KPM — Kronaxis Polling Method

Public, falsifiable UK election forecasting with cryptographic pre-registration.

This repository contains the code, methodology, hashes, and full audit trail behind every Kronaxis election prediction. Each prediction is committed to git with a SHA-256 hash before the event happens, so anyone can verify after results land that nothing was retroactively adjusted.

The companion public scorecard at https://kronaxis.co.uk/scorecard displays every prediction alongside its actual outcome — hits and misses with equal prominence.

Why this exists

UK polling and AI forecasting are full of claims that can't be verified after the fact. The KPM project is the inverse: every prediction is hashed and committed before the event, code is open-source, results are auto-scored against actuals, and misses are reported with the same prominence as hits.

This is the foundational pattern Kronaxis applies to every product — software vulnerability claims, behavioural simulations, system performance — but elections are the cleanest test case because outcomes are unambiguous and timing is fixed.

What's in this repo

Directory	Purpose
scripts/kpm2/	KPM-2.2 v15.1 — rule-based fragmentation override on KPM-1's vote shares. Hand-crafted, hash-anchored.
scripts/ward_data/	KPM-v17 ward-level methodology (v17.0 → v17.10 ensemble). Per-ward UNS + Reform-target detection + ONS Census + Hanretty Brexit. Honest negative results documented alongside the positive findings.
scripts/scorecard/	Public scorecard generator. Seeds + computes metrics + publishes JSON for the website to render.
scripts/byelection/	Continuous by-election engine. ALDC RSS ingest → v15.1 prediction → hash → scorecard append → outcome scoring.
scripts/llm_test/	The honest LLM-failure experiments. Three approaches tested, all failed, full results published.
data/kpm/	Hand-verified May 7 2026 actuals + historical NOC priors + full session gold log.
data/ward_data/	Per-ward election results from Democracy Club JSON API (36 councils with prior-history data + 4 Wikipedia-filled held-out councils + 26 pre-2024 priors for attempted cross-cycle backtest).
data/brexit2016/	Hanretty 2016 Brexit Leave % per constituency (Harvard Dataverse).
data/census2021/	ONS NOMIS bulk ward-level demographics (5 tables: age, ethnicity, religion, tenure, qualifications).
data/scorecard/	Source-of-truth scorecard JSON + computed metrics JSON.
data/byelection/	By-election calendar (auto-populated from ALDC RSS).
tests/	Reproducibility tests — verify both methodology hashes + reproduce the 59.2% v15.1 and 72.5% v17.10 numbers.

The KPM-1 → KPM-2.2 → KPM-v17 trajectory

• KPM-1 (separate repo: https://github.com/Kronaxis/kpm1-election-projections) — synthetic-panel forecasting via DYNAMICS-8 personas. Pre-registered before May 7 2026 with SHA-256 hash committed at https://github.com/Kronaxis/kpm1-election-projections/commit/.... Result: 28.5% on broader 130-council sample, ~PNS MAE 1.64pp on national vote share.
• KPM-2.2 v15.1 (this repo, scripts/kpm2/) — hand-crafted rule-based fragmentation override on KPM-1's vote shares. Six rules (LD-incumbent retain, NW Lab retain, historical NOC prior, LD strong-leader, Reform metro sweep, fragmentation NOC). Result: 59.2% on the same 130 councils, +30.7pp lift over KPM-1.
• KPM-v17 ward-level methodology (this repo, scripts/ward_data/) — per-ward UNS projection + Reform-target detection + ONS Census 2021 demographics + Hanretty 2016 Brexit Leave estimates. v17.10 production hash 2ea86b8d1e25ee68…. Result on n=40 sample (36 DC + 4 Wikipedia-held-out): 72.5%, tied with v15.1 at 72.5%. Bootstrap 95% CI on the difference is [-17.5, +17.5]pp — we cannot reject the null hypothesis that v17.10 = v15.1. The substantive structural finding: v17.10 has ~2× Reform UK recall (83% vs 46%) on the broader v15.1 sample, at comparable precision. Both methodologies have 0% recall for Conservative and Green council wins (14 of 130 cases — systematic blind spot).

Methodology hashes

v15.1 (council-fragmentation):

SHA-256: 52df676e792c29c6c893382a5c390c9b9790663e6e7e0a9c7edb1a54ac6c0741
Schema:  kpm2-fragmentation-v15.1
Verify:  python3 -c "import scripts.kpm2.rules as r; print(r.fragmentation_hash())"

v17.10 (ward-level ensemble):

SHA-256: 2ea86b8d1e25ee68ebf66c6f59496e2480a43dba2a81ff17012ce0094531f018
Schema:  kpm-v17.10-ensemble
Verify:  python3 -c "import scripts.ward_data.methodology_v17_10_ensemble as m; print(m.methodology_hash())"

Both hashes are frozen. Any rule change → new hash → new methodology row in the scorecard.

