04

Teyit — document intake with an explicit confidence boundary

In development2026 —Document AI · Turkish · Arabic · English · Human-in-the-loop · Self-hosted

Designed and specified, not yet built. The problem and the decisions below are real; there are no screenshots or results, because there are none to show.

The problem

The best-documented reason automation projects fail is not integration. It is that nothing defines what the system does when it is not sure. A deployment without that boundary either proceeds silently, which is dangerous, or blocks on a person nobody appointed to review it — a bottleneck wearing automation's clothes.

There is a second failure underneath it. The longer an automatic system runs without visibly breaking, the less anyone checks it, so confidence in the output drifts upward while its accuracy does not. Any design that treats human review as a formality inherits that drift.

Invoices are where this gets expensive and measurable. Manual handling runs roughly five times the cost of an automated pass, Türkiye's e-invoice regime makes structured invoice data effectively compulsory, and document tooling is overwhelmingly English-first — so Arabic script and Turkish diacritics are exactly where the off-the-shelf accuracy degrades and the boundary matters most.

How it works

Three exits from adjudication, and deliberately no fourthA document is read, then extracted into typed fields, each carrying a confidence and a reference to the region it was read from. Deterministic checks then run over that output: line items summing to the subtotal, subtotal plus tax matching the total, a legal tax rate, a valid supplier identifier, a plausible date, and no duplicate. Adjudication has exactly three exits. A document is committed automatically only when confidence clears the threshold and every applicable check passes. If a check fails or confidence falls short it is held, with the specific field and the specific failed check named. If it is not a document of this kind at all it is rejected. There is no fourth path, so nothing leaves in an undefined state.01ReadScan or photo,any of threelanguages02ExtractTyped fields, eachwith a confidenceand a source box03CheckSums, tax rate,checksum, date,duplicateCommittedConfident and every check passedHeldThis field, this failed checkRejectedNot a document of this kind
Three exits, and no fourth. Every document leaves as committed, held or rejected, and carries the reason either way.

Decisions

  1. What makes a non-deterministic extractor trustworthy?

    ChoseDeterministic checks over the model's output

    Instead ofA confidence score from the model alone

    A model's own confidence is a self-report, and a fluent extractor is confidently wrong in exactly the cases that matter. Arithmetic is not: line items either sum to the subtotal or they do not, a tax rate either legally exists or it does not, an identifier either passes its checksum or it fails. The model proposes and the arithmetic disposes, which converts an opaque probability into a specific, nameable failure.

  2. What does the system do when it is not sure?

    ChoseHold it, and name the field and the failed check

    Instead ofCommitting anyway, or flagging the whole document as low confidence

    These are the two documented ways this fails. Proceeding silently is how a wrong number reaches an accounting system. Handing back a whole document marked uncertain makes a human re-read all of it, which is the work the automation claimed to remove. Naming one field and the check it failed turns review into a decision instead of a re-keying job.

  3. How do you stop people over-trusting the automatic path?

    ChoseSample the auto-committed output for spot-checking

    Instead ofReviewing only what the system flags

    Automation complacency is the failure where a system that has not visibly broken stops being checked at all. If humans only ever see the documents the machine already doubted, nobody is measuring the errors inside the confident set — which is the only error rate that matters, and the one nobody publishes. A deliberate sample keeps trust calibrated by evidence.

  4. Cloud document AI, or self-hosted?

    ChoseSelf-hosted OCR and layout analysis

    Instead ofA per-page cloud document API

    Three reasons pointing the same way: financial documents never leave the client's premises, which is often the objection that decides the sale; there is no per-page cost, so volume does not change the economics; and the multilingual open-source models handle Arabic script and the full Turkish diacritic set, which is the specific place English-first tooling degrades.

Next project

05Vardiya — a scheduler that encodes the rules instead of approximating them