Published: 2026-05-03
Comparative analysis across 12 elections (2009-2026): flash-memory anomalies, turnout, protest vote, risk neighborhoods and polling.
Update (2026-05-12): see the follow-up article for new aggregation tools added since this piece — the composite Election Risk Index, per-section risk screening (/risk-score), and per-party Benford 2BL test (/benford). All three are reachable from the new/risk-analysisdashboard.
Dataset: 12 elections, 2009-07-05 through 2026-04-19. All metrics extracted from /public/<date>/ — national summary, problem sections, suspicious settlements, section reports and polls.
Glossary of indicators used below:
suemg_removed) — sections where the official paper protocol records fewer votes than what the machine flash drive logged.suemg_added) — the inverse: the protocol records more votes than the flash drive.suemg_missing_flash) — sections that ran machines but submitted no flash drive at protocol time.suemg) — sections where the flash drive and the protocol agree on the total machine vote, but the per-party split differs (votes moved between parties).supports_noone) — valid votes for the "I support no one" option (protest vote).additional_voters) — voters added to the supplementary list on Election Day.concentrated) — top party's share in a section/settlement.problemSections) — sections in the eight tracked risk neighborhoods.| metric | 27.10.2024 | 19.04.2026 | delta | 04.2026 votes impacted |
|---|---|---|---|---|
| Voter turnout | 38.81% | 50.70% | +11.89 pp (largest jump in the series) | — |
| Actual voters | 2,568,992 | 3,360,330 | +791k | — |
| Registered voters | 6,619,877 | 6,627,747 | +8k (rolls essentially unchanged) | — |
| Machine vote share | 38.04% | 47.61% | +9.57 pp | — |
| Top party | GERB-SDS 26.39% / 66 seats | PB 44.59% / 131 seats | first single-party majority in the series | — |
| New entrant top finish | — | PB (zero in prior election) | first time a new party wins outright | — |
| Sections with at least one anomaly | 3,353 | 3,331 | -22 | — |
| Sections with "protocol above flash" discrepancy | 126 | 88 | -38 | 190 (vs 330 in 10.2024) |
| Sections with "protocol below flash" discrepancy | 2,685 | 3,077 | +392 (all-time high) | 11,979 (vs 9,050; +2,929) |
| Sections with reallocated votes between flash and protocol | 158 | 226 | +68 | 338 (vs 194; +144) |
| Sections with completely missing flash drive | 33 | 62 | +29 | 9,104 (vs 3,146; ×2.9) |
| Sections in risk neighborhoods | 137 | 138 | +1 | 32,544 (vs 25,191; +7,353) |
| Suspicious settlements — concentration >80% | 218 | 145 | -73 (lowest since 2017) | 18,537 top-party (vs 38,719; -20,182) |
| Suspicious settlements — invalid >10% | 57 | 87 | +30 | 1,670 invalid ballots (vs 814; +856) |
| Suspicious settlements — additional voters >10% | 225 | 429 | +204 (series high) | 9,666 day-of additions (vs 5,346; +4,320) |
Sections in risk neighborhoods are the single largest integrity-flagged bucket — 32,544 votes, ~0.97% of national turnout — even though the section count (138) is one of the smallest in the table. The next-largest buckets are settlements where the top party took >80% of the vote (18,537 top-party votes — though down from 38,719 in 2024, the biggest absolute drop on the board) and the "protocol below flash" discrepancy (11,979 votes — only ~0.37% of turnout despite affecting 24.2% of machine sections, much smaller than the section share suggests).
Two flags grew faster in vote terms than in the count of affected sections and settlements: completely missing flash drive tripled (3,146 → 9,104 votes, ×2.9) on a doubling of sections (33 → 62), meaning the affected sections are larger on average; day-of list additions >10% added 9,666 votes (vs 5,346, +81%) on roughly double the settlement count. By contrast, reallocated votes between flash and protocol is essentially negligible in vote terms — 338 votes across 226 sections — a real but cosmetic signal.
