Property documents filed to the right unit, or held for a person.
Docket reads incoming leases, invoices and notices, and matches each one to a known unit. It files a document on its own only when two independent clues agree. A language model can suggest a unit to a reviewer; it never files a document.
Watch a test document get filed or held
The recorded results did not load. Reload the page to try again.
Lease document Synthetic property-management fixture. Record: UNIT-0116 Address: 1510 Birch Street Apt 16 Building: BIRCH_06 Tenant: Paul Logan Tend tax sit later talk couple sound little pass control land.
Read the text
Find candidate units
Score the clues
Decide
Recorded held-out results, played back here. Nothing is uploaded.
Try it live: ask for an invite code- 28 of 28 automatic filings were correct in the held-out test.
- On held documents, the model's unit suggestions were less accurate than the rules' top pick. See the comparison
- All data is synthetic: 200 units and 180 held-out test documents. The live test missed 2 of its 4 gates. What it missed
How a document is decided
Docket looks for clues in the text that point at one of the known units. Each kind of clue carries a weight. A close spelling counts for part of its weight.
A document is filed on its own only when at least 2 different clues point at the same unit, and their weights add up to a score of about 0.586. If the clues name more than one unit outright (by unit number, full address or full tenant name), it waits for a person. A building code shared by many units does not count as naming one. One clue alone, even the unit number, is never enough, so a document cannot file itself just by naming a unit.
Every document files, waits for a person, or stops with a reason: a file that cannot be read at all (empty, encrypted, damaged or not a PDF) is rejected, and a document whose type question gets no answer from the model fails. Everything else that does not file waits on the review page, where a person sees the PDF, the clues, the rules' top pick and the model's suggestion side by side, and picks the unit or dismisses the document.
The model labels the document type and may suggest a unit. Its suggestion is shown to the reviewer and recorded in the trace; it is never used to file. If its type answer cannot be read, the document waits for a person, so the model can hold a document back but never file one.
| Clue | Weight |
|---|---|
| Unit number in the text | 0.30 |
| Street address | 0.30 |
| Tenant name | 0.25 |
| Building code | 0.10 |
| Unit number in the file name | 0.05 |
Held-out results
Measured on 180 synthetic test documents that played no part in choosing any setting: 28 filed, 112 held for a person and 40 rejected as unreadable files. "Rules only" is the same filing policy with no model at all.
| Docket | Rules only | |
|---|---|---|
| Automatic filings that were correct | 28 of 28 | 38 of 38 |
| Documents that should file, filed | 28 of 39 | 38 of 39 |
| Held documents | Count | Model right | Rules' top pick right |
|---|---|---|---|
| Belong to a known unit; model asked. Right means naming that unit. | 62 | 3 of 62 | 42 of 62 |
| Belong to no unit; model asked. Right means naming none. | 20 | 20 of 20 | 0 of 20 |
| No candidate unit, so the model was not asked | 20 | not compared | not compared |
| The model's suggestion call failed | 10 | not compared | not compared |
What it missed. The model's safety filter refused every very long test document: 40 of 440 answers across both test sets, against a limit of 2%. Those documents wait for a person instead of filing, so in the held-out set Docket filed fewer of the documents it should have than the rules alone. That missed 2 of the live test's 4 gates. Nothing was filed to a wrong unit. The full results and both missed gates
Terms and method
- Held-out test
- Two sets of generated documents. The tuning set, 180 documents, chose the filing threshold, 0.5857142. The held-out set, 180 documents, chose nothing. Every other figure on this page comes from the held-out set, except the live-test gates, which count the model's answers across both.
- Model
- Claude Haiku 5.5 on the Anthropic API, answers recorded once on 2026-11-13 and replayed with no model call.
- Precision and recall
- "Correct automatic filings" is precision; "should file, filed" is recall.
- Uncertainty
- The report gives a 95% Wilson interval for its overall figures, such as filing precision, recall and suggestion accuracy; the split rows in the suggestions table have none. These figures measure synthetic documents, not real mail.
The review page
Held documents wait in one queue. Each shows the PDF, why it is waiting, every candidate unit with its score, the rules' top pick and the model's suggestion. A reviewer approves either, picks another candidate, or dismisses the document with a reason. Every step is in the document's trace.
It needs a reviewer token. To try it, run Docket locally with the quickstart, then open the review page.