What Ledger does not claim
The limits, stated as plainly as the results.
A tool built to stop people quoting unearned confidence has to hold itself to the same rule. So here is exactly where Ledger is not yet measured, and what the guarantee does and does not cover.
- What does “labelled data” mean, and why do you need it?
- Labelled means you give us the correct answers for a sample of your documents — about 20 to start — so accuracy can be measured on your data, not a benchmark. It is a hard requirement: without it the pipeline refuses to auto-accept anything, because a guarantee it cannot measure would be exactly the unearned confidence Ledger exists to prevent.
- Do I need a developer to run it?
- To run the pipeline yourself, yes — it is delivered as code (a CLI, JSON/CSV output, and for higher tiers a Docker image and a FastAPI service) that a developer runs in your environment. If you would rather not, the Production tier is delivered as a running service. A pilot needs almost nothing from your side beyond the documents and their labels.
- Where are my documents processed?
- Extraction runs on your own Claude subscription, or on an SDK backend you control for production — there is no paid third-party OCR or vision service in the path. The Production tier adds PII handling, redaction and retention controls. Exact data-handling for your setup is agreed during the pilot.
- Does Ledger read handwriting?
- Not yet measured. The evaluation corpus is degraded printed text. We do not claim a handwriting result we have not measured.
- Can Ledger handle multiple documents on one page?
- Not yet supported.
- Are line items (tables) certified?
- Not yet. On CORD, cell-level exact match is 94.9% and whole rows entirely correct is 84.2%, but line items are delivered as reviewed output, never straight-through. We state this plainly because hiding it would be the one thing the product exists to prevent.
- What does the guarantee actually cover?
- The fields Ledger auto-accepts, at your chosen risk level α, measured on a held-out slice of your own documents. It is a certificate you can check, not a marketing number.
These limits are not a disclaimer bolted on. A tool built to stop confident, wrong data has to be the first to say where it isn’t sure — so it is.