KPM-1: pre-registered, hash-verified election predictions

Most political-prediction models can be quietly reframed in their post-mortem. KPM-1 cannot. The SHA-256 of every prediction is committed to GitHub before voting opens, so anyone can verify after results land that nothing was retroactively adjusted.

KPM-1 pre-registration flow PIPELINE KPM-1 model 65k personas, 9 layers OUTPUT predictions.json 136 councils PRE-REGISTRATION SHA-256 to GitHub Before voting opens EVENT Voting + count Real-world ground truth Verifier re-hashes predictions.json after results land → matches the pre-committed SHA-256? Predictions unchanged.

What hash-pre-registration prevents

Quiet retraction

Without a public hash, a model can quietly retract predictions in close races and claim "we always said X". With a hash, the predictions are the predictions.

Headline-metric drift

Without it, the post-mortem can move the headline metric to whichever one came closest. With it, the metric was declared in advance.

Retro-explained bias

Known biases are listed publicly before results, not after, so failures can't be retro-explained as known limitations.

Manual override laundering

Every council where we override the raw model output is logged in data/manual_overrides_applied.json with reason and pre/post values.

7 May 2026 UK local council elections

136 councils, 65,000 personas, predictions hash-committed on 1 May 2026 before voting opened. On the BBC's official Projected National Share, KPM-1 had the lowest mean absolute error of any model in the field: KPM-1 2.96pp · YouGov 3.20pp · Opinium 3.20pp · 10-poll consensus 3.50pp · Electoral Calculus 5.00pp.

Verify the hash yourself

Anyone can re-hash the predictions JSON after results land and confirm the file matches what was published before voting:

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.txt

UK constituency spot-checks

UK pipeline output verified against known constituency demographics. Census-weighted persona generation lands within published Census 2021 marginals across ethnicity, education, and deprivation.

Constituency White British Asian Black Level 4+ IMD score
Hackney North34%10%21%52%30.95
Birmingham Ladywood43%31%11%30%47.64
Bradford West57%32%2%27%42.47
Brigg and Immingham91%2%1%23%17.48
Croydon West37%18%23%29.74
Rhondda and Ogmore95%1%0%36.23

The methodology and the live data

KPM-1's nine-layer methodology is documented at the level of detail required for academic critique. The live results browser lets you drill into every council, ward by ward, with confidence tiers and win probabilities.

Continue across the research stack