The honest LLM-failure record

Three LLM approaches were tested before settling on hand-crafted rules. All three failed. Full report in scripts/llm_test/SUMMARY.md.

Method	20-council holdout
Always-NOC (trivial baseline)	65.0%
KPM-2.2 v15.1 hand-crafted rules	55.0%
Approach 1: Gemini 2.5 few-shot (110 labelled examples)	50.0%
Approach 3: LLM rule mining	30.0%
KPM-1 LLM panel	25.0%
Approach 2: RAG with Wikipedia	1/5 (the one win was a Wiki post-event leak; honest 0/5)

The LLM has no useful signal for council-level UK election prediction. It remains useful for national PNS calibration (1.64pp MAE — competitive with major pollsters). Reported here so that future readers can challenge the conclusion or repeat the experiment.

Reproducing the published numbers

v15.1 (59.2% on n=130):

git clone https://github.com/Kronaxis/kpm.git
cd kpm
python3 -m pip install --quiet -e .  # no third-party deps; pure stdlib
python3 -m tests.test_backtest
# Expected: KPM-2.2 v15.1 = 77/130 = 59.2%

v17.10 (72.5% on n=40, tied with v15.1):

python3 -m tests.test_v17_10_backtest
# Expected: KPM-v17.10 ensemble = 29/40 = 72.5%
#           KPM-2.2 v15.1         = 29/40 = 72.5%
#           Difference            = +0.0pp (tied)

The v15.1 dataset (scripts/llm_test/dataset.json) is the merged 130-council set: 52 hand-verified actuals + 78 from a strict scrape of the Wikipedia 2026 results table, with hand-verified winning on overlap. The v17.10 dataset (data/ward_data/*_history.json) is real per-ward results from the Democracy Club JSON API, augmented with Wikipedia for 4 held-out councils where DC data is incomplete.

Honest negative results

In the spirit of falsifiable research, this repo also documents what didn't work:

• scripts/llm_test/ — three LLM approaches (few-shot, RAG, rule-mining) all under-performed always-NOC baseline. Full results published.
• scripts/ward_data/FINDING_n36_honest_plateau.md — v17.10 ties v15.1 at n=40. The earlier +25pp lift at n=15 was a Reform-heavy sampling artefact. Both methodologies have 0% recall for Conservative and Green council wins.
• scripts/ward_data/FINDING_2024_backtest_data_limit.md — cross-cycle 2024 backtest attempted on 26 metropolitan boroughs; 81% had ward boundary review 2022→2024 making ward-level retrospective validation empirically impossible for that cycle. The 5 stable-boundary cases were uninformative (Lab landslide year).
• scripts/ward_data/FINDING_v17_per_party_failure_modes.md — the methodology over-predicts Reform UK in Brexit-voting Lab strongholds with low prior Reform share (Wigan, Wolverhampton, Dudley false positives).

The credibility of any positive claim rests on these negative results being equally visible.

Falsifiability principles

1. Hash-and-commit before the event. Every prediction is in git history with a SHA-256 hash before the result is known.
2. Open-source the code. This repo. MIT-style permissions on the code; CC BY 4.0 on the methodology and data.
3. Auto-score against actuals. When results land, the scorecard updates automatically with hit/miss markers.
4. Misses with equal prominence as hits. The Hackney/Lewisham/Waltham Forest Green sweep miss is as visible as any hit. The 1/8 lean-track is reported alongside the 59.2% v15.1.
5. Methodology evolves in public. Each successor version (v14 → v15 → v15.1 → v16) shows its lift over the prior version transparently. Backtests are marked as backtests, never claimed as pre-registered predictions.

Next-up roadmap

• Week 4: Ward-level FPTP simulator (KPM-3). Per-ward swing model + Monte Carlo + seat counting. Target: 70-75% on broader council sample.
• Month 2: Parliamentary by-election predictor (MRP-light). Mayoral/PCC predictors (AV/SV simulators).
• Month 3: Scotland 2026 + Wales 2026 devolved elections. Same scorecard, same hash discipline.
• Ongoing: Every Thursday a UK council by-election happens. Every one gets predicted, hashed, scored. Within 12 months the scorecard will hold 250-300 predictions vs ~5 from any major pollster.

Licence

Code: BSL 1.1 (Business Source Licence) — converts to Apache 2.0 on 10 May 2031. See LICENSE. Data and methodology: CC BY 4.0.

Contact

• Project lead: Jason Duke jason@kronaxis.co.uk
• Issues: https://github.com/Kronaxis/kpm/issues
• Public scorecard: https://kronaxis.co.uk/scorecard
• Methodology paper: KPM-2 Methodology document forthcoming