How to read this page
This page defines the intended product boundary; it does not claim that a public application is available today. Implemented controls, tested release behavior, deployed-service controls, and independent assurance will remain separate evidence classes. Tensor Cortex does not currently claim SOC 2, ISO 27001, a completed penetration test, or third-party security certification.
The local project is authoritative
- Creating and running a local project should not require a Tensor Cortex account.
- Dataset sources, cleaned revisions, model files, checkpoints, eval results, and packages remain on the Mac unless the user explicitly exports or submits them.
- Automatic fixes create new revisions rather than silently rewriting source data.
- Project manifests should avoid embedding credentials and should keep machine-specific secrets outside portable artifacts.
- The first public beta sends no product telemetry or automatic crash reports. Any later diagnostic report must be generated and previewed on the Mac, sent only by a report-specific user action, and exclude workload content and persistent device identifiers.
Major data classes and intended defaults
| Data | Default location | Boundary |
|---|---|---|
| Source and prepared datasets | Local project | Project content remains local unless the user explicitly exports it. |
| Base models and adapters | Local model storage | License, source, revision, and checksum should travel with provenance. |
| Experiment configuration | Portable manifest | Configuration is versioned; secrets and provider credentials stay outside the manifest. |
| Checkpoints and results | Local artifacts | Retention and deletion remain under the project owner’s control. |
| Evaluation records | Local evidence | Scores, cases, hardware, and known regressions are attached to the experiment. |
| Website contact email | Email service provider | Handled separately under the public Privacy Policy. |
Network operations stay narrow
The first release permits only documented, user-initiated model metadata inspection, confirmed model downloads, manual signed model-catalog refresh, and signed-update requests. Inspecting a repository does not download its weights. Those boundaries identify:
- The destination, requested model revision or update artifact, and ordinary connection metadata exposed to the recipient.
- That datasets, prompts, outputs, adapters, checkpoints, and project paths are not attached to those requests.
- The user action or consent that initiates each request and how to disable scheduled update checks.
Models, datasets, and third parties
An open-model name does not settle license, retention, training-data, export-control, or commercial-use questions. Tensor Cortex Studio is intended to record source and license metadata, but users remain responsible for confirming that their dataset, base model, teacher model, and output use are permitted.
Report a security issue
Send the affected public surface, reproduction steps, and potential impact. Do not include live secrets, private datasets, or model credentials.
Please report privately and allow reasonable time to investigate before public disclosure. Include the affected version and minimal reproduction. Do not access other people's data, disrupt services, or test third-party systems. This reporting channel does not grant permission for intrusive testing or promise a bounty, response deadline, or legal safe harbor.