SDFT-style continual qualification evidence (2026-07-28)

Scope

This prerelease preserves the clean three-seed qualification of YOLOZU's
SDFT-style checkpoint-distillation continual-learning lane at source commit
d54ea3d2e8647c4742562e639872b2d9c4b83b7c.

Result

• Methods: naive sequential fine-tuning and SDFT-style checkpoint distillation
• Seeds: 11, 22, 33
• Evaluation: real pycocotools COCOeval over all 128 images per task
• Protocol: identical data, order, budget, and initial task state per seed
• Result: all mAP50-95 matrix cells were 0.0; all method deltas were 0.0
• Decision: hold
• Efficacy: not_established

This is a bounded repository-local Gaussian-blur diagnostic, not an external
benchmark or an independent reproduction. It validates execution, provenance,
and fail-closed promotion behavior; it does not support an efficacy claim.

Integrity

• Archive SHA-256:
  5dddd416f6b65c817cbdb4cd2bcef81d59a67a329c4f6497ea6dbd16e68e8bba
• Embedded checksums: 434 files, all verified before and after clean extraction
• The archive includes the machine-readable qualification summary, COCO
  reports, checkpoints, run provenance, promotion decisions, and checksums.