Platform
Fast enough to change how your team works. Careful enough that they’ll trust it.
Every figure your team sees has already been read, recomputed independently, and checked against the rest of the document it came from. If a number can’t be backed up, it never reaches a reviewer’s desk — and nothing leaves your organisation until someone on your team says it can.
How it’s built
Reading is one job. Arithmetic is another. We never let one pretend to be the other.
A language model is extraordinary at reading a messy document and drafting language — and it is the wrong tool to trust with a sum. So the platform is built in layers, each with exactly one job, and the layer that reads never gets to also be the layer that decides a number is correct.
- 01
Ingestion
Documents arrive exactly as your team already produces and receives them — statements, filings, confirmations, contracts. Nothing is reformatted before it is read; the original is what gets checked.
- 02
Reading
A language model reads the document and drafts a first pass — the figures it found, the commentary it would write, the terms it extracted. This layer only ever transcribes and drafts. It never has the final word on a number.
- 03
Validation & computation
Every figure is recomputed independently, in plain deterministic code, and cross-checked against the rest of the document — arithmetic, cross-statement agreement, matching labels and periods. If the two layers disagree, the disagreement is what surfaces, not a silently corrected number.
- 04
Human sign-off
Nothing produced by the layers above reaches anyone outside your team until a named person on your side approves it. That approval is the release — not a formality bolted on afterwards.
Underneath all four, one more layer runs constantly rather than as a step of its own: every extraction, check, computation and approval is written to an audit ledger the moment it happens.
Sign-off
Nothing leaves without a name attached.
Every release — a reconciliation, a statement review, a credit memo, or any workstream we add after them — passes through approval by someone on your team before it goes anywhere. That approval records who signed off, when, on which draft, and what produced it, written to a record that cannot be edited afterwards. Authority sits exactly where it already does today: whoever could approve a reconciliation break or a statement analysis before this platform existed is the same person who approves it now. There is no new kind of machine authority to define, and no new signature to grant.
It keeps working when a provider doesn’t
Reconciliation, computation and every check keep running with no AI provider reachable, because none of them depend on one. Only reading and drafting queue until service is restored. The platform gets slower. It does not get less accurate.
A record that survives an audit
Every extraction, validation, computation and approval is written to an append-only, hash-chained ledger. Altering a past entry invalidates every entry written after it — which is what makes the record usable by an internal auditor, rather than merely available to one.
Stated as plainly as what it will do
What it will not do.
This is the more useful list to anyone signing off on this — a boundary we hold to, not a feature we haven’t built yet.
- It will not move money, initiate transactions, or hold credentials that can.
- It will not write to your systems of record.
- It will not release a figure without a person approving it.
- It will not replace your core operational systems.
- It will not read a document it was not given.
- It will not produce a number it cannot show the working for.
See it against your own documents.
The clearest way to judge this is to bring a real case and watch it work.