Open methodology

Methodology 2026.1

Every rule, threshold and statistical test behind the findings. Versions are frozen once used, so any finding can be traced to the exact parameters.

Published: 27 Oct 2026
Parameters hash: f468a887e74d21e45f38d7a013ead418f10bda857ae499aa50d5d21f4aeccce1
Notes: First ElectraLens methodology. Thresholds are proposals, tuned on synthetic data only, and frozen once a run uses them.

Rules

Rule	What it checks	Parameters	Version
R01Exact duplicate entries	Identical entries in the same part.	—	1.0.0
R02Likely duplicate entries	Entries with similar names, the same relative’s name and close ages.	matching = phonetic_key_v1 max_age_gap = 2	1.0.0
R03Age outside the possible range	Ages outside the possible range on the qualifying date.	max_age = 110 min_age = 18	1.0.0
R04Age change inconsistent with time elapsed	Ages that changed more than the time between the rolls.	tolerance_years = 1	1.0.0
R05Unusual deletion rate	Deletion rate far above similar parts.	min_count = 20 min_effect = 2.0	1.0.0
R06Unusual addition rate	Addition rate far above similar parts.	min_count = 20 min_effect = 2.0	1.0.0
R07Part appeared or disappeared	Large transfers into one part, mostly from one other part.	—	1.0.0
R08Many electors at one address	Unusually many electors at one address.	percentile = 0.999 min_electors = 20	1.0.0
R09Deletions differ sharply between women and men	Women deleted at a very different rate from men.	min_count = 10 min_effect = 1.5 min_shift_per_1000 = 50	1.0.0
R10Missing or incomplete details	Many entries with missing details or deletions with no reason.	min_share = 0.05 percentile = 0.95	1.0.0
R11Unusual mix of deletion reasons	One deletion reason far more common than in similar parts.	min_count = 10 min_effect = 2.0 min_deletions = 20	1.0.0
R12Unusual number of first-time voters	Unusually many new 18–19 year old electors.	min_count = 10 min_effect = 2.0	1.0.0

Statistical rules compare each part with similar parts. A finding is reported only when it is both rare and large enough to matter.

Record-level checks (R01–R04, R07, R08, R10) report data quality, not a statistical signal.

Baseline: each part is compared with the other parts of its constituency. Unusual parts are trimmed before the baseline is estimated, so one extreme part can't hide another.
Spread: a beta-binomial model with a robust dispersion estimate allows for the normal variation between parts.
Many tests: every part is tested, so p-values are corrected with the Benjamini–Hochberg procedure (q-values) to control false discoveries at 0.05.
Uncertainty: rates come with Wilson 95% intervals. Gender comparisons use Fisher’s exact test; deletion reasons use a binomial test per reason with a Bonferroni correction.
Small parts: parts with fewer than 50 electors are not tested statistically.

Measured against the anomalies the generator injected, including innocent look-alikes. No accuracy target is claimed for real data.

No benchmark is attached to the latest publication.

Statistics show how unusual a pattern is, not why it happened.
Deletion reasons are used as recorded and are not checked independently.
Name matching across spellings is approximate.
Thresholds were tuned on synthetic data only and will be reviewed with real data.