Changelog
Including what was taken out. A changelog with no reversals in it is a changelog nobody believes.
v0.6
2026-08-10The period close, and an export that refuses
- AddedPeriod close as a checklist of counts rather than a readiness score, with every count linking to the documents behind it.
- AddedExports return 409 while the period has blocking items. A file that silently omits four unposted documents looks complete and gets filed.
- RemovedThe readiness percentage. Ninety-four per cent ready is not a state anybody can file from, and a number nobody can act on trains people to ignore the screen.
v0.5
2026-07-28Posting autonomy per supplier
- AddedDraft → suggest → auto, earned per supplier and per account from that pairing's own record.
- AddedDemotion on the first correction, with the reason recorded and shown on the ladder.
- ChangedLedger entries no longer store whether they balance. Debit and credit totals are computed on read, so an entry that does not tie out cannot claim it does.
v0.4
2026-07-06The comparison layer
- AddedEight checks that compare a document against everything else held: duplicate, bank account, delivery match, price drift, arithmetic, identifiers, due dates, missing delivery note.
- ChangedThe review queue sorts by money at risk rather than by extraction confidence. A 97%-confident bank account that changed outranks a 61%-confident reading of a twelve-euro line.
- RemovedTwo checks that fired on nothing across all five sample workspaces. Every check now shows its own hit count beside its toggle, so this stays visible rather than becoming a thing somebody has to remember to verify.
v0.3
2026-06-15Two readers instead of one
- ChangedEvery document now goes through two extraction engines from different vendors. Two runs of the same model agree, confidently, on the same wrong digit.
- AddedSettlement: arithmetic, check digits, known values and template memory, tried in that order before a person is asked. Each field records which one settled it.
- FixedOIB and IBAN validation now runs the real MOD 11,10 and MOD 97 algorithms rather than a length test. Measured on the fixtures: 693/693 and 612/612 single-digit corruptions rejected.
v0.2
2026-05-20Refusing rather than guessing
- AddedDocuments that cannot be read are refused with a reason and an instruction the person holding the phone can follow, instead of being filled with the best guess available.
- ChangedMoney is stored as integer cents throughout. Not one amount in the product is a floating-point number.
v0.1
2026-04-28First working pipeline
- AddedIngest from e-mail, upload, scanner and phone; classification, extraction and line items; the document queue.
- AddedWhatsApp as a first-class channel through EnterChat, with the conversation attached to the document.