Skip to content

Q-R1A Sample Corpus Extracts — MANIFEST

Three real PITH extracts (RTOpacks corpus, D1 rto-nrt-db 1249760d + R2 tga-content), per RTOPACKS-CORRECTION-01 §C-5. Real bytes off the real path — the TGA-API ingestion path (structured store + R2 raw-JSON archive).

PITH read-only compliance: every D1 query ran --remote and returned changed_db:false, rows_written:0; R2 objects were get only (--remote). Nothing in PITH was written.

The three units

Unit Title Selected as Why (one line)
SITTIND003 Source and use information on the tourism and travel industry CLEAN Current, plain-text content (numbered lists, \n), no HTML; the known-quantity control.
HLTAHA049 Recognise impact of health conditions MIDDLING Current, well-formed, but carries HTML entities (&) in content — a minor real-world mess.
MSL975041 Perform fire assay techniques UGLY Genuine mess two ways: performance_criteria is NULL (missing core field) AND HTML fragments + malformed spacing in knowledge_evidence. The authentic long-tail, not a curiosity.

Files + SHA-256 (9 files)

Per unit: {code}.structured.json (units row + qualification placements + one parent qual's packaging-rules context incl. retained raw_input + confidence), {code}.raw.json (R2 current raw TGA JSON), {code}.raw.{date}.json (R2 dated snapshot, the sync date the row was drawn from).

SITTIND003

  • SITTIND003.structured.json — ea5620197923aa3064ab6e5684f316322dcc282d0ba909f7d8a778ee919c2a20
  • SITTIND003.raw.json — d4eed3e53ccfdcdf3578c279919d94442b7dfc167d0d1559db0a864a6fc9c59a — R2 key training/SITTIND003/raw.json
  • SITTIND003.raw.2026-06-28.json — d4eed3e53ccfdcdf3578c279919d94442b7dfc167d0d1559db0a864a6fc9c59a — R2 key versions/training/SITTIND003/2026-06-28.json

HLTAHA049

  • HLTAHA049.structured.json — 483cb49486fbaf4693e9f8c7cbb98063bb96bf31ebc4a440956b209c979ddd09
  • HLTAHA049.raw.json — b8840e977d6a8bf433996408e63771d9077a0bd9d4fc9824c856f7ff2da3c3d1 — R2 key training/HLTAHA049/raw.json
  • HLTAHA049.raw.2026-06-28.json — b8840e977d6a8bf433996408e63771d9077a0bd9d4fc9824c856f7ff2da3c3d1 — R2 key versions/training/HLTAHA049/2026-06-28.json

MSL975041

  • MSL975041.structured.json — 3535e9f246a0e3ae8cfcf119203ac349fb474033984daba874ef193fcd074fab
  • MSL975041.raw.json — 8a9108028d37edf3e063ccd16ca346026df42dd6dc6e82a12a3c6a7712d80317 — R2 key training/MSL975041/raw.json
  • MSL975041.raw.2026-06-27.json — 8a9108028d37edf3e063ccd16ca346026df42dd6dc6e82a12a3c6a7712d80317 — R2 key versions/training/MSL975041/2026-06-27.json

Hash-chain mapping (RTOPACKS-CORRECTION-01 §C-3)

  • raw_document_hash = SHA-256 of the dated snapshot (versions/training/{code}/{date}.json) — dated, immutable once written, reproducible. (Note: current raw.json is byte-identical to the latest dated snapshot, so their hashes match here — the current pointer is the latest dated write.)
  • normalised_document_hash = SHA-256 of the {code}.structured.json extract (parsed content row + qualification packaging context) — the durable attestation of normalised state at job time (our structured layer is current-state-only; not reconstructable later — §C-3 disclosure).

Data-shape notes (real bytes, honestly)

  • release_number is NULL for every unit in the units table (all 15,198 current) — release/version lives in tga_training_releases, not on the unit row. So "NULL release number" is a data-shape fact, not per-unit corruption.
  • The structured (normalised) layer is remarkably clean corpus-wide: of 15,198 current units, only ~4 carry raw < HTML in PC/KE, ~51 carry & entities, ~90 are sparse. The ugly sample sits in that long tail deliberately.

Query shapes (reproducibility)

  • structured: SELECT * FROM units WHERE unit_code=?; SELECT * FROM qualification_units WHERE unit_code=?; SELECT ...,raw_input,confidence FROM qualification_packaging_rules WHERE qual_code=? (parent qual).
  • raw: wrangler r2 object get tga-content/training/{code}/raw.json --remote; dated: ... tga-content/versions/training/{code}/{date}.json --remote.