Skip to main content
  • Status: accepted · Date: 2026-07-17 · Deciders: Davor Runje

​
Context

Datasets differ in whether they can be committed, auto-fetched, or only manually acquired. The author wants a private copy of every dataset regardless. A standards pass grounded the policy (FAIR, licensing).

​
Decision drivers

  • Redistribution rights (may we republish bytes?) and access automation (can a machine fetch?) are orthogonal and both matter.
  • FAIR A1.2/A2: metadata stays open even when bytes are gated.
  • Legal correctness: a private mirror does not confer redistribution rights.

​
Considered options

  1. Three tiers — A committed (small + redistributable), B auto-retrievable, C manual/gated — with a per-project private mirror across all tiers.
  2. Two tiers (in-repo vs external).
  3. No tiering; commit everything or fetch everything.

​
Decision

Option 1. Tier set by license × access; tier proposed by the skill but confirmed by the human. Hard rule: tier/redistributable are set by the license, never by mirror presence; never promote to Tier A because a mirror exists.

​
Consequences

  • Reproducibility (Tier A/B) + legal safety (Tier C) + durability (mirror).
  • Tier assignment is a human-confirmed judgement.

​
Rejected alternatives

  • Two tiers — conflates the license and access axes.
  • No tiering — either bloats the repo or breaks on gated/large/non-redist data.
sub-spec 3 §3; digest dataset-management-standards.md.