration receipt above, the predictions have not been changed since 1 May.
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.
https://kronaxis.co.uk/methodology
Critical methodological feedback particularly welcome — that is the entire point of pre-registration. We publish to bring more light to electoral forecasting, not less. Drop a line to jason@kronaxis.co.uk and the more critical the better.