Meddies/meddies-pii-v3-skeleton

Meddies PII v3 — Skeleton Corpus

Typed-placeholder document skeletons for multilingual synthetic PII/PHI training data. Every PII-shaped slot is a ⟦label.subtype#id⟧ placeholder — no real values exist in this corpus by construction. A deterministic filler substitutes values downstream.

• 299,065 documents · 20 languages (incl. Khmer, Javanese, Sundanese)
• 9 coarse labels: address, company_name, date, email_address, human_name, id_number, phone_number, private_url, secret — plus op:/neg:/trap:/ link:/frag/mod annotations for contextual judgments
• Generated across 16 free-tier provider lanes via meddies-llm-runtime
• accepted/violations flag contract quality; provenance marks corpus/pilot/matrix/smoke; content_sha256 dedupes

Every published row passed the current gate and carries ownership contract v2 (contract_version), which says whose value each slot holds. Rows made under earlier contracts are withheld.

Known limitations

Organisation names are meant to be ⟦company_name.<kind>⟧ slots, but some rows still write one as plain text in the letterhead, sender, signature or stamp block. A blinded audit (2026-10-30, 300 rows, 75 per layout, two AI judges from different model families, agreed verdicts only, 296/300 agreement) measured these rates of rows with at least one plain organisation name there:

Layout	Leak rate	95% Wilson interval