Read it once. Know which answers to check.
Two clerks and a hundred forms
Two readers with the same accuracy are not worth the same. The one that tells you which answers to check has given you your afternoon back.
Two clerks read the same hundred forms. The first is right 95 times and swears by all 100. The second is right 95 times and hands you 16 forms marked "look at these", with a finger on the exact line. Only the second one has given you your afternoon back. Same accuracy; one of them is a product.
Anyone can rent a model that reads documents at about 94%. That is a commodity, priced by the page. What a business cannot buy off the shelf is accountability: for every value the system filled, whether it may be used without a person, and if not, where on the page the person should look.
The job is sorting, not reading
A reader that says honestly "these 84 are certain, these 16 need your eyes" removes 84% of the work at the same accuracy.
Consider a form with a hundred fields. A reader that is right on 95 of them and equally confident about all 100 has not reduced the reviewer's work, because the reviewer cannot tell the 95 from the 5 without checking all 100.
The same reader, if it could say honestly "these 84 are certain, these 16 need your eyes", would have removed 84% of the work. The second system is worth more than the first at the same accuracy. Its value is in the sorting.
The paper that offices read all day and type into systems is the target: tax forms, invoices, statements, filings. The number that matters to the office is reviewer minutes per document.
Read, point, judge
Every value comes back with a rectangle on the page and one of three verdicts: accept, accepted after a second reading, or a person should look.
The system does three things, and each is measured. Reading means answering a question about a page. Pointing means returning the rectangle on the page where the answer was read, so the reviewer's eye lands on the evidence rather than on a blank form. Judging means attaching one of three verdicts to every value.
A citation is a rectangle in page coordinates around the printed words a value was read from. A verdict is one of accept, accepted after a second reading, or a person should look. Everything that follows serves one claim: that the verdicts are measured, not asserted.
Two benchmarks, because one does not exist
Reading is scored on the public DocVQA benchmark. No public benchmark scores a citation or a verdict, so one was built.
DocVQA is the public benchmark for answering questions about document images: 5,349 validation questions over scanned forms, letters, tables and reports. The score is ANLS: full credit for an exact match, partial credit for a near match, nothing for an answer that differs from the gold answer by more than half its characters. We report it over every validation question with nothing removed, and beside it the stricter share of answers identical to the gold answer.
No public benchmark scores a citation rectangle or a trust verdict, so we built one: generated business documents (invoices, statements, forms, tables) in which every field's value, location and arithmetic relations are known. It gives a composite of two parts, ground (does the rectangle cover the right words) and trust (was the verdict right), with a wrong auto-fill penalised far more than an unnecessary review.
The discipline behind the numbers
One change at a time, judged by a rule written before the first run, and decided by questions nobody designed against.
Every change to the system was one experiment: a single change, run against the fixed benchmark, and kept or discarded by a rule written down before the first run.
For reading, a change was kept only if it improved the development split, improved the 2,865 held-out questions, and the chance that the gain over all 4,365 scored questions was luck fell below ten percent. Discarded experiments stay in the log beside the kept ones.
Every paid model call passes through a cache with a ledger, so a repeated run is free and the cost of every result is known.
Reading at human level
97.53 ANLS on all 5,349 questions. The best published entry scores 97.29; people score 98.11.
Over all 5,349 validation questions the system scores 97.53 ANLS, with nothing removed. 98.35% of its answers are accepted as matching, and 93.56% are identical to the gold answer.
For context, the best published entry on the official test leaderboard scores 97.29, and human performance is 98.11 ANLS, or 94.36% exact match. The system is level with the best published systems and within half a point of people.
By split, against the 2025 system: development 95.42 to 97.74 on 1,500 questions, holdout 94.80 to 98.04 on 1,201, and 96.72 on a second holdout of 1,664 that the earlier system was never run on.
The two numbers a buyer asks for
93.3% of fields filled with no person at 100.00% precision, and 99.5% of citations on the right ink.
In the terms a buyer uses: 93.3% of fields are filled with no person in the loop at 100.00% precision. That is zero wrong auto-fills in 1,181 fields, from 83.6% coverage with two wrong auto-fills at the baseline.
99.5% of citations land on the right ink: six wrong rectangles and none missing, from 106 wrong rectangles at the baseline.
The composite grounding score rose from 79.1 to 93.5 on the development documents and from 79.7 to 90.9 on held-out documents, over 36 kept changes. The two parts moved together: trust from 81.3 to 92.9 and ground from 76.9 to 94.1. Seven audits on freshly generated documents were all positive, from +4.7 to +9.0 points.
No larger model
Not one point came from a bigger model. The reader sets the ceiling, disagreement is a signal, and the holdouts decide.
None of the gains came from a larger model. Several larger models, tried as readers, scored below the system rather than above it. The points were earned by the engineering around the model, and each was confirmed on questions no change was designed on.
The reader sets the ceiling. Everything built around a reader can only recover points the reader itself nearly had. Disagreement is a signal, not a failure. When independent readings of the same page differ, the difference is the system telling the reviewer where to look.
Design on training questions, confirm on unseen ones. Ideas chosen by inspecting the hardest development questions failed on unseen questions every time they were tried. Rules designed on the training split carried over every time.
What it costs, and where it stops
Under half a cent per question. The grounding benchmark is generated, and the reading results are on the validation set.
At the reported accuracy the reading pipeline costs under half a cent per question, with every call cached so that a repeated question costs nothing. The system records the cost and the latency of every document it processes.
The limits belong on the same page. The grounding benchmark is generated, not collected, and its development score sits 2.6 points above its holdout score, which suggests some fitting to the development documents. The reading results are on the validation set, as all public comparisons for this benchmark are, and not on the hidden test set.
The wording normalisation is fitted to the benchmark's convention of the shortest printed answer and would be re-fitted for a customer's schema.
What is public, what is not, and where to go next
The measurements are public. The method is in a companion paper, shared on request.
This work is proprietary. What was measured, on which questions, and why the measurements are credible is public, here and in the paper.
It does not describe the readers, the settings, the normalisation rules, the verdict thresholds or the 36 changes to the grounding engine. A companion paper does, with every kept and discarded experiment and the cost ledger, and is available on request.
Step 2: the paper