Zero-Data-Retention (ZDR) Privacy Policy
1. Strict Zero Data Retention (ZDR)
oneshot (oneshot.computer / api.oneshot.computer) operates exclusively as a stateless, Zero-Data-Retention (ZDR) inference provider:
- Zero Prompt or Completion Persistence: Prompts, messages, system instructions, tool calls, and generated completions are processed strictly in volatile GPU/host RAM for the duration of the HTTP stream and are immediately discarded upon stream completion.
- Zero Model Training or Fine-Tuning: Customer inputs and outputs are never used to train, fine-tune, distill, or evaluate any model.
- Zero Payload Logging: Edge gateways and GPU worker nodes have request-body and response-body logging disabled at the application and reverse-proxy layers.
2. Ephemeral In-Memory Prefix Caching
To minimize Time-To-First-Token (TTFT) on multi-turn coding and agentic workloads, oneshot utilizes volatile in-memory KV prefix caching (vLLM / SGLang):
- KV cache tensors reside exclusively in volatile GPU VRAM / host DRAM and are automatically evicted via Least-Recently-Used (LRU) memory pressure or node termination.
- Edge routing hashes only a non-reversible cryptographic digest (
FNV-1a / SHA-256) of the prefix header for replica affinity; raw text is never stored at the router layer.
3. Operational Telemetry
For billing reconciliation, rate-limiting, and hardware autoscaling, oneshot records only non-content operational counters: timestamp, model identifier, token counts (prompt_tokens, completion_tokens, cached_tokens), latency (TTFT / tok/s), and HTTP status code.
4. Contact
For security, compliance, or router verification inquiries: hello@oneshot.computer.