Section-share-wise, the "protocol below flash" trend is monotonic since flash-drive auditing began — share of all machine-running sections:
The vote impact translates this into ~12,000 votes in 2026 — a process-quality issue worth investigating, but well below outcome-changing thresholds for any party in the result.
| election | total | protocol above flash | protocol below flash | missing flash | problem sections |
|---|---|---|---|---|---|
| 2009 - 04.2021 | 125-136 | 0 | 0 | 0 | 125-136 |
| 07.2021 | 188 | 0 | 1 | 53 | 135 |
| 11.2021 | 177 | 9 | 7 | 27 | 135 |
| 10.2022 | 154 | 0 | 0 | 19 | 135 |
| 04.2023 | 2,148 | 288 | 1,710 | 37 | 135 |
| 06.2024 | 2,432 | 309 | 1,917 | 92 | 137 |
| 10.2024 | 3,353 | 126 | 2,685 | 33 | 137 |
| 04.2026 | 3,331 | 88 | 3,077 | 62 | 138 |
Pre-2023 zeroes are absence of comparable flash-drive data, not absence of anomalies — CIK began publishing comparable flash datasets only from 2023.
International context: Bulgaria's machine-flash-vs-paper-protocol reconciliation is a direct audit trail. Most countries lack the equivalent and have to rely on statistical fingerprints — the Klimek-Yegorov-Hanel-Thurner method (PNAS 2012) plots section-level turnout against winning-party share looking for clusters near 100%/100%, which appear in elections widely flagged as fraudulent (Russia 2011-12, Uganda) but not in clean ones (Switzerland, Finland, Spain). The Bulgarian flash audit is more granular than that statistical signature — and a 24% protocol-vs-flash mismatch rate is the kind of process-quality indicator most democracies cannot even measure.
| party | sections where it is the biggest single loser | net votes lost in those sections (protocol − flash) |
|---|---|---|
| PB (winner) | 1,418 | −4,156 |
| PP-DB | 614 | −1,582 |
| GERB-SDS | 234 | −350 |
| MECh | 165 | −197 |
| Vazrazhdane | 117 | −136 |
| DPS | 80 | −125 |
The winning party is the largest net loser from the protocol-vs-flash reconciliation. The pattern tracks machine vote share: machine-heavy parties get shaved proportionally. PB's 4,156-vote shave is 0.29% of its 1,444,920 total — far from outcome-changing, and the wrong direction for any "fraud favouring the winner" hypothesis.
Largest single-section discrepancies in 2026:
All three discrepancies hit the parties that won the section — direction-wrong for any fraud hypothesis.
| election | mean section turnout | sections >100% | >95% | >90% | <5% |
|---|---|---|---|---|---|
| 2017 | 67.34% | 628 | 830 | 916 | 5 |
| 11.2021 | 84.89% | 1,275 | 1,516 | 1,617 | 55 |
| 04.2023 | 99.05% | 1,117 | 1,336 | 1,402 | 30 |
| 06.2024 | 110.91% | 1,087 | 1,334 | 1,400 | 49 |
| 10.2024 | 126.70% | 1,039 | 1,250 | 1,319 | 30 |
| 04.2026 | 75.34% | 987 | 1,228 | 1,321 | 19 |
Despite a +11.89 pp national turnout rise, sections with turnout above 100% fell to 987 — the lowest since 10.2022. A fabrication scheme would normally show the opposite pattern.
International context: the Klimek et al. (PNAS 2012) "election fingerprint" methodology treats section-level turnout >95% combined with high winning-share as the canonical ballot-stuffing signature — present in Russia 2011-12 and Uganda 2011, absent in Switzerland, Finland, Spain. In Bulgaria these high-turnout sections have stable, structural explanations (mobile sections, hospitals, abroad sections with day-of registration, small remote villages). The fact that the >95% section count decreased from 1,250 to 1,228 while national turnout rose by 11.89 pp moves Bulgaria 2026 farther from the Klimek fraud signature, not closer.
Settlement-level day-of additions partially contradict this: 429 settlements with additional voters >10% of actual (vs 225 in 10.2024) — series record. Top case: с. Vehtino (Smolyan) at 161.82%, mathematically possible only for very small sections with high churn but worth flagging individually. Concentration in Smolyan, Kardzhali and Vratsa oblasts.
| election | mean section "support no-one" % | sections >5% | >10% |
|---|---|---|---|
| 2014 (option introduced) | 5.28% | 4,448 | 1,135 |
| 10.2022 | 3.50% | 2,449 | 82 |
| 04.2023 | 4.16% | 3,862 | 145 |
| 10.2024 | 3.41% | 2,075 | 80 |
| 04.2026 | 1.74% | 84 | 1 |
*Methodology note: the per-section mean and the >5% / >10% thresholds are computed by the same formula the section-level report uses — 100 × ⌈ (noOne ÷ totalActualVoters) × 100 ⌉ ÷ 100 — i.e. each section's percentage is rounded up to the nearest whole percent before counting and averaging. Without the ceiling rounding, the means would be ~0.3-0.4 pp lower, but the downward trend is the same.*
Sections with significant protest vote (≥5% "no one") fell from 2,075 to 84 — a 96% reduction. Combined with the +11.89 pp turnout rise, the most parsimonious read is previously demobilised voters returned with a positive choice — PB absorbed the protest vote. This pattern is hard to fabricate.
