Proven core
You want to build your own, so this page shows where the prototype stops being one.
AI makes it sound easy. You can prototype a lending system in a week, with customers, a loan product, interest, fees and instalments, and it's tempting to ask how hard the rest can be.
The trouble is that the happy path is a tenth of the system. The edge cases are the killer, and they tend to arrive one at a time, in production, on a live book. We know the feeling. Our founder spent years on his own rewriting this engine from scratch, and without the people to finish it, it stayed a mile away from a commercial product.
Customers don't pay
You need arrears chasing with promise-to-pay tracking, a collections case with an owner, bad debt provisioning, write-off and recovery, and every one of those is a workflow, a screen and a set of postings.
Money arrives and you don't know what it's for
A deposit lands in your bank account with a reference nobody recognises, and now you need allocation rules, a suspense account and a case, a worklist for the people who work it, and an ending that is either a match or a refund.
Payments move on T+n
An electronic transfer isn't final when you send it. Return files arrive days later, and the debit order you marked as collected turns out not to have been. Fees and interest then recalculate from the date the money should have landed, and the customer's statement has to explain all of it.
The list goes on, and this is where a quick prototype turns into a nightmare in production, because each item ends up as a data fix on somebody's account.
All of the above runs in the demonstration tenants today, and each link opens the guide page a clerk would read.
| What a ledger must get right | What goes wrong when it does not | Status here |
|---|---|---|
| Value dating: trading date and value date on every movement, every property a date series, back-dated corrections re-accrued | A rate stored as a column; interest wrong by a day, on every account, discovered at month-end | Core, event-sourced, tested |
| Accrual and compounding across day counts and conventions | Rounding drift that never reconciles | Core, five conventions from the ISO code set |
| Day-end that can run forward and replay | Forecasts disagree with what actually posts | One code path, parity tested |
| A sub-ledger that balances per entity and per currency, and reconciles to your general ledger | Manual journals, FX result in the wrong place, a month-end that never ties | Core, posting rules configured, journal export |
| Four-eyes control on the operations that matter | A single clerk moves a limit | Deny by default, authorisation eligibility derived |
| Multi-currency revaluation and period close | A profit that disappears when the rate moves | Core, in the books |
| Schedules on irregular frequencies with multiple calendars | Payments on public holidays | Core, configured per account |
| Structures and legal entities | The deal you administer on a spreadsheet | The reason this engine exists |
Architecture tests
The build fails on a cross-feature reference, a static handler or an untyped id.
The rulebook →Permission catalogue
A command with no permission entry fails the build, so it can never become a hole.
Reachability gate
A handler that no test dispatches fails the gate.
Standards register
Every deviation from a standard is written up as a decision record.
Decision records
There are 883 of them, and together they are the design history of the engine.
A greenfield build gets you a ledger in months and the edge cases over years.
The edge cases are the product. The mistakes surface in month-end runs, audits and customer statements, and each one is a data fix on a live book. Twenty-five years in production at one of the world's safest banks produced a handful of true defects, and that record is the real asset. The code is just how it gets delivered.
You get the years along with the code, and your AI works inside the rules those years wrote.