Drop in an Indian bank statement PDF. Ithamar reads every row and checks it against the statement's own running balance. A row that holds is proven. A row that does not is explained, repaired only when the repair is earned, or handed to you. It all happens in this tab; nothing is uploaded.
23 rows19 proven3 repaired1 left for youClosing balance 4,04,878.87 matches
The balance above, plus what came in, minus what went out, equals the balance printed. Ithamar walks that line down the page the way an auditor runs a ruler down a ledger. When it holds, the row is ticked as proven. Nothing is guessed.
₹64.27 was read as money coming in. The balances say it went out. Move it to the debit column and this row closes, and so does the next. A missing debit of ₹128.54 would also balance it, but only by coincidence, so that alternative is logged and set aside.
The debit came out as 42,701.59. The balances on either side need 42,706.59: a single 1 that should be a 6. That is the kind of mistake text extraction actually makes, so it is the only kind of amount Ithamar will correct.
When a balance itself is misread, its row breaks and so does the row below, because the next row starts from the wrong figure. Ithamar recognises the pattern and corrects the one cell. Both rows close. The corrected row is marked repaired, not proven: those are different claims.
Between the 16th and the 18th, ₹1,76,090.10 arrived and no row records it. A missing row can explain any break, which is exactly why it proves nothing. The gap and its amount are handed to you, printed VOID the way a cheque marks a copy.
The closing balance matches. Nineteen rows are proven as printed, three are repaired with every change logged, and one is left for a person. A tool that quietly smooths that last row over is not auditing anything.
Across 24 generated statements, 855 rows and 39 injected errors, in three layouts modelled on Indian banks. Every swapped column, misread digit and misread balance was repaired to its true value. Every missing row was flagged and left alone.
These statements come from Ithamar's own generator, so their truth is known exactly. That proves the verifier. Real bank PDFs are the next test, and the result will be published the same way, failures included.
Your file is read inside this tab. It is never uploaded; there is no server to upload it to. When it is done, your statement gets its own seal.
This is the sum of the tabernacle, even of the tabernacle of testimony, as it was counted, according to the commandment of Moses, for the service of the Levites, by the hand of Ithamar, son to Aaron the priest.
Exodus 38:21, King James Version
The first audit in the record was kept by Ithamar, and its silver can still be checked. Exodus 38 counts 603,550 men giving half a shekel each, and separately reports the total as 100 talents and 1,775 shekels. At 3,000 shekels to the talent, the two records agree.
The same engine ships as a Python command. It exits 0 when every row is proven, 1 when rows were repaired, and 2 when a person needs to look, so it can guard a script.
pip install git+https://github.com/victorisaacchinta-eng/ithamar
ithamar audit story.pdf --csv rows.csv
ITHAMAR 0.2.0 story.pdf
ASHOKA LEDGER BANK
✓ 3 2026-04-06 UPI/746814538401/PAYMENT/imran.sharma@ 12,837.61 2,74,589.36
↻ 4 2026-04-07 UPI/307666193451/PAYMENT/lakshmi.kumar 64.27 2,74,525.09
✓ 5 2026-04-07 NEFT CR-SHFL0323205-KONARK LOGISTICS-S 4,738.18 2,79,263.27
...
✓ 23 2026-04-23 UPI/333123214616/RENT/asha.patel@upi/A 53,094.00 4,04,878.87
↻ row 4: direction swapped. 64.27 is a debit, not a credit
considered and rejected: a debit of 128.54 may be missing above this row
↻ row 9: amount misread. debit reads 42,701.59; the balances need 42,706.59
considered and rejected: a debit of 5.00 may be missing above this row
↻ row 14: balance misread. balance reads 2,84,383.97; the chain gives 2,84,382.97
✗ row 19: missing row suspected
a credit of 1,76,090.10 may be missing above this row
19/23 proven as printed, 3 repaired, 1 broken, 0 unchecked. Closing balance matches.
NEEDS REVIEW