Model development and your data
Paresium improves Clara using real runs. Unless you turn it off, a generalized, stripped-down version of your organisation's data is eligible to be used for that work. Turning it off is one setting, takes effect right away, and changes nothing else: your records, your retention, and the service you get all stay exactly the same.
This page is for everyone, with schema links for developers.
What is shared, and when
Your runs are stored in full, at full detail, for as long as your retention policy keeps them. See The Record and retention.
Nothing is shared as your run happens. Sharing is a separate, later step: when Paresium prepares data for model development, an export step reads eligible runs and, at that moment, generalizes them before they go anywhere. Full-detail data never leaves storage for this purpose. Only the generalized copy, made at export time, is used.
What generalization actually removes
At export, the wording of the original request is replaced with a fixed marker that says only that a request happened, not what it said. The detailed working, the step-by-step findings behind an answer and the summary that fed the final write-up, is collapsed the same way: down to a flag that says a step ran, not what it found or concluded. What is left is shape, not substance: enough for Paresium to learn from how a run was handled, nothing that reads back as your business or your customer's words.
The export step is default-deny. If a run cannot be placed in an eligible category, or a check for identifying content fails, that run is excluded rather than shared. The bar is "prove it is safe to include," not "include unless someone objects."
What is never shared, on any setting
Direct quotes from your documents are kept out of the shared knowledge base Clara draws on for evidence, regardless of this setting. When Paresium indexes a published policy or playbook, it deliberately leaves out the exact quoted sentence and keeps only the structured decision made from it. This is a standing design choice, not something this setting turns on or off, and it means the data behind Clara's improvements does not become a searchable archive of your customers' words.
Enterprise is excluded entirely
If your organisation is on the Enterprise plan, none of your data is eligible for model development, generalized or not. This is not a setting to check: Enterprise runs are blocked from that path outright, and an attempt to route one through it is treated as a failure, not a quiet skip.
How to turn it off
Open Settings in the portal and turn off the control for including your organisation's data in model development. It takes effect immediately: from that point on, none of your organisation's runs are picked up by a future export.
Turning it off is forward-looking. It stops your data from being used in export runs from that moment forward. It cannot reach back into a model that was already trained before you turned it off and remove your data's contribution to it.
The same forward-only limit applies to the datasets that sit between an export and a trained model. An export run writes its generalized rows to a dataset file, and turning this setting off does not rebuild or remove your organisation's rows from a dataset that already exists on disk. Each row carries its own origin and plan tier, so your organisation's rows remain identifiable within a dataset even after you withdraw, the same way they would be identifiable in a model already built from one. What is written there is the generalized form described above, not your original wording.
If you cannot find the setting, or want this commitment in writing as part of a contract, contact us.
What turning it off does not change
Turning this off affects one thing only: whether your organisation's generalized data may be used for model development. It does not touch:
- Your records. Every governed run still produces its permanent Record, exactly as before. See The Record and retention.
- Your retention. How long your data is kept is unaffected.
- The service you get. The same checks run, the same way, at the same cost. Nothing about how Clara treats your organisation's runs changes.
For developers
Whether a run is even a candidate for this path depends on its plan tier,
carried on every trace. See the tier field on
RatedTrace for the enum this is built on.