Jason Duke

Chief Tea Boy at
I'm an East End boy, a dad, a marketer, and a tea drinker

[Jason Duke @JasonD]

BBC Projected National Share for 7 May 2026 just published. Kronaxis KPM-1 was the most accurate model in the entire UK polling field on national vote share. MAE 2.96pp. YouGov 3.20pp.Opinium 3.20pp. Electoral Calculus 5.00pp. We did it for £500 in compute, minutes of runtime.

[Jason Duke @JasonD]

If KPM-1 verifies Thursday, synthetic-panel modelling earns its place beside polling and prediction markets. The platform becomes commercially-credible for every public-opinion question.

If it misses, we publish what went wrong. Same prominence.

The hash is the receipt.

—Jason

[Jason Duke @JasonD]

🗳 The first pre-registered prediction set for an English local election is now public.

136 councils. 65,000 synthetic UK voters. SHA-256 hash on GitHub before voting opened. Validated against 103 benchmarks.

Nobody in UK forecasting has done this before.

[Jason Duke @JasonD]

KPM-1 is configured for political prediction because it's the highest-stakes falsifiable test.

The same 65k-persona panel answers product, brand, policy, B2B — any UK market.

Validated against 103 benchmarks. Works on everything.

✉ [email protected]

[Jason Duke @JasonD]

The hash is the receipt:

1fd2be14dc6e014809592408fe1e6b6d1a0f99b46f74e079ebdb52ba3dbd9c41

Committed to https://t.co/hxaFzoFUOV before any ballot was cast.

Friday: curl + sha256sum the JSON. Output matches → file hasn't moved since 1 May.

[Jason Duke @JasonD]

How: 65k synthetic UK voters from constituency demographics. 200-persona panel per council. Two questions each. Calibrated to 2024 GE, anchor-capped.

1.4pp avg gap on 103 cross-domain benchmarks.

Validation paper: https://t.co/9gEfRcNqbb

[Jason Duke @JasonD]

The other 126 races are too close for the 4-seed ensemble to call. Most pundits manufacture a winner from a 1pp margin. We don't.

That isn't weakness. It's the discipline that lets pre-registration mean something.

Polling papers over fragmentation. KPM-1 maps it.

[Jason Duke @JasonD]

The 5 Lab holds applied a documented manual override (LAB_GE_PREV_OVERRIDE), published openly before voting opened.

If those flip Lab-out anyway, the override hypothesis is proven wrong — and we publish that too.

Falsifiable both ways.

[Jason Duke @JasonD]

The 10 calls KPM-1 will be measured on:

Lab holds: Bradford 17pp · Harrow 16 · Sandwell 15 · Blackburn w. Darwen 13 · Wolverhampton 13 Con holds: East Surrey 10 · West Sussex 7 · Hart 6 Reform flips: NE Lincs 10 · Thurrock 7

[Jason Duke @JasonD]

Pre-registered before voting? ❌ polls ❌ markets ❌ pundits ✅ KPM-1 All 136 councils? ❌ ❌ ❌ ✅ Public SHA-256 hash? ❌ ❌ ❌ ✅ 103-benchmark validation? n/a n/a n/a ✅ Per-persona reasoning trace? n/a ❌ ❌ ✅ Public post-mortem? partial ❌ ❌ ✅
We did the it

[Jason Duke @JasonD]

KPM-1's call: historic three-way fragmentation.

Lab 62 pluralities. Reform UK 46 (from 0). LD 15. Con 11 (smallest EN footprint since 1979). Grn 2 (first-ever Green pluralities).

Most fragmented EN local result since the 1980s — captured in advance, hash-verified.

[Jason Duke @JasonD]

The hash is the receipt. No retro-fitting.

If KPM-1's first test verifies on 7 May, synthetic-panel modelling earns its place. If it misses, we publish what went wrong.

—Jason

https://t.co/hxaFzoFUOV

[Jason Duke @JasonD]

🗳 PRE-REGISTERED: our council-by-council predictions for the 7 May 2026 local elections.

136 councils. Every party. SHA-256 hash committed to GitHub on 1 May, before any ballot was cast.

📊 https://t.co/HpVH2lgwFA

[Jason Duke @JasonD]

Limits flagged:
• Manual override on Bradford, Sandwell, Harrow, Blackburn, Wolverhampton (all Lab restored — audit at /methodology)
• Reform over-prediction in strong-Leave seats
• LD floor uncertain where 2024 local data is patchy

https://t.co/1UfXVwOZGv

[Jason Duke @JasonD]

Honest framing: only 56 of 136 are confident calls (≥70% bootstrap win-prob). 16 leans. 64 noted three-way toss-ups within 2-3pp.
Hold us on the 72 confident calls. The 64 toss-ups are fragmentation indicators, not firm predictions.

[Jason Duke @JasonD]

Reform UK 58
Labour 43
Conservative 22
Lib Dem 11
Green 2
National share: Ref 26%, Con 21%, Lab 22%, LD 15%, Grn 13%.
Most fragmented local-elections result projected since the 1980s.

[Jason Duke @JasonD]

5/ Every other forecaster gives you numbers. We give you 12,224 individual reasons.

Full interactive results with reasoning traces for every council: https://t.co/LeTCL7yveO

Pre-registered. Verifiable. Open methodology.

[Jason Duke @JasonD]

We asked 12,224 synthetic voters WHO they'd vote for in the May 2026 local elections. Then we asked WHY.

The numbers: Reform 28.3%, Con 23.4%, Lab 20.1%, LD 18.4%

The story is in the reasoning.

https://t.co/Ib9PPTtaOH

[Jason Duke @JasonD]

4/ What voters actually talk about:
- Council tax: 63%
- Brexit/Reform: 65%
- Tactical voting: 36%
- Housing: 25%
- Cost of living: 23%
- Immigration: 2.4%

Immigration is 2.4%. Council tax is 63%. The gap between media coverage and voter reasoning is enormous.