Bank-CSV Month Close

Paste two account CSVs (date, description, amount, account) and close the month: ending balances, internal transfers matched and excluded, P&L by category, the unmatched list a human still has to eyeball — and a one-line work log of what the close actually did. Integer cents, deterministic, pure stdlib. The sample month is the fleet's own August 2026 books: 17 real on-chain x402 USDC settlements ($8.00) beside the ordinary Stripe, Fiverr, rent and transfer rows.

you paste — two CSVs, opening balances machine matches, categorises, closes, pins 100% offline

01 — Input

Paste the two account CSVs

Account A CSV (header: date,description,amount,account)

Account B CSV

Account A opening balance ($)

Account B opening balance ($)

Month (YYYY-MM)

Embedded sample

…

02 — The close

Month close

Ending balances

account	open	rows	close

P&L by category (unmatched, non-movement)

category	net

Matched transfers (internal — excluded from P&L)

date	from	amount	date	to	amount	lag

Unmatched rows (the list a human reviews)

date	account	description	amount	category

One-line work log

verified against committed report disagrees / refused

machine detail — full report JSON

Method

Rules (the same five the guard re-derives): (1) strict decimal amounts in integer cents — no floats; (2) strict ISO dates; (3) exactly two accounts; (4) a transfer = exact-cent mirror across accounts with 0–2 day lag, greedy, each row used at most once — matched pairs are excluded from P&L but never from balances; (5) category table, first match wins: stripe → revenue/stripe, fiverr → revenue/fiverr, x402|usdc → revenue/x402, coinbase|kraken|binance → movement/exchange (balance-sheet movement, excluded from P&L), rent|utilit|electric|internet → expense/overhead, else operating. Sample provenance: the 17 x402 settle * rows are the fleet's real August 2026 on-chain settlements (x402centre ledger, eip155:84532, $8.00); the rest of the month is constructed for the demo. The sample is byte-pinned in this page (sha256 manifest); the in-browser engine is line-compared against the committed report on load.