Privacy Policy &
Data Governance.

3.1 Directly Provided Data

• Identity: Name, email, SSH public keys (for git integration), and API tokens.
• Workspace Context: Project file trees, dependency graphs, and environment configurations.

3.2 Technical Telemetry & AI Context

• Local-First Architecture: Dropstone operates primarily on the client device. Indexing, AST parsing, and vector embedding generation occur locally.
• Ephemeral Context Windows: When you trigger an AI agent, only the specific code snippets relevant to the prompt (the "Context Window") are transmitted to inference endpoints. We do not upload entire repositories.
• Operational Metrics: Latency, token usage rates, and GPU/CPU thread utilization for performance debugging.

5.1 Zero-Retention Inference Architecture

• Inference (Processing): Code sent to our model partners (OpenAI, Anthropic) follows a strict stateless protocol. Data is held in volatile memory only for the duration of the generation request and is cryptographically wiped immediately after.
• No Third-Party Training: We have executed "Do Not Train" agreements with our model providers. Your code is never used to train their foundation models. This clause is about them; for how Dropstone itself may use session content, see 5.2.

5.2 Improving Dropstone's Own Models

Consumer accounts (Free, Pro, and Max) participate in two consent-based programs. Both are on by default and both can be turned off at any time in dashboard settings. Enterprise accounts are excluded from both, regardless of setting, and that exclusion is enforced when data is written rather than by configuration.

• Improve Dropstone (outcome metadata). Anonymous, metadata-only signals used to train our routing and verification models: whether an edit was accepted, whether tests passed, the turn count, a coarse failure category, and the model and effort used. This never includes your code, prompts, diffs, completions, or tool arguments — the table storing it has no column capable of holding content.
• Full coding sessions. The session text — your task description, the model's diffs, and tool or test output — used to post-train our own open-weight model on real, test-verified work. This is text only; raw images are never collected. This program captures the content of your work, so if your sessions involve confidential or proprietary code, turn it off before you begin.

Turning either program off stops future collection; it does not by itself delete samples already stored. You may request deletion of previously collected samples at any time by contacting [email protected], and we will action it without requiring a reason.

5.3 Memory & Learned Context

To give the assistant continuity between sessions, we store context derived from your work against your account. This is retained rather than ephemeral, and it is separate from the training programs in 5.2. It consists of:

• Knowledge. Documents and reference material you or the agent have surfaced, kept with their title, content, and source reference.
• Lessons. Short generalisations about what worked — a situation, the strategy applied, and why — so the agent repeats what succeeded and avoids what did not.
• Summaries. Condensed records of prior sessions.

Each item is stored with a vector embedding — a numerical representation of its text — so it can be retrieved by relevance later. Embeddings are derived from your content and are treated as your content.

This memory is scoped to your account. It is retrieved to inform your own sessions and is not served into another user's assistant. You can review and delete individual memory items from your dashboard, and request deletion of all of it at [email protected]. Deleting your account removes it.

5.4 User-Owned Fine-Tuning

If you choose to fine-tune a model on your proprietary codebase (an Enterprise feature), the resulting model weights are owned exclusively by you and are siloed from other users.

12.1 Infrastructure Security

• Encryption Standards: Data at rest is encrypted using AES-256 (GCM mode). Data in transit is secured via TLS 1.3 with forced HSTS.
• Key Management: Encryption keys are rotated periodically via a hardware security module (HSM) equivalent.

12.2 Enterprise Data Isolation

For Enterprise Tier customers, we offer logical tenant isolation. Your vector indexes and usage logs are tagged with a unique Tenant ID, ensuring that your data is logically separated from other customers at the database level.

12.3 Vulnerability Management

We conduct regular static code analysis (SAST) and dependency scanning to identify vulnerabilities. Security patches are deployed within 72 hours of critical disclosure.