Books ready the day the month ends

A signed-in human who is an owner or admin of the workspace. The close endpoint refuses API keys and service principals, the close commit is never exposed as a model or MCP tool, and the MCP close-flow tools deliberately have no commit. Automation can suggest that a period is ready. It cannot close it.

No. There is no reopen path: a correction is a new adjustment entry in an open period. Any posting aimed at a closed period halts, and the halt is recorded with its reason.

The close runs its gates in order and stops at the first failure. The named ones: source coverage is clean, no draft remains, every entry balances to the cent, every entry has at least one active line, and FX is finalized on every line. Drafts are validated in the same transaction as the lock, so a draft never jumps straight to locked.

A model reads the invoice and the candidate bank lines and returns, for each pair, a settlement probability between 0 and 1 plus one sentence naming the evidence it used. The evidence it is told to weigh is published: the amount, where a few percent of bank fees or FX spread still counts as a full settlement, the bank memo, the payment date against the invoice issue date, the direction of the movement, and whether the two counterparties already resolve to the same company record. It is also told that a shared amount, currency or date alone is never enough, and to return no match rather than guess. Two deterministic guards then run on that answer and can only lower it: a batch guard, when one payment is proposed against several invoices whose allocated amounts add up past what the payment was worth, and a party-consistency guard, when both counterparties are resolved and disagree with no invoice reference on the line to explain it. What survives decides the band. At 0.7 and above the payment links to the invoice. From 0.55 to 0.7 it waits for you. Below 0.55 it is dropped.

No. Well ships no writeback adapter for Xero or Pennylane: the register-diff gap poster registers QuickBooks only, and nothing in Well's own close or accounting code posts a journal entry to either ledger. No scope Well requests on either OAuth grant authorizes a journal write. The hand-off is the export: a FEC file or a CSV of a closed period.

A gapless legal sequence per fiscal year, one append-only audit row per close that locks entries (and per override) with the actor and the reason, and on every entry who validated it and when. For a closed period, a FEC file (France, every entry balanced, every account mapped) or a CSV.

Every category comes from one catalog of 46. When you correct the same counterparty to the same category on three distinct transactions, that becomes a rule and the classifier is no longer consulted for it. A correction to a different category resets the count. A category you set never carries a confidence, so your decision is never confused with a guess.

By email (Gmail and Outlook), by bank feed (Plaid and Qonto), from a collect task in the Chrome extension, from a file dropped in a session, or through the REST API and the MCP upload tools. Every channel writes to the same context graph.