The full report
Tested on real documents. Every count, every mistake.
Takarda read two real bank statements and a public company report, and every result was counted against an answer key taken from each document's own text. These are the figures of the version tested. What went wrong is listed beside what went right.
- real documents
- 4
- pages read
- 87
- transactions counted
- 983
Headline results
Four figures, each with its count.
A figure is always k of n: how many were right, out of how many were counted. The note under each one says what it counted and where it fell short.
983 of 983
Transactions found
On two real bank statements, 48 pages. None missed, none invented.
Counted against the answer key's transactions.
1,966 of 1,966
Amounts and balances right
Money in or money out, and the running balance, on every row.
1,638 on the 36-page statement exactly; 328 on the 12-page statement by size (92 kept their printed minus sign)
1,358 of 1,365
Boxes in the right place, 99.5%
When Takarda shows where a value sits, it points to the right spot.
But it showed a box for only 1,365 of 3,932 values. The rest said Not on the page.
19 of 19
Fields right on one report
A public report of 37 pages. Each field had its box on the page, and the check that total liabilities plus total equity equals total assets passed.
Read in 1 min 42 s from upload to ready.
982 of 983
Dates on the right day
Transaction dates on the two statements.
944 of 983
Descriptions word for word
96.0%; the misses are listed under each statement.
The three documents
One section for each, mistakes included.
Two bank statements that belong to the founder, shown with every character blurred, and a public quarterly report of a listed Nigerian bank. Nothing a person could be identified by is published here: only counts, rates, times and costs.
Document 1 of 4
A fintech wallet statement. 819 transactions, 36 pages.
Two months, two accounts in one file (a wallet and a savings account), phone-bank PDF.
- Pages
- 36
- 819 transactions
- Upload to ready
- 14 min 8 s
- Cost
- ₦1,768
- about $1.11; about ₦49 a page
| Measure | Result | Share | What it counts |
|---|---|---|---|
| Transactions found | 819 of 819 | 100% | transactions in the answer key0 missed, 0 invented. |
| Amounts right to the kobo | 819 of 819 | 100% | money amountsMoney in or money out, to the kobo. |
| Running balances exactly right | 819 of 819 | 100% | balances after each transaction |
| Money in and money out swapped | 0 of 819 | 0% swapped | transactions (lower is better) |
| Dates on the right day | 818 of 819 | 99.9% | transaction dates |
| Descriptions word for word | 781 of 819 | 95.4% | transaction descriptions28 of the 38 misses differ only in spacing; 6 lost the first line where the row starts at the bottom of a page; 4 have misread characters. |
| Account summary values right | 10 of 10 | 100% | summary values at the top of the statement |
| Boxes in the right place | 1,064 of 1,071 | 99.3% | boxes shown, against the answer key's place for that value |
| Values that had a box | 1,071 of 3,276 | 32.7%A weak spot | values; the rest said Not on the pageA weakness of the version tested: most right values had no box and said Not on the page. |
Every character is blurred on purpose. The coloured boxes are where Takarda found each value.
Document 2 of 4
A naira account statement from a payments app. 164 transactions, 12 pages.
Nine months, newest first, one signed amount column, phone-bank PDF.
- Pages
- 12
- 164 transactions
- Time reading
- 2 min 51 s
- Cost
- ₦293
- about $0.18; about ₦24 a page
| Measure | Result | Share | What it counts |
|---|---|---|---|
| Transactions found | 164 of 164 | 100% | transactions in the answer key0 missed, 0 invented. |
| Amounts right in size | 164 of 164 | 100% | money amounts92 money-out amounts were stored with their printed minus sign. |
| Running balances exactly right | 164 of 164 | 100% | balances after each transaction |
| Dates on the right day | 164 of 164 | 100% | transaction dates |
| Descriptions word for word | 163 of 164 | 99.4% | transaction descriptions |
| Account summary values right | 10 of 10 | 100% | summary values at the top of the statement |
| Boxes in the right place | 294 of 294 | 100% | boxes shown, against the answer key's place for that value |
| Values that had a box | 294 of 656 | 44.8%A weak spot | values; the rest said Not on the page |
Every character is blurred on purpose. The coloured boxes are where Takarda found each value.
Document 3 of 4
A public quarterly report of a listed Nigerian bank. 19 fields, 37 pages.
Interim consolidated accounts, 37 pages, amounts in millions of naira; read with a field list of 19 fields written for it (no ready list).
- Pages
- 37
- 19 fields
- Upload to ready
- 1 min 42 s
- Cost
- ₦175
- about $0.11; about ₦5 a page
| Measure | Result | Share | What it counts |
|---|---|---|---|
| Fields right | 19 of 19 | 100% | fields in the field list written for this report |
| Fields with a box on the page | 19 of 19 | 100% | fields |
| A check that adds up | Passed | Not a count | Total liabilities plus total equity equals total assets. |
Read from a picture of one page
Two single pages were also given as pictures. Every value on the page was right, and every field the page does not hold said Not on the page.
| Page given as a picture | Values on the page right | Fields not on the page |
|---|---|---|
| Page 2, the income statement page as a picture | 12 of 12 | 7 of 7 said Not on the page |
| Page 3, the balance sheet page as a picture | 10 of 10 | 9 of 9 said Not on the page |
Each tile is one field of the field list written for this report. A tick means the value was right and had its box on the page.
