Security
What is implemented, and what is not yet.
Unpublished manuscripts are among the most sensitive documents a research organisation holds. This page describes the controls that exist in Trace today and marks everything else as roadmap. It does not claim any certification.
No certification is claimed. Trace holds no security or privacy certification.
There is no SOC 2 report, no ISO 27001 certification, and no third-party attestation of any kind. Trace is not a HIPAA covered entity or business associate and makes no HIPAA compliance claim. Do not upload protected health information or other regulated personal data.
The controls described below are engineering controls that exist in the product. They have not been audited by an independent party. If a procurement process requires attested compliance, Trace does not currently meet that bar, and it is better for both parties to know that at the start.
Implemented
Controls that exist in the product today.
- 01Encryption in transit
- All traffic between clients, the application, and upstream providers is served over TLS. Plaintext transport is not offered as an option.
- 02Encryption at rest
- Stored documents, derived text, and database contents are encrypted at rest using the managed encryption provided by the underlying storage and database platforms.
- 03Tenant isolation
- Every record carries its workspace. Data access is scoped by workspace at the query layer, and cross-workspace reads are not an available code path in the application.
- 04Role-based access control
- Roles separate analysts, reviewers, and administrators. Provider configuration, retention settings, and deletion are administrator actions.
- 05Audit logging
- Measurement runs, exports, approvals, provider configuration changes, and deletions are recorded with actor, timestamp, and the affected object.
- 06Scoped credentials
- Upstream provider credentials are held in server-side configuration, are never exposed to the browser, and are scoped to the single provider they authenticate.
Roadmap
Controls that do not exist yet.
Listed so that a security review can see the gaps directly rather than infer capability from silence.
- SSO and SAML
- Single sign-on and SAML federation are planned and not available today. Authentication is managed inside the application in the current build.
- SCIM provisioning
- Directory-driven user provisioning and de-provisioning are planned.
- Regional data residency
- Region-pinned hosting is planned. Do not assume a specific region is available for a pilot without asking.
- Customer-managed keys
- Customer-managed encryption keys are planned and not available today.
- Independent penetration testing
- A third-party penetration test is planned. No such test has been completed, and no report exists to share.
- Third-party attestation
- Formal audit and attestation work is a future undertaking. See the statement below on certifications.
Data lifecycle
Retention, deletion, and portability.
- 01Retention
- Retention windows are configurable per workspace. Where no window is set, content persists until it is deleted by the workspace. Retention settings are administrator-controlled and their changes are logged.
- 02Manuscript deletion
- Deleting a manuscript removes the stored document and the text derived from it, including extracted sections and embeddings, and records the deletion in the audit log. Aggregate measurements that no longer reference the document may be retained; those that do are removed with it.
- 03Export and portability
- Workspace measurements, diagnostics, and reports are exportable as CSV and JSON at any time, so leaving does not mean losing the record of what was measured.
- 04Backups
- Backups inherit the same encryption as primary storage. Deletion propagates to backups on the backup rotation schedule rather than instantly, and that lag is stated here rather than glossed over.
Third parties
Provider data policies and subprocessors.
Upstream data sources and infrastructure vendors are treated as a governed registry rather than an implementation detail.
- 01Provider policy registry
- Every upstream data source carries an explicit policy entry covering commercial use, full-text storage, derivative embeddings, redistribution, and attribution. A provider with no approved entry cannot run in production.
- 02Rights-aware storage
- Full text is stored only where reuse rights are confirmed. Where they are not, the analysis runs on metadata and the report says so on its face.
- 03Honest measurement labelling
- An API model with web search is not a consumer product surface, and measurements are labelled for exactly what they ran against. Customer-supplied exports from a licensed enterprise surface are ingested and normalized, not collected.
- 04Simulation has no network egress
- Social-diffusion and stakeholder simulation run in a sandbox with no path to any live platform. There is no capability to post, message, or act on a real network.
Subprocessors
Categories are listed here; the specific vendor list is provided under a pilot agreement.
- Cloud hosting and storage — runs the application and stores encrypted documents and database contents.
- Managed database — stores workspace records, measurements, and audit logs.
- Language-model providers — process text for analysis, under zero-retention and no-training terms, or replaced by a customer-controlled endpoint.
- Bibliographic and availability sources — public scholarly infrastructure queried for metadata and open-access status.
- Operational tooling — error monitoring and transactional email.
No vendor is named on this page because the list is not yet stable enough to publish and a public list that lags reality is worse than none. It is supplied in full, with names, before any pilot processes customer content.
Model providers
Your manuscripts are not training data.
- 01No training on customer content by default
- Customer manuscript content is not used to train models. This is the default and it is not conditional on plan or price. Any change requires an explicit, revocable, workspace-level opt-in that is recorded.
- 02Provider retention settings
- Where a hosted model provider is used, zero-retention and no-training options are configured before unpublished manuscript content is sent. Where a provider cannot offer those terms, it is not used for manuscript content.
- 03Customer-controlled endpoints
- Model calls can be pointed at a customer-operated, OpenAI-compatible endpoint so that manuscript text stays inside an environment you already govern.
- 04Deterministic local mode
- The product runs end to end against a deterministic local model with no external API keys, which is how demonstrations and tests are run.
The default, stated plainly
Disclosure
Reporting a vulnerability.
Coordinated disclosure
Reports are read by an engineer, not a ticket queue.
Send security reports to security@trace.science. Include the affected URL or endpoint, the steps to reproduce, and what you were able to access. If you need to send anything sensitive, say so and an encrypted channel will be arranged.
- Acknowledgement within three business days.
- An assessment and a remediation plan within ten business days, or an explanation of why it will take longer.
- Credit in the fix notes if you would like it, and no legal action against good-faith research conducted without accessing other customers’ data, degrading the service, or exfiltrating content.
- There is no paid bug bounty programme at present. That is a resourcing fact, not a judgement about the value of the work.