2023 presidential electionNationwide auditPublished 24 September 2026

BVAS or phones? What metadata in Nigeria’s 2023 presidential result sheets shows

Mundx inspected 164,659 accessible polling-unit result records from all 36 states and the FCT. We found the EMP2920 model associated with INEC’s BVAS in 52,947 records, another identifiable device model in 50,452, and no recoverable camera model in 61,260.

The finding measures camera metadata in the published result image. It does not, on its own, identify who owned the camera or whether the metadata was edited.

Headline numbers

Nearly half of identifiable records carried a model other than EMP2920

The broadest denominator is every result record whose linked document could be opened and inspected. Records without a recoverable camera model remain visible instead of being assigned to either device group.

164,659Polling-unit result records successfully inspected93.1% of the 176,846-record inventory
52,947Records carrying EMP2920/BVAS camera metadata32.16% of inspected records
50,452Records carrying another identifiable model30.64% of inspected records
61,260Inspected records with no recoverable camera model37.20% of inspected records
EMP2920 / BVAS52,947 records · 32.16%
Other identifiable model50,452 records · 30.64%
No recoverable model61,260 records · 37.20%
Among the 103,399 records where a model was recoverable, the split was 51.21% EMP2920 and 48.79% other models.

The other-model total is an upper bound for phones. It also includes some tablets and unresolved labels, while the no-model group cannot be classified from the available metadata.

INEC’s prescribed workflow

INEC told Presiding Officers to use BVAS for the IReV upload

Clause 38(ii) of INEC’s 2022 Regulations and Guidelines prescribed BVAS for uploading the scanned polling-unit result, Form EC8A, to the INEC Result Viewing Portal.

The 2023 Manual for Election Officials reinforced this in sections 2.9.0–2.9.2. Its online process begins on the BVAS homepage and directs the official to photograph and upload the result. The manual’s offline workflow also runs on the BVAS, retaining the submission until connectivity is available.

The original paper EC8A still travelled to collation. BVAS was therefore the prescribed device for the digital IReV copy, not the whole result-management chain.

Other identifiable devices

Infinix, TECNO and Samsung dominate the non-EMP2920 metadata

The five largest manufacturer groups account for 42,949 of the 50,452 known non-EMP2920 records, or 85.1%. These are record counts, not counts of individual physical phones.

Infinix13,387
TECNO13,251
Samsung11,025
Xiaomi3,181
Huawei2,105
OPPO1,728
itel1,331
Nokia / HMD969
vivo867
Apple584

The ten most frequent non-EMP2920 model strings

RankRecorded manufacturerRecorded modelResult recordsExample evidence
1InfinixInfinix X688B1,12932/09/07/001 · Rivers →
2InfinixInfinix X657C87132/03/02/004 · Rivers →
3SamsungSM-A135F82132/06/01/002 · Rivers →
4TECNOTECNO KD773932/03/05/002 · Rivers →
5TECNOTECNO KG7h58502/02/01/009 · Adamawa →
6InfinixInfinix X6511B57132/06/11/001 · Rivers →
7InfinixInfinix X655C56529/19/06/007 · Osun →
8TECNOTECNO CG656232/05/01/004 · Rivers →
9TECNOTECNO KC854832/15/05/019 · Rivers →
10InfinixInfinix X68353612/08/03/015 · Edo →

The audit inventory contains 805 normalized manufacturer/model combinations derived from 802 distinct model strings. A small number of model strings appeared under more than one manufacturer label.

Geographic distribution

Non-EMP2920 records appeared nationwide, with Lagos contributing the largest share

Lagos accounted for 6,285 known non-EMP2920 records, or 12.46% of the national total. The five largest states together accounted for 30.64%; every state and the FCT had at least one.

Lagos6,285
Kano2,593
Bauchi2,256
Plateau2,165
Katsina2,161
Delta2,053
Niger2,048
Edo1,906
Kaduna1,876
Akwa Ibom1,617
The distribution was not concentrated in one state.

Lagos was the largest single contributor, but 87.54% of known non-EMP2920 records came from the other 36 state/FCT geographies.

Votes in the known non-EMP2920 subset

APC recorded the largest vote total in polling units with known non-EMP2920 metadata

Mundx matched all 50,452 known non-EMP2920 polling units to its presidential result-observation layer. Reportable OCR values were available for 46,611 units, covering 92.39% of the subset; 3,841 units remained unavailable.