International context: the same "protest-absorbed-by-new-movement" pattern appeared in Italy 2013 (Five Star Movement) where blank/null ballots and abstention dropped sharply as M5S took 25.6%; in France 2017 (Macron's En Marche absorbed both Socialist and Republican defectors and depressed blank ballots in round 1); and in Slovakia 2020 (OĽaNO surge). The fingerprint of fraud is the opposite: when protest votes are erased numerically (e.g. by reclassification or ballot substitution) without a movement-party surge, you typically see falling protest vote alongside falling turnout — not rising turnout, as Bulgaria 2026 shows.
In Bulgarian electoral analysis, the eight tracked risk neighborhoods are conventionally used as a fairness barometer: they have the highest historical incidence of organised vote-buying, ballot stuffing, and coerced "controlled" voting. Three signals are typically watched together — invalid-ballot rate, machine-vote share, and same-day list additions. Persistent extremes on these signals indicate a captured electorate; convergence toward national averages indicates a freer one.
International context: the OSCE/ODIHR final report on Bulgaria's June 2024 elections documented that the Ministry of Interior conducted raids in these specific neighborhoods explicitly to fight vote-buying, planned to publish a list of "risky polling stations" largely overlapping with the same areas, and opened 88 vote-buying investigations by the eve of election day — independent confirmation that the fairness concern in these specific neighborhoods is well-founded and predates this analysis. The academic literature on Bulgarian vote-buying (Mares & Young 2019, Conditionality & Coercion, Oxford UP) likewise treats these neighborhoods as the canonical case of broker-driven turnout in EU member states.
| neighborhood | sections | 2017 | 04.2021 | 10.2024 | 04.2026 | Δ vs 2024 |
|---|---|---|---|---|---|---|
| Stolipinovo (Plovdiv) | 71 | 44.2% | 37.3% | 28.9% | 31.6% | +2.7 pp |
| Fakulteta (Sofia) | 10 | 27.0% | 26.2% | 6.1% | 19.1% | +13.0 |
| Filipovci (Sofia) | 5 | 42.4% | 58.9% | 32.2% | 40.7% | +8.5 |
| Nadezhda (Sliven) | 8 | 38.4% | 22.9% | 5.3% | 10.4% | +5.1 |
| Pobeda (Burgas) | 8 | 37.8% | 27.5% | 18.2% | 20.7% | +2.5 |
| Gorno Ezerovo (Burgas) | 3 | 48.0% | 45.6% | 21.8% | 29.0% | +7.2 |
| Dolno Ezerovo (Burgas) | 7 | 48.4% | 45.0% | 27.7% | 46.8% | +19.1 |
| Maksuda (Varna) | 26 | 51.6% | 46.9% | 37.2% | 53.5% | +16.3 |
All 8 saw turnout increases. Maksuda and Dolno Ezerovo are back at or above their 2017 baselines — a 19 pp and 16 pp jump respectively in a single cycle. Whether that reflects genuine remobilisation or organised mobilisation is the question the next three sub-sections address.
Vote-buying schemes typically pay voters to deliver a photographed marked ballot and then tear up a substitute, or pay them to spoil their ballot intentionally. Both produce elevated invalid rates — a long-known pattern in these neighborhoods. National 2026 rate: 3.87%.
| neighborhood | invalid % 10.2024 | invalid % 04.2026 | × national 2026 |
|---|---|---|---|
| Fakulteta | 12.69% | 17.95% | 4.6× |
| Nadezhda Sliven | 12.31% | 14.17% | 3.7× |
| Gorno Ezerovo | 12.94% | 11.48% | 3.0× |
| Pobeda Burgas | 10.58% | 10.17% | 2.6× |
| Filipovci | 7.83% | 9.38% | 2.4× |
| Stolipinovo | 11.29% | 7.74% | 2.0× |
| Dolno Ezerovo | 3.94% | 5.93% | 1.5× |
| Maksuda | 2.70% | 3.20% | 0.83× (below national) |
The picture is mixed. Stolipinovo improved sharply (-3.5 pp), and Maksuda is now actually below the national invalid rate. But Fakulteta got worse (12.7% → 17.9%) — almost 5× the national average, the worst reading among the tracked neighborhoods in this dataset. Five of eight remain at 2-4× national, indicating that paid-vote and coerced-spoil patterns persist where they have always been concentrated, even as turnout rose.
