What we store
Uploading a whole-slide image creates a sample: the persistent Strand entity that carries organization ownership, status, jobs, and results. We store two kinds of data for that sample:- Uploaded whole-slide image: the source image you upload for prediction.
- Marker predictions: the per-marker outputs generated from that image, including the image pyramids served back to you.
De-identified input requirement
Slides must be de-identified before upload. Do not upload protected health information (PHI) or other regulated identifiable health information unless a separate written agreement with Strand AI expressly authorizes that processing. This requirement covers the slide pixels, embedded label and macro images, scanner metadata, and filename. Whole-slide images often carry identifiers in an embedded label image, a macro image, scanner-written metadata fields, or the filename. Remove these identifiers before upload. If identifiers are baked into the tissue region, crop or re-scan the slide before uploading.Optional de-identification at ingest
Automated slide de-identification is off by default and runs only when Strand AI enables it for your organization. When enabled, the ingest pipeline removes embedded label and macro images and known PHI-bearing metadata fields, then validates the result before permanent storage. Unsupported formats or failed validation stop that enabled path. The enabled path does not retain an unprocessed copy in permanent storage. Automated de-identification does not scan or redact identifiers in the tissue image or filename. It does not replace de-identification before upload, and it is not a representation that data satisfies HIPAA or any other legal de-identification standard.Encryption
Your data is encrypted in transit and at rest. Strand AI runs on enterprise-grade cloud infrastructure with encryption applied to stored objects and database records, and TLS protecting data moving between you and our services.Retention & deletion
Retention is under your control. Expiration is opt-in: samples never expire until an owner or admin sets a policy (see Sample expiration for how to configure it). When a sample expires, it moves to “Trash”, where it is held for 7 days. During that window you can restore it. After the window closes, the sample and its marker predictions are permanently deleted 7 days after expiry. Both the stored objects and the corresponding database records are removed.Deletion happens over a window rather than instantly. A sample is permanently
deleted 7 days after it expires, once it has passed through the Trash grace
period, and remains restorable until then.
Logs are retained separately
Your sample data (uploads and the predictions derived from them) is deleted according to your retention policy as described above. Operational and audit logs are kept separately and are retained beyond your sample data for security, billing, and compliance purposes. These logs do not contain your slide images or prediction outputs; billing history is preserved even after the underlying samples are deleted.Access control & audit logging
Access to samples is scoped to your organization and governed by role-based permissions (see the permissions table in Sample expiration). Retention actions (automatic archival on expiry and permanent deletion after the grace window) are audit-logged, recording what was deleted and when.We do not train on your data
We do not use your data to train our models. If that ever changes, we will notify you in advance, and any such use will be opt-in.MCP connections and explicit conversation context
The hosted MCP endpoint authorizes each action against the connected user and organization, using OAuth or an organization-bound API key. Slide transfer goes directly to GCS rather than through MCP. Core upload/result tool responses contain expiring bearer URLs in model-visible context; keep them out of prose, logs, and screenshots. Do not use chat to relay multi-GB slides or PHI. The in-chat Lattice viewer is read-only, and result pixels never enter the model’s context.strand_ai.view_sample requires an organization-scoped sample
ID returned by list_samples; the same public slide can have a different ID in
another organization. Opening a sample gives the model a bounded sample card
(ID, ownership, name, readiness, MPP, segmentation status, the latest job’s
status, progress, and markers, and a platform link) and a result summary (count,
and the newest result’s status, markers, and completion time); the result IDs
the viewer switches between stay in the viewer. Core tools such as get_sample
and get_job return job IDs to the model as usual. The viewer reads result files
with a short-lived, read-only permission for one result, checked on every read
against the live connection. A compliant host keeps the app-only
strand_ai.viewer_session and its result away from the model; the server cannot
enforce a malicious host’s model delivery choices. Conversation-host policies
govern sample metadata shared in chat separately from Strand AI sample
retention.
Candidate paid approval state contains only bounded normalized arguments, a fresh
UUID, and exact patch/marker/credit estimate, encrypted and bound to audience,
principal/organization, method, tool, and arguments for 600 seconds. It contains
no bearer credential or sample display name. The approved REST replay mapping
lasts while the associated job remains stored; permanent deletion of the
job/sample/organization/account removes it. This is not a claim that current
production already provides native approval or exactly-once submission.