0001-immutable-derived-labels.md

ADR 0001: Keep source labels immutable

• Status: accepted
• Date: 2026-08-15

Context

Teacher predictions may be useful evidence but are not ground truth. Writing them directly into source labels makes rollback difficult, obscures provenance and can silently contaminate later experiments.

Decision

The recovery pipeline treats the source labels/ tree as read-only. Apply mode copies labels into labels_autofill_v1, appends accepted AUTO candidates there, and can materialize a separate trainable dataset. Every candidate remains represented in an audit CSV.

Consequences

• Experiments are reversible and comparable.
• Extra disk space is required for derived labels; hardlinked images limit the larger image cost.
• Consumers must intentionally select the derived data.yaml rather than assuming the source changed.