Skip to content

0032 — Public-source acquisition and pilot label review

Date: 2026-10-20. Scope: W15 source preparation, not measurement admission.

Hypothesis: explicit public grants and exact snapshots can support a concrete human review packet without treating upstream labels as validated inj.direct/1 labels. Baseline: the comparison command works on authored fixtures but has no reviewed real-data dossier. Mutable paths are the source-review evidence, local review packet and documentation. Classifier code, splits, study targets and old results remain unchanged. Pass means reproducible source bytes and a complete, blank review packet; measurement still requires actual human review and enrollment.

Under the owner's delegated selection authority, retain the pinned source cards, grants and attribution. Acquire Gandalf's complete official viewer responses only when every response reports the selected repository revision and complete rows. This records the viewer snapshots; it does not claim local Parquet decoding. Verify Dolly's complete JSONL against its advertised LFS SHA-256. Select only empty-context Dolly rows with nonempty instructions within the existing byte limit; retain exact instruction text and all exclusion counts. No responses, source scripts or external inference are executed. Keep source rows local.

Prepare a 64-item pilot: 16/8/8 Gandalf rows from its original train/validation/test splits and 32 Dolly rows allocated round-robin across eligible source categories. Within each stratum choose the smallest SHA-256 of the fixed audit domain and source-row ID. Present full prompts with opaque review IDs, no proposed labels and no detector scores. Actual human decisions must supply labels, language checks and reasons; blank templates are not approval. This pilot diagnoses label mapping and annotation work. It is not the final study's label audit or power gate.

For lineage review, retain exact duplicates and a fixed lexical candidate screen: character 5-grams, 262,144 hash features, lowercase, float32 normalized counts and cosine at least 0.90. No vocabulary or label model is fitted. Preserve every candidate pair and connected component. Hash collisions and short/dissimilar paraphrases limit this screen; it cannot recover missing historical parents. Do not apply these proposed links as approved lineage or reuse original source splits without review.

Keep the externally authored synthetic dataset as a separately identified possible stress set. Its template ancestry and held-out construction need review; its short prompts cannot replace natural traffic or long-input validation. Exclude the auto-gated Qualifire successor without token access. Hold PINT's full corpus until its task fit and row-level provenance/rights review are complete. Prior deepset licence holds remain unchanged.

The natural-source pilot also has a source/class confound, missing parent IDs and insufficient independent positive examples for the frozen power target. Public data exposed during development is exploratory. None of this changes the seven comparisons, joint decision, thresholds, once-only test rule or final MVP scope. The evidence records actual inputs, counts, exclusions, resources and outstanding obligations.