A public document read as a test. The bank's name, logo and page pictures are not shown here.
Document 4 of 4
Two shop receipts photographed on a phone. 10 item rows, every value right.
A supermarket and a body care shop in Abuja: thermal paper with pen marks across the shop names, creases and shadows, taken with an iPhone.
- Photos
- 2
- 10 item rows
- Upload to ready
- About 20 s
- each receipt
- Cost
- ₦28
- about $0.02 for both
| Measure | Result | Share | What it counts |
|---|---|---|---|
| Printed values right | 10 of 10 | 100% | shop, receipt number, date, total and how it was paid, on each receipt |
| Not printed, said Not on the page | 8 of 8 | 100% | Tax ID, subtotal, VAT and currency, which neither receipt prints |
| Item rows found | 10 of 10 | 100% | rows of things bought, none missed, none invented0 missed, 0 invented. |
| Item cells exact | 40 of 40 | 100% | quantity, item, unit price and amount on every row |
| Values with a box on the photo | 50 of 50 | 100% | every value read, header and item rows |
The photos as Takarda straightened them. Each coloured box is where it found a value: yellow for the shop, date, receipt number, total and payment, the other colours for the item rows.
The screens
The screens the figures were read and checked on.
These are real screens of the working product, not drawings: what Takarda is asked for, the box beside every value, and the table it returns. The statements above are never shown here beyond their blurred pages.
The receipt photos are from a public set: CORD v2, NAVER Clova, CC BY 4.0.
How it was measured
The answer comes first, from the document itself.
Nobody marks Takarda's work by eye alone. The answer key is made from the document's own text, proved against the totals the document prints, and checked twice, before a single Takarda value is compared.
An answer key from the document's own text
Built from each PDF's own text, with the position of every character, never from Takarda's reading of it.
Proved against what the document prints about itself
Transaction counts, totals, opening and closing balances and every running balance had to agree with the key.
Checked by two independent checkers
Two checkers compared the key with the page pictures. Neither found a problem.
Every cell compared by value
Takarda's rows were matched to the key's rows by date, amount and balance. Then each cell was compared on its own.
How each cell was compared
- Money
- To the kobo.
- Dates
- By the day.
- Descriptions
- Word for word, after spacing is set aside.
- Boxes
- Against the answer key's place for the same cell.
- Rows
- Matched by date, amount and balance.
Changed since this test
Figures of the version tested. Later changes are not counted.
Takarda has changed since these documents were read. The changes below have not been measured again, so none of them is counted in any figure on this page.
Most right values had no box and said Not on the page.
More boxes on statement rows.
Two accounts in one file were checked as if they were one.
One check for each account.
Dates printed with a time of day were not recognised as dates.
A change to how dates with a time of day are read.
Money out amounts kept the minus sign the statement prints.
Amounts without their minus sign.
Reading calls ran out of room on dense pages, and were paid for twice.
Fewer wasted reading calls.
A description that runs over a page break lost its first line.
A change to descriptions that run across a page break.
Measured in public
Real receipt photos, a public set.
A second test, on a public set of real receipt photos that anyone can download. It tests where Takarda puts the box for each value. It is not a score for every kind of document.
96.8%
Box on the right words
On 200 real receipt photos, the box sat on the words it points to.
1,342 of 1,387 boxes
likely between 95.2% and 97.8%
100%
Missing totals reported as missing
Where a receipt printed no total, Takarda said Not on the page every time.
181 of 181 values
likely between 98.3% and 100%
73.0%
Values given a place
The rest are marked Not on the page. Nothing is guessed to fill the gap.
1,387 of 1,900 values
likely between 69.8% and 76.0%
2.7%
Boxes on the wrong text
A box that touched none of the words it should have marked.
38 of 1,387 boxes
likely between 1.8% and 4.2%
| Measure | Words as printed | With 6% of the numbers misread |
|---|---|---|
| Values given a place on the page | 73.0% (1,387 of 1,900 values; likely between 69.8% and 76.0%) | 70.8% (1,346 of 1,900 values; likely between 67.6% and 73.9%) |
| Box on the right words | 96.8% (1,342 of 1,387 boxes; likely between 95.2% and 97.8%) | 95.8% (1,289 of 1,346 boxes; likely between 94.1% and 97.0%) |
| Box on the wrong text | 2.7% (38 of 1,387 boxes; likely between 1.8% and 4.2%) | 3.3% (45 of 1,346 boxes; likely between 2.3% and 4.9%) |
| Missing total reported as Not on the page | 100% (181 of 181 values; likely between 98.3% and 100%) | Not measured |
200 public receipts with their annotated words. Credit: CORD v2, NAVER Clova, CC BY 4.0.
Read a document of your own.
Name the fields you want. Takarda returns one row for each file, with the box on the page beside every value.