Skip to content

Latest commit

 

History

History
33 lines (21 loc) · 1.45 KB

File metadata and controls

33 lines (21 loc) · 1.45 KB

Source-Task Forensics

Why this audit exists

The initial simple final-confirmation eligibility predicate reconstructed 20,829 source pairs, while the canonical v1b train/development manifests contained 20,800 pairs.

A source-only forensic audit was created to identify that discrepancy before proceeding with official validation.

Forensic boundary

The forensic audit:

  • reads source train manifests and source train annotations only
  • does not open official validation annotations
  • does not scan official validation images
  • does not load a model

Finding

Forensic field Value
Source manifest pairs 20,800
Baseline reconstructed pairs 20,829
Missing source pairs 0
Unexpected reconstructed pairs 29

No exact non-data-derived filter reproduced the canonical source pair set. The audit identified a data-derived category filter based on categories observed in the source manifest.

The final confirmation uses the source-rule reconstruction contract rather than assuming that a simpler predicate is equivalent. The official-validation manifest summary records a successful source-train-rule reproduction contract with 20,800 source train/development pairs.

Operational implication

For dataset-derived tasks, eligibility must be treated as part of the experiment specification. A plausible-looking predicate is not sufficient evidence that a final-validation sample construction matches the source task exactly.