The voting machine produces a paper receipt and a tamper-evident flash record — both of which are independently auditable. Paper-only sections cannot be cross-checked against flash data. Higher machine share in a captured area therefore limits the available manipulation surface. National 2026: 47.61% machine.
| neighborhood | machine % 10.2024 | machine % 04.2026 | shift |
|---|---|---|---|
| Maksuda | 55.2% | 66.0% | +10.8 pp |
| Dolno Ezerovo | 49.2% | 61.4% | +12.2 |
| Stolipinovo | 33.6% | 54.3% | +20.7 |
| Gorno Ezerovo | 27.0% | 47.6% | +20.6 |
| Filipovci | 30.6% | 39.6% | +9.0 |
| Pobeda Burgas | 17.0% | 27.7% | +10.7 |
| Nadezhda Sliven | 20.0% | 26.6% | +6.6 |
| Fakulteta | 26.7% | 24.8% | -1.9 |
Three neighborhoods (Maksuda, Dolno Ezerovo, Stolipinovo) are now above the national machine share — which is unusual; historically these neighborhoods skew heavily paper. This is a substantive integrity improvement: a higher fraction of these votes is now auditable end-to-end. Fakulteta and Nadezhda Sliven remain heavily paper (75-77% paper), and they are also the two with the worst invalid rates — paper-heavy and spoil-heavy.
Election Day list additions (additional_voters) are how patron-managed voters are typically processed: they aren't on the standing roll, they show up with a paid courier, get added on the spot, and vote. Persistent high additions in a low-turnout area are a strong fairness flag.
| neighborhood | additions % 10.2024 | additions % 04.2026 | shift |
|---|---|---|---|
| Nadezhda Sliven | 13.99% | 7.28% | -6.71 |
| Fakulteta | 9.96% | 4.80% | -5.16 |
| Pobeda Burgas | 6.93% | 5.27% | -1.66 |
| Gorno Ezerovo | 5.60% | 4.71% | -0.89 |
| Stolipinovo | 3.80% | 4.29% | +0.49 |
| Maksuda | 2.53% | 2.12% | -0.41 |
| Filipovci | 2.16% | 2.90% | +0.74 |
| Dolno Ezerovo | 1.60% | 1.82% | +0.22 |
Despite turnout rising, **same-day additions are down in 5 of 8 neighborhoods**, with two large-magnitude declines (Nadezhda Sliven -6.7 pp, Fakulteta -5.2 pp). This is the opposite of what controlled-mobilisation would produce — the additional turnout was largely from voters already on the roll, not from courier-driven on-the-day registrations.
Captured-electorate signatures usually show one party with an outsized share. National PB share: 44.59%.
| neighborhood | top party 10.2024 (share) | top party 04.2026 | PB pp above national |
|---|---|---|---|
| Stolipinovo | GERB-SDS (41.5%) | PB 50.5% | +5.9 pp |
| Fakulteta | GERB-SDS (27.2%) | PB 53.7% | +9.1 |
| Filipovci | DPS–NB (29.3%) | PB 40.7% | -3.9 |
| Nadezhda Sliven | DPS–NB (36.4%) | PB 59.7% | +15.1 |
| Pobeda Burgas | GERB-SDS (44.5%) | PB 40.2% | -4.4 |
| Gorno Ezerovo | GERB-SDS (44.9%) | PB 56.5% | +11.9 |
| Dolno Ezerovo | GERB-SDS (28.3%) | PB 60.1% | +15.5 |
| Maksuda | GERB-SDS (29.6%) | PB 49.4% | +4.8 |
Two patterns coexist:
The neighborhoods where PB looks anomalously strong are also the neighborhoods where the structural fairness signals (invalid, paper share) are most degraded. The most parsimonious read is that whatever local broker network historically ran turnout in Nadezhda Sliven, Dolno/Gorno Ezerovo, and Fakulteta has migrated its allocation from the previous patron parties (GERB-SDS, DPS–NB) to PB. The headline national result is not driven by these four — they account for ~5,000 votes in total — but they are the cells where the cleanest fraud-style signature shows up, and they should be the focus of any post-election integrity review.
