AI & Code Processing Disclosure
AI & Code Processing Disclosure — ORA by Aurem
Last updated: February 12, 2026 Effective: February 12, 2026
ORA by Aurem reads your GitHub repository, generates fixes with large language models (LLMs), and commits code directly. This document explains — in engineering-grade detail — exactly what happens to your code when it enters our system.
This disclosure supplements our Privacy Policy, Subprocessor List, and Terms of Service.
1. What We Access
When you connect a GitHub repository:
| Data Type | Purpose | Retention |
|---|---|---|
| Repository metadata (name, default branch, languages, size) | Task routing & UI | Until repo disconnected |
| Selected file contents (specific files needed for the task) | Fed to LLM for context | Purged after task completion (see §4) |
| Diff/patch produced by the LLM | Preview & commit | Stored 90 days for undo/audit |
| Task inputs (your prompt/instruction) | Model input | Stored 30 days for support & abuse review |
| Task outputs (model response, commit SHA) | Task history & billing | Stored while your account is active |
| GitHub Personal Access Token (PAT) | Reading/writing to your repo | Encrypted at rest (HKDF-Fernet), scope-scoped |
We do not clone your entire repository to our servers. We fetch only the specific files an active task requires.
2. Where Your Code Goes (LLM Routing)
ORA uses a multi-model "Parliament" router. A single task may be evaluated by 2–3 models in parallel for consensus. The models currently in production:
| Model | Provider | Purpose | API Endpoint |
|---|---|---|---|
| Claude Sonnet 4.5 / 4.6 | Anthropic PBC (via OpenRouter or direct) | Council A — primary code generation | api.anthropic.com / openrouter.ai |
| DeepSeek Chat | DeepSeek | Council B — validation & lightweight fixes | api.deepseek.com |
| GLM-5.2 | Zhipu AI (via OpenRouter) | Fallback when LongCat unavailable | openrouter.ai |
| Groq-hosted models | Groq Cloud | Low-latency utility tasks | api.groq.com |
Which model sees which code is deterministic per task type. See our Subprocessor List for full inventory.
2.1 Data Flow
Your Browser → Aurem backend (Canada/US region) → LLM provider API
↳ Response → Aurem → GitHub commit
Every hop uses TLS 1.2+ encryption in transit.
3. Training & Model-Improvement Opt-Out
Your code is never used to train LLMs. We contractually enforce this in the following ways:
| Provider | Training Opt-Out Status | Evidence |
|---|---|---|
| Anthropic | ✅ Opted out via API Zero-Data-Retention default | Anthropic API Data Usage |
| OpenRouter | ✅ We set X-OR-Train: false and use privacy-preserving routes |
OpenRouter Privacy |
| DeepSeek | ✅ Enterprise API — no training on inputs | DeepSeek Terms |
| Groq | ✅ Zero-retention API tier | Groq Privacy |
| Zhipu (GLM) | ✅ Via OpenRouter — no training on API traffic | Contractual (OpenRouter) |
If any upstream provider changes their training policy, we will notify you 30 days before continuing to route traffic to them, and we will provide an opt-out.
4. Data Retention at LLM Providers
Even without training, some providers cache request/response pairs for abuse monitoring. We use the shortest retention available:
| Provider | Retention | Our Configuration |
|---|---|---|
| Anthropic | 30 days (abuse monitoring only) | Zero-Data-Retention enabled where eligible |
| OpenRouter | 30 days (opt-out available for paid tier) | Opt-out flag set |
| DeepSeek | 30 days | Standard API |
| Groq | 24 hours | Standard API |
Aurem's own retention: Selected file contents are purged from our servers within 24 hours of task completion. Only the diff, commit SHA, and task metadata are retained per §1.
5. What We Do NOT Do
- ❌ We do not sell your code or code-derived data.
- ❌ We do not use your code to train our own models (we don't train any models).
- ❌ We do not share your code with LLM providers for their own training.
- ❌ We do not access files outside the scope of an active task.
- ❌ We do not read repositories you have not explicitly connected.
- ❌ We do not persist your PAT in plaintext or in logs.
6. Sensitive-Data Detection
Before sending file contents to an LLM, we run a pre-flight secret scan using detect-secrets patterns to redact:
- API keys (AWS, Stripe, OpenAI, GitHub, etc.)
- Private keys (RSA, EC, PGP blocks)
- Database connection strings with credentials
- JWT tokens with recognisable structure
Redaction replaces the match with [REDACTED_SECRET_<type>] before the LLM sees it. This is best-effort — you remain responsible for what is committed to your repo.
7. Human Review
- Your code is not read by Aurem staff except when you explicitly file a support ticket that requires it. Such reviews are logged, and you are notified.
- Anonymised/aggregated task metadata (model latency, error rates, no code content) is reviewed by engineering for reliability and cost optimisation.
8. Model Output Warranty
LLM output is probabilistic. Aurem does not warrant that generated code is:
- Bug-free
- Optimal
- Free of licensing conflicts with third-party dependencies your project pulls in
You are the final reviewer. Every fix goes through a preview + commit step. We recommend keeping automated tests + code review in your workflow. See Terms §7 — Limitation of Liability.
9. Rights & Requests
Under GDPR, CCPA/CPRA, DPDP Act, and PIPEDA, you may:
- Request a copy of the AI-processed history on your account
- Request deletion of your task history
- Request the specific LLM provider that handled a given task ID
- Withdraw consent and terminate LLM processing (this ends your ability to use ORA fixes but preserves your account)
Email auremcto.com/support with your account email and task ID(s).
10. Changes
Material changes to LLM routing, subprocessor list, or retention windows trigger email notice at least 30 days before taking effect.
11. Contact
Polaris Built Inc Incorporated in Canada AI/data inquiries: auremcto.com/support Engineering security: security@auremcto.com