⚖ Working draft — under counsel review; not legal advice. Contact: [email protected]
Architecture. Web/app plane (VPS, EU) is fully separated from the scan plane (GPU workers).
Model weights never touch the web server.
Ingestion modes. Public HF reference (default, no upload) · scoped read-only HF token
(used once, permanently purged from the job record on completion).
safetensors-only. Pickle-based checkpoints (.bin/.pt/.pth/.ckpt) are refused — they are an
arbitrary-code-execution vector (CVE-documented). Repo code files are not even downloaded;
trust_remote_code is never enabled.
Private-job routing. Jobs with private tokens are dispatched ONLY to wipe-guaranteed,
single-purpose cloud workers — never to shared infrastructure.
Egress. Workers speak only to the model registry and our result endpoint (restricted;
full per-job network isolation is on the hardening roadmap and stated honestly here).
Deletion attestation. Weights are wiped before report delivery; every job closes with an
Ed25519-signed receipt (shard SHA-256 manifest + timestamps) — verify against our
published key.
Self-serve report deletion. You can permanently erase any report from your dashboard
(two-step confirmation). Erasure removes the report, metrics, images and the job record immediately
and irreversibly — we keep no copy and no content log, by design: your models and results stay yours.
Contractual commitments. No training on customer models; no telemetry; no report-external
data egress.
Access control. Founder-only production access at this stage; least-privilege tokens;
signed cookies; rate-limits; strict security headers.
Compliance roadmap. GDPR DPA + KVKK; independent pentest before enterprise onboarding;
SOC 2 Type II on the roadmap (stated honestly — not yet certified).
Tetracta · private beta · your model is deleted from the
worker after processing (signed attestation per job) · deleting a report from your dashboard erases it
permanently — we keep no copy · docs · security ·
terms · privacy