Rent Ledger NowCreate a ledgerHow the rent ledger parser works.
The first pass turns pasted notes, spreadsheet rows, or a CSV into reviewable ledger entries inside your browser. This page documents what the parser does, how we test it, and where human review is still required.
01 / Privacy boundary
Local parsing comes first
The default parser runs in the browser. It identifies candidate dates, amounts, transaction types, recurring ranges, and common CSV columns without sending the pasted ledger text to Rent Ledger Now, PostHog, or an AI provider. AI parsing is a separate, optional step that requires explicit permission.
Analytics records workflow metadata such as input mode, line count, and result count. It does not include names, addresses, amounts, or pasted ledger contents.
02 / Benchmark design
1,000 labeled cases, including a holdout
The deterministic corpus contains generated names, dates, amounts, and references rather than copied tenant records. Eight hundred cases guide development and every fifth case forms a 200-case regression holdout.
- Standard prose and terse landlord notes
- Several dated transactions on one line
- Payment references and running balances
- Reversals, NSF returns, fees, waivers, and concessions
- Conventional and split-column CSV exports
- Signed amounts, recurring ranges, and noisy text exports
03 / Exactness
What a passing case requires
A case passes only when the transaction date, balance effect, and amount all match the labeled result. The benchmark also measures entry-level precision and recall and treats unresolved blocking issues as failures.
The development and held-out cases currently pass exactly. This demonstrates regression coverage for the documented patterns; it does not demonstrate perfect accuracy on arbitrary real-world records.
04 / Known limits
Where the benchmark stops
The corpus does not yet establish accuracy for handwriting or OCR, non-English prose, every locale's debit and credit conventions, malformed spreadsheet quoting, or records whose accounting meaning is genuinely ambiguous. Always compare the result with the lease, receipts, bank records, and written adjustments before relying on the PDF.
The next meaningful evidence is a privacy-safe, opt-in set of redacted real failures that can be labeled and kept separate from development. Until that exists, AI remains optional rather than the default parser.
05 / Research basis
Representative public references
These sources informed terminology and document patterns. Their records were not copied into the benchmark.
- Rent ledger formMohave County Courts
- Returned and bounced payment terminologyCalifornia Department of Real Estate
- CSV debit and credit conventionsLedger CLI documentation