One caveat: of the four, only Dolno Ezerovo has a long-running pattern of swinging heavily toward each new movement party — ITN in 04.2021 (25.9%) and 07.2021 (36.5%), PP in 11.2021 (31.8%), now PB. Its 60% PB share in 2026 is partly consistent with ordinary movement-party adoption rather than a clear broker pivot. The other three — Gorno Ezerovo and Fakulteta voted GERB-SDS top in seven straight elections (04.2021 through 10.2024) and Nadezhda Sliven oscillated GERB-SDS / DPS without ever picking a movement party — making their abrupt 2026 pivot to PB harder to explain without a broker-network shift.
Final polls for 19.04.2026 — averaged across all agencies whose fieldwork ended within 14 days of election day (06.04 - 16.04.2026). Six agencies fall in this window: Market Links, Trend, Alpha Research, CAM (Center for Analysis and Marketing), Myara, and Sova Harris.
Methodology note. Two of the six agencies (ML, CAM) publish raw electoral attitudes with an explicit "undecided" share rather than an already-redistributed forecast. To compare apples-to-apples against the official ЦИК result (which reports vote share among ballots cast), the /polls leaderboard rebases those agencies' shares by proportionally redistributing the undecided share across the named parties — share × (1 + undecided / Σshares) — before scoring. The poll-mean column below uses these normalised values; raw published numbers are noted where they differ materially.
| party | poll mean | actual | error |
|---|---|---|---|
| PB | 35.35 | 44.59 | +9.24 pp |
| GERB-SDS | 19.86 | 13.39 | -6.47 |
| PP-DB | 12.15 | 12.62 | +0.47 |
| DPS | 9.89 | 7.12 | -2.77 |
| Vazrazhdane | 7.02 | 4.26 | -2.76 |
| BSP-UL | 4.09 | 3.02 | -1.07 |
PB's raw published shares (before rebase) were lower: ML 30.8%, TR 33.2%, AR 34.2%, CAM 32.1%, MY 34.6%, SH 33.6% — mean 33.08%, giving a raw miss of -11.51 pp. The proportional rebase above absorbs the explicit "undecided" pools (~14-17% each) from ML and CAM into the named parties, which shrinks the apparent PB miss to -9.24 pp but enlarges the GERB-SDS overshoot. ML's own published forecast for this cycle (column G in Market LINKS's results summary — median of four prognostic models with expert weighting, used directly in the table above since ML's genre tag is forecast) put PB at 38%, the closest single-agency reading; CAM's mechanically-rebased 38.51% landed nearly identical. The take-away is the same regardless of methodology: all six pre-election polls materially under-pegged the new entrant — by 6.08 pp (CAM) at the closest, by 10.99 pp (SH) at the farthest.
For comparison, final polls for 27.10.2024:
| party | poll mean | actual | error |
|---|---|---|---|
| GERB-SDS | 25.26 | 26.39 | +1.13 |
| PP-DB | 16.98 | 14.21 | -2.77 |
| Vazrazhdane | 14.18 | 13.36 | -0.82 |
| DPS–NB | 11.08 | 11.51 | +0.43 |
| BSP | 8.42 | 7.57 | -0.85 |
In 2024, every party landed inside ±2.8 pp; in 2026 the winner deviated -9.24 pp (after rebase) and GERB-SDS deviated +6.47 pp in the opposite direction.
The mean absolute error (MAE) across the parties shown:
The largest meta-analysis of pre-election polls — Jennings & Wlezien, "Election polling errors across time and space" (Nature Human Behaviour, 2018), 30,000+ polls in 351 elections across 45 countries (1942-2017) — finds the eve-of-election MAE at ~2 percentage points internationally. The American Association for Public Opinion Research notes that real-world error tends to run closer to 3.5 pp once non-sampling sources are included, roughly double the textbook ±3 pp sampling margin for an n=1,000 poll.
So the 2024 BG aggregate MAE (1.20 pp) was unusually good — better than the international historical average. The 2026 BG aggregate MAE (3.80 pp) is roughly at the real-world baseline — not actually exceptional in aggregate.
