A checked text format for process models
tok: a textual concrete syntax of BPMN and DMN for humans and AI agents · arXiv:2611.07945
In plain words
Process models show the steps and choices in a business process. They are often shared as XML files, which include a lot of technical structure that can be hard to read or write.
tok is a plain-text way to describe those models. Its engine checks the text and converts it to standard files.
In tests, four language models wrote valid tok models for 85.7% to 100% of descriptions, depending on the model and dataset; the same models wrote valid XML for every description with a one-page reference.
The research
A process model has to be readable and maintainable while still working with tools that exchange standard files. A shared text notation could reduce the gap between those forms and give language models a checker while they write.
The results make that possibility concrete for the tested faults and description-writing tasks, while showing that checker coverage and model validity vary by case.
tok is a plain-text notation for BPMN processes and DMN decisions, with an engine that checks models and translates them to standard files.
The checker comparison used the same injected faults in tok and corresponding BPMN XML, counting a new warning as well as an error as a signal.
Tok check signalled more of those variants than the XML Schema and bpmnlint correctness rules together, though two specific gateway faults were often missed.
Four language models then wrote models from descriptions in two datasets. An independent judge found every XML output valid and tok validity between 85.7% and 100%, depending on model and dataset. This supports tok as a workable checked format in these tasks, with results varying by model.
The tests cover specified injected faults, four language models, two English description sets and one simulation setup.
A structurally valid model can still misread its description, and the study did not evaluate editing existing models, sustained human-model collaboration or executing compiled processes.
What the evidence shows
- One text, standard models
tok is a plain-text notation for Business Process Model and Notation (BPMN), a standard for describing processes, and Decision Model and Notation (DMN), a standard for decisions within them. Its engine checks the text against represented standard rules, imports standard files and compiles tok back to them.
The text holds the model while the engine generates diagram layout, so the same model can be handled as text and exchanged in standard formats.
- More injected faults signalled
On the same 1,411 injected faulty variants, tok check signalled more cases than the XML Schema and bpmnlint correctness rules together. Because a new warning counted as a signal, this comparison measures warning coverage as well as errors; it does not mean every fault was blocked from export.
- Specific faults remain difficult
The checker’s overall rate hides fault types it often missed. It signalled 15 of 27 variants of an exclusive split with no conditions or default, and 1 of 5 variants of an exclusive split joined by a parallel join.
These cases involve behaviour across connected parts of a process; they help bound the result to the tested fault types rather than establish general coverage of modelling errors.
- Models learn the notation
In a separate test, four language models turned descriptions from two English datasets into process models. An independent judge checked structural validity, including schema rules and whether flows connected known nodes.
With a one-page reference and up to four attempts, tok validity ranged from 85.7% to 100% across model-and-dataset groups, while BPMN XML was valid for every description. The range matters: the result differed across models and datasets.
- Guidance changes token cost
The token cost changed with the agent skill package: the lean and distributed setups produced substantially different run costs for Claude Sonnet 5.5. The independent judge found all lean-skill models structurally valid on both datasets, compared with 83.3% on PET and 93.3% on PMo under the distributed skill.
Since the packages changed the reference-reading procedure and tools together, this comparison cannot attribute the difference to either change alone.
Deep
- Original paper authors
- Marek Deja
- Submitted to arXiv
- License
- CC BY-SA 4.0