2,016,277Bola TinubuAPC · 32.72% of all recorded votes
1,712,193Atiku AbubakarPDP · 27.79% of all recorded votes
1,611,328Peter ObiLP · 26.15% of all recorded votes
352,788Rabiu KwankwasoNNPP · 5.73% of all recorded votes
APC2,016,277
PDP1,712,193
LP1,611,328
NNPP352,788
The four candidates account for 5,692,586 of the 6,161,562 recorded votes in this reportable subset.

These are polling-unit research totals, not an official collation. Only 811 units had reportable values without OCR warnings; 45,800 had reportable values with caveats. Candidate totals sum recorded numeric cells, while blank cells remain missing instead of being filled with zero.

Camera-clock dates

Known non-EMP2920 records peaked on 28 February 2023

A camera date was recoverable for 49,962 of the 50,452 non-EMP2920 records. Only 4,162 were dated on election day; 45,787 carried a date after 25 February. The largest daily count was 14,179 on 28 February.

These are editable camera clocks, not authenticated upload timestamps.

The date may reflect the original photograph, a later rephotograph, or an incorrectly configured device clock. It should not be read as proof of when the result first reached IReV.

Method and limits

A complete inventory audit, with unknowns kept separate

The saved national inventory contains 176,846 polling-unit records. Every record was accounted for; device conclusions were drawn only from files that opened and retained usable metadata.

01 · Inventory

176,846 polling-unit records across all 36 states and the FCT, from a saved IReV inventory snapshot extracted on 3 September 2026.

02 · Accessible files

164,659 linked result records were successfully opened and inspected. Another 9,403 had no document listed, 2,721 contained no usable document URL, and 63 remained inaccessible.

03 · Metadata extraction

The audit read EXIF and embedded image metadata from PDFs and direct images. Full-size image counterparts were checked when thumbnails had stripped camera fields.

04 · Validation

The range-reading method reproduced make, model, software, capture time, identity fields and dimensions across 6,093 complete local files with zero mismatches.

05 · Counting rule

Counts refer to polling-unit result records carrying a model label. They are not counts of distinct physical devices, owners or uploads.

06 · Evidence limit

EXIF can be edited or removed. A non-EMP2920 label identifies the image-producing camera or rephotograph, not necessarily the device that transmitted the file to IReV.

07 · Vote join

All 50,452 known non-EMP2920 polling-unit codes matched the Mundx presidential result-observation layer. No geographic or name-based fuzzy matching was used.

08 · Vote coverage

Vote totals use 46,611 units with reportable OCR values. The 3,841 unavailable units and every blank candidate cell remain excluded rather than being converted to zero.

The result files remain hosted at their source URLs. Mundx publishes counts, representative links and machine-readable summaries so readers can inspect the basis of the analysis.

Frequently asked questions

What the device metadata can—and cannot—answer

Did INEC require BVAS for uploading 2023 polling-unit results to IReV?

INEC’s operational instructions did. Clause 38(ii) of the 2022 Regulations and Guidelines told Presiding Officers to use BVAS to upload a scanned copy of Form EC8A to IReV. The election manual provided online and offline upload procedures on the BVAS.

How many unique device models appeared in the metadata?

The audit found 802 distinct recorded model strings across the 103,399 records where a model could be recovered. Counting normalized manufacturer-and-model pairings produces 805 combinations because a few model strings appeared under more than one manufacturer value.

Does a non-BVAS model prove that a personal phone uploaded the result?

No. It shows that the published image contains camera metadata associated with another model, subject to the possibility of rephotography or edited metadata. EXIF does not identify the network device that transmitted the file or the camera’s owner.

Which non-BVAS phone models appeared most often?

The leading non-EMP2920 model strings were Infinix X688B, Infinix X657C, Samsung SM-A135F, TECNO KD7 and TECNO KG7h.

Was the non-EMP2920 pattern concentrated in one state?

No. Lagos had the largest count at 6,285, representing 12.46% of the known non-EMP2920 total. Every state and the FCT had at least one known non-EMP2920 record.

Which major candidate recorded the most votes in the known non-EMP2920 subset?

Across the 46,611 units with reportable OCR values, APC recorded 2,016,277 votes, PDP 1,712,193, LP 1,611,328 and NNPP 352,788. These are research-layer totals from recorded numeric cells, not an official collation.

What is EMP2920?

EMPTECH identifies EMP2920 as its BioTab8 biometric tablet. INEC’s election manual shows the EMP2920 test application on the election device used in the BVAS workflow.

Explore the underlying records

Follow the result trail from national patterns to polling units.

Open the 2023 presidential map, review Mundx’s method, or browse the source-linked election record.