On October 1, 2026, Tonic.ai shipped Fabricate 4.32 with a feature called Simulations, in private preview for Enterprise customers. It generates a set of linked records that tell a story over time, such as the documents, emails, and Slack posts around a product design review, or the appointments, patient notes, and insurance responses around a patient’s treatment. Each result comes with a scenario overview, the cast of personas involved, and a timeline written as a script of acts and scenes.
Documents and messages can connect like real work, and a critical situation can still be missing from the test. If your scenarios never include a launch date changed in a later email, your test may miss an assistant that reports the old date.
Key takeaways
- Synthetic data use cases now include work scenarios: documents, messages, and records generated as one workflow, with time order and personas.
- A scenario can look real and still leave out a situation where the AI gives the wrong answer in actual use, such as a decision changed later in another tool.
- Test scenario data against the task: list the situations that matter, check that the data contains them, and run the task under fixed model and evaluation conditions, on real data where you can.
- If you could not test on real data, say so, and note what still needs checking in real work.
What is new in synthetic data use cases: whole work scenarios
Simulations generates several documents and messages as one workflow. What you check in that output goes beyond individual values to the order of events and the points where a decision changes. That suits AI work that reads across a thread, such as an assistant that summarizes a project from documents and messages, or an agent that acts on what an email chain decided.
The table below compares what generation focuses on. The two columns are not separate kinds of data. The use cases listed are our interpretation, not an exhaustive classification.
| Aspect | Focus: values and distributions | Focus: events across records (scenarios) |
|---|---|---|
| What it generates | Records whose values, types, and distributions follow a source | Linked documents, messages, and records that follow a timeline and a cast of personas |
| Common use | Development and test databases, analytics, model training and testing | Testing assistants, retrieval, and agents that read across documents and conversations |
| What has to hold | Value ranges, relationships between fields, the rare cases that drive results | Event order, the conditions that apply, and which later message changes an earlier one |
| What looking real can still miss | Rare cases and relationships that drive the result | The decision change or exception the answer has to reflect |
Similar distributions do not prove task fit for tabular data either. A rare case or a relationship that drives the result can drop out while the overall shape stays the same. With scenarios, you also check the order of events, the conditions that apply, and the decisions that change along the way.
One email that moved the launch date
Take a team testing an assistant that writes the status summary after a product design review (illustration). The scenario set looks right: a spec document, an email chain, a Slack channel, five personas, three acts. The assistant passes. In production it reports the wrong launch date.
- 01Review
The design review sets a launch date in the spec document.
- 02Change
Two weeks later, a reply in a forwarded email moves the date.
- 03Gap
Nobody updates the spec, and none of the test scenarios included a change like this.
- 04Summary
The assistant reports the date from the spec, which is now wrong.
The scenarios were realistic, but none contained a decision that changed later in a different tool. A generator writes the stories it is asked to write, so you have to ask for these turns and check that they are there. This is the same point we make about AI data preparation starting with the task: data can look complete and still fall short of AI-ready data for a specific task.
How to test whether scenario data fits your AI task
Start from the task. These checks work for scenario data from any source, including scenarios your own team writes.
- List the situations the task must get right. Pull them from incidents and reviewer notes: a reversed decision, an approval given in a direct message, two people with the same first name, a figure changed in an attachment.
- Check coverage before running anything. Check that each situation you listed appears in a scenario. However plausible the whole set looks, a situation that is missing is a situation you did not test.
- Run the task on real data where you can. Use the same model, prompt, and evaluation criteria on the scenario set and on a small slice of real data, and compare the results situation by situation.
- State what real data did not confirm. If you tested only on reconstructed data, record that. Note which situations you checked and what still needs checking in real work.
- Keep the data version with the result. When the scenarios, the data, or the model change, rerun the same comparison so you can see what moved.
Pick one AI task that reads across documents or messages, write down five situations it has to get right, and check whether your test data contains each one.
Related product: Syntitan
For structured data, CUBIG’s Syntitan supports use-case-specific preparation and validation. Teams prepare data for a defined use case, validate it with a selected model or agent under fixed evaluation conditions, and release it as a data version. Each version carries the data state before and after preparation, the evaluation conditions, the results, and a validation result of qualified, not qualified, or inconclusive for the task. Explore Syntitan.