What is exceptional is the dispersion: a -9.24 pp miss on PB (after rebase; -11.5 pp on raw published shares) and a +6.47 pp miss on GERB-SDS pulling in opposite directions. Comparable single-party misses in major recent elections:
| election | party | poll | actual | miss |
|---|---|---|---|---|
| BG 2026-04 (post-rebase) | PB (new) | 35.4% | 44.6% | +9.2 pp |
| BG 2026-04 (raw published) | PB (new) | 33.1% | 44.6% | +11.5 pp |
| France 2017 round 2 | Macron | +22 lead | +32 lead | +10.0 pp |
| Italy 2013 | Five Star Movement (new) | ~20% | 25.6% | +5.6 pp |
| US 2016 (state polls avg) | Trump | — | — | 5.2 pp |
| US 2016 (national polls avg) | Trump–Clinton margin | — | — | 3.1 pp |
| UK 2010 | Lib Dems (overestimate) | ~28% | 23% | -5 pp |
The PB miss (between +9.2 and +11.5 pp depending on methodology) sits in the same range as the 2017 Macron round-2 underestimate (+10 pp) and the 2013 Five Star Movement underestimate (+5.6 pp). Both of those were also new-or-movement parties peaking late, and both produced equivalent national hand-wringing about polling failure that turned out, on review, to reflect known methodological weaknesses rather than fraud.
Likely explanations for the 2026 miss:
A polling miss of this magnitude is not evidence of fraud — Jennings & Wlezien's headline finding (summarised by the University of Southampton press office: "claims of a crisis in the accuracy of election polling are false") is that single-party 5-10 pp misses occur in roughly one in ten national elections studied, almost always when a new party surges. The 2026 BG miss fits the well-documented "new-movement-party" failure mode rather than indicating something irregular happened on Election Day.
PB won 31 of 32 oblasts. Exception: Kardzhali — won by DPS at 56.1%.
PB shares:
No oblast where PB wins with a patron-driven >65% share. The most-suspicious-on-its-face number is the low central Sofia figure, not a high one — the expected anti-populist district.
| election | mean | sections >5% | >10% | >25% |
|---|---|---|---|---|
| 2014 | 6.62% | 6,745 | 1,917 | 120 |
| 10.2024 | 3.13% | 2,244 | 376 | 16 |
| 04.2026 | 3.70% | 2,990 | 489 | 21 |
Modest uptick vs 10.2024, well within historical norms — half the 2014 peak. Top extreme cases all involve very low total-vote sections (small-N noise).
International context: invalid-ballot rates are a well-established vote-buying proxy in the comparative-politics literature. Stokes, Dunning, Nazareno & Brusco's Brokers, Voters, and Clientelism (Cambridge UP, 2013) documents the mechanism: paid voters are required to deliver a spoiled or photographed ballot to the broker as proof, which inflates section-level invalid rates 2-5× the national baseline. Bulgaria's 3.7% national rate is in line with EU peers (Romania 2-4%, Slovakia 3-5%, Italy 2-3%); the risk-neighborhood multiplier of 2-5× national in this dataset matches the literature's predicted signature in captured cells almost exactly.
The 19.04.2026 result reads, on the integrity metrics this dataset captures, as a genuine but historically unusual surge rather than a manipulated outcome.
supports_noone collapsed -96% — protest voters absorbed by the winning party.Listed in order of votes impacted, with section/settlement counts as secondary context. Total turnout was 3,360,330 — a useful denominator for scale. Each item also flags the top beneficiary — i.e. the party that captured the most votes in the affected units, with raw counts so the partisan tilt of each signal is visible.
The ~9.2-11.5 pp miss on the winner (depending on whether one applies the proportional-redistribution rebase) is a forecast failure, not an integrity finding, but it is the largest in the series and warrants a methodology post-mortem from polling agencies — particularly turnout modelling, incumbent anchoring, and how to weight a brand-new entrant that polled at zero in the previous cycle.
This analysis was generated by Claude Opus 4.7 (Anthropic, May 2026), working from the structured datasets in /public/<date>/ — national_summary.json, problem_sections.json, dashboard/suspicious_settlements.json, the per-section reports under reports/section/*, and the polls dataset under polls/. Every numeric claim in the article is computable from those files using the queries in the underlying repository; section anomaly counts follow the definitions in scripts/reports/nationalSummary.ts (the union of recount, recountZeroVotes, suemgAdded, suemgRemoved, suemgMissingFlash, and problemSections per election).
External references (academic, OSCE/ODIHR, polling-agency archives) were retrieved via web search at time of writing and are linked inline at point of use. No new survey, interview, or scraping work was performed for this article — it is entirely a re-analysis of pre-existing public datasets through the lens of the international election-forensics and polling-error literature.
Reproduction notes: the polling-MAE figures are computed by averaging the four April 2026 final polls per party and subtracting the official result; section-level distributions (>5%, >10%, >100% etc.) are direct counts from the corresponding reports/section/*.json files.