Objective and design constraints
Dataset Foundry establishes lineage, deduplication, assessment and evidence envelopes while preserving historical artefacts. A later R1.1 research corpus addressed a discovered row-specific identity/split coupling problem through regenerated material, cross-split grammar coverage, provenance checks and independent comparison.
Research qualification pipeline
Validation and current state
Dataset Foundry R0 preserved 44 source executions and canonical candidates while reporting zero training-ready targets and a training NO-GO. Separately, the R1.1 corpus and sealed holdout passed defined regeneration, provenance, disjointness and oracle checks. The corpus is curated rather than exhaustive.
- Claim
- Dataset qualification and holdout discipline were demonstrated for bounded research material.
- Method
- Regeneration, provenance, split, leakage and independent comparison checks.
- Result
- Defined R1.1 training/development and a separate sealed holdout were verified.
- Limitation
- R0 has no target ready; the qualified material is not exhaustive and does not establish a successful model or deployment readiness.
- Source period
- 2026 qualification records.
Failure, recovery and related systems
The discovered identity flaw was retained as a research lesson, not hidden. Corrected material does not erase the need for a fresh authorisation before any later research step. The current model result remains RESEARCH ONLY, with no production or superiority claim.
Nano Factory provides traceable campaign machinery; Evidence and External Memory carries the provenance and decision boundary.
Sources: Dataset Foundry qualification, evidence-contract and R1.1 corpus/holdout result records. Dataset contents, local locations and experimental identifiers are excluded.
Related revision
See KR-001: Dataset Identity Coupled Meaning to Row Metadata for the public correction record.