Data Classification#
Use ETH's PUBLIC, INTERNAL, CONFIDENTIAL, and STRICTLY CONFIDENTIAL levels. The information owner classifies the project and approves systems. Unpublished research and data awaiting publication may be confidential; do not classify it yourself merely because you created the file.
AI Tools#
- Check the current ETH assessment for the exact tool and account mode.
- Strictly confidential information must not enter external cloud/AI services.
- Confidential, personal, partner, licensed, or unpublished material requires explicit approved handling.
- Minimize prompts and repository context.
- Never provide credentials, private keys, or recovery information.
- Review output for correctness, fabricated claims/citations, licenses, bias, and inappropriate disclosure.
- Agree AI use and disclosure expectations with the supervisor.
If classification or service approval is unclear, ask the named project information owner or supervisor privately before uploading. For a suspected exposure, use the private incident route.
Approval is contextual. A tool approved for one classification or account mode is not automatically approved for every project or plugin/integration.
Personal AI Accounts And Costs#
- The manual workflow and GitHub Copilot Student route require no personal AI purchase for eligible verified students.
- Zed with personal OpenRouter or any other paid personal service is optional, not a passport requirement.
- Unless explicit written approval is given before a purchase, IDEAL Lab does not provide credits, reimburse charges, administer a personal account, or recover unused credits.
- The personal account holder is responsible for usage, charges, key security, monitoring, auto-recharge settings, taxes/fees, and current provider terms.
- A spending limit reduces financial risk; it does not approve project data, guarantee technical correctness, or replace usage monitoring.
- Lab-funded access for eligible PhD students, postdocs, or staff follows the separate internal lab handbook. Do not assume personal costs will be moved or reimbursed later.
The complete student-facing warning and setup are in Optional Zed and personal OpenRouter.
AI-Agent Execution Boundaries#
- Start read-only when explanation or review is sufficient.
- Grant access only to the repository, files, and tools required for the task.
- Prefer one-time approval for a reviewed command over broad persistent access.
- Never grant an agent access to credentials, private keys, recovery material, or unrelated home/project directories.
- An agent may write files, install software, use the network, commit/push Git changes, or start remote/Slurm work only when a human explicitly requests that action and reviews its scope.
- A sandbox limits impact; it does not prove that a command or result is safe.
- Inspect the diff and verification output before accepting or publishing work.
Repository And Storage Boundaries#
- Commit source, tests, small fixtures, configuration examples, and docs.
- Do not commit real
.envfiles, credentials, datasets, checkpoints, generated results, or local editor/Obsidian state. - Put project data in the supervisor-approved NAS/Euler location.
- Record which copy is the main approved copy, which copies are temporary, and which final results must be kept.
Incident Rule#
If protected material or a credential enters Git or an unapproved AI service, stop sharing, contain the credential/access risk, preserve useful evidence after removing credentials, personal information, private paths, and research data, and notify the information owner and lab IT promptly.