Goal: choose authoritative, working, temporary, and final locations for a research project before large files are created.
Prerequisites: the project's supervisor/information owner and intended compute environment are known.
Execution context: planning exercise. Do not move or delete real data while completing it.
Scenario Inventory#
For each artifact, choose a location and justify it:
| Artifact | Questions to answer |
|---|---|
| Python source and tests | Who reviews it? Does it belong in GitHub? |
| Raw dataset | Is it published, internal, confidential, or restricted? |
| Processed training data | Can it be recreated? Does computation need fast I/O? |
| Model checkpoints | Which are temporary and which support a result? |
| Logs and metrics | How long are they needed and who reads them? |
| Final figures/tables | What must be preserved for a thesis/publication? |
| Software tokens | Which credential store handles them? |
| Temporary job files | What happens when the job or scratch lifetime ends? |
Complete The Location Map#
Information owner:
Classification:
Source code:
Authoritative raw data:
Euler working data:
Temporary data:
Checkpoints:
Logs and metrics:
Final results:
Long-term handover/archive:
Backup and restore owner:
Retention/cleanup date:
Use system choices, the NAS guide, and Euler storage.
Add Repository Boundaries#
Review .gitignore. It should normally exclude local secrets, data,
checkpoints, caches, environments, and generated output while keeping source,
tests, configuration examples, and documentation visible.
Do not add a broad pattern such as *.csv if small source-controlled test
fixtures legitimately use that extension. Prefer project directories such as
data/local/, checkpoints/, or results/generated/.
Verification#
Ask the supervisor/information owner to confirm:
- the classification and approved services;
- the authoritative raw-data copy;
- which working locations are backed up;
- which temporary locations may be purged;
- the required final handover artifacts.
Common Failures And Safe Recovery#
- Large file already committed: stop pushing, notify the repository owner, and move it to approved storage before coordinated history cleanup.
- Only copy is in scratch or Blade
D:: copy and verify it in approved durable storage immediately. - Collaborators need write access: use the approved group/ACL procedure; do not make the directory world-writable.
Understand Before Accepting AI Output#
- The agent did not infer data classification or ownership.
- I verified storage behavior from current policy/source documentation.
- I know which files are authoritative and which are disposable.
- No secret value appears in the location map.
Evidence#
Submit the approved map, a safe .gitignore diff when relevant, and one
sentence explaining how final results can be restored or reproduced.
Ask For Help When#
Classification, licensing, consent, ownership, approved cloud use, or retention is unclear.