Sheaf
Sheaf vs general-purpose LLMs
Point GPT-5, Claude or Gemini straight at a PDF and you get a fluent answer. Nothing in that answer tells you whether the page supported it.
General models are superb readers and unusable as a system of record. Sheaf uses them deliberately narrowly: a model reports what a document says about itself, and a deterministic engine does the matching, the filing and the coverage arithmetic. The parts you would want to audit are the parts that do not vary between runs.
Sheaf vs a general model at a glance
| Requirement | Sheaf | GPT-5, Claude, Gemini |
|---|---|---|
| Where geometry comes from | The reader — text layer, or optical reading | The model’s own description |
| Citations | File, page and span, checkable | Prose references |
| Same file run twice | Deterministic matching and filing | Answers can vary between runs |
| When the page is silent | Recorded as not stated, distinctly | A fluent answer either way |
| Split a mixed bundle | Determined boundaries that tile the file | No boundary contract |
| Missing documents | Named by requirement, recomputed live | Whatever you thought to ask |
| Cross-document consistency | Names, dates and amounts checked automatically | Whatever you thought to ask |
| Straight-through processing | Per-requirement thresholds you control | Yours to build |
| Filing and approval | Ships as the product | Yours to build |
| Audit trail | Every outcome walks back to a page and a person | Yours to build |
Where Sheaf wins
- Geometry comes from the reader Sheaf never asks a model for coordinates or page numbers. The text layer supplies them where one exists and optical reading supplies them where it does not, and the answer is bound to that geometry. A value that cannot be located on a page does not become a citation.
- Abstention is a feature When the page does not state something, Sheaf records that the page does not state it — distinctly from unread and from not applicable. An empty field is ambiguous, and those three cases lead to three different actions.
- The deciding logic is deterministic The model contributes facts. It does not get a vote on the outcome. Send the same file twice and you get the same boundaries, the same filing and the same coverage number.
- It is a product, not a prompt Filing, approval, withdrawal, live completeness, cross-document validation, an exception queue and an audit trail. None of that comes out of a chat window, and all of it is what an examiner asks for.
Where a general model is strong
- Reading comprehension Genuinely excellent and improving fast. Sheaf is built on language models precisely because they read well.
- Flexibility Ask anything, in any shape, with no schema. For exploration that is exactly right.
- Nothing to buy A capable model is already on your desk. For a one-off question about one document, that is hard to argue with.
Two failures, and they are different
- Invention A value that is not on the page, returned with the same confidence as one that is. Usually a plausible number in the right format, which is what makes it expensive.
- Drift The same file, run twice, cut into different documents or read into different fields. Nothing failed, and the two answers disagree.
Invention is caught by grounding. Drift is caught by taking judgement away from the model wherever a rule can make the decision instead. They need different fixes, and a system that only addresses one is still unusable for work that gets audited. Sheaf addresses both at the architecture level.
Choose Sheaf if
- The output flows into a system that treats it as fact.
- Somebody will eventually have to prove where a value came from.
- You need the same file to produce the same answer twice.
Choose a general model if
- You are reading, exploring or drafting, and a person checks the result immediately.
A general model is optimised to be helpful. Document work needs a reader optimised to be checkable, which sometimes means being unhelpful on purpose. Sheaf is checkable first — and faster in practice, because the deciding logic never has to be run twice.