Prediction records include
- Timestamp and request ID
- User and model version
- Prediction and confidence
- Source and error status
Full input retention is optional and can be configured.
Data that powers a model should be handled with the same discipline as the model itself.
AscentML validates and screens every dataset before it is used, keeps an immutable version history, and only removes data permanently when you ask or when the contract requires it.
AscentML is intended for non-regulated business data.
Before a dataset is uploaded, the user attests that it does not contain:
That attestation is the first control. It is not the only one.
Files are uploaded directly to your organization’s isolated storage using short-lived, single-purpose upload links. Storage locations are assigned by the platform, never by the client.
Before a dataset moves forward, the upload is verified server-side for existence, size, and integrity.
Each dataset is profiled for common quality problems before it is used for model development:
Users can identify the target column and any known leakage columns during upload so they are handled correctly from the start.
Every dataset goes through automated screening for possible sensitive or prohibited information.
The dataset is rejected and does not proceed.
The dataset is quarantined. It is released only after both your organization and an AscentML administrator approve.
Dataset versions are immutable. Once a version is created, it does not change.
Models record the dataset versions and feature definitions they were built from. Derived features are stored as versioned definitions, so lineage covers both the raw data and the transformations applied to it.
When someone asks what a model was trained on, there is a precise answer.
Customer data is not automatically purged during an active contract.
Deletion in the platform is soft by default, so records can be recovered if something is removed by mistake.
Permanent deletion happens only at verified customer request or at contract termination. Purge operations are explicitly authorized, audited, and cover related records, stored files, model artifacts, cached explanations, and batch outputs.
Operational logs exist to run and support the service. They are not a second copy of your data.
Full input retention is optional and can be configured.
AscentML is not currently positioned for workloads that require HIPAA, FedRAMP, or similar regulated-environment assurances.
If your use case sits near that boundary, we would rather discuss it directly than leave it to interpretation.
Tell us what your organization expects from a vendor handling its data. We will walk through how AscentML meets those expectations and where the current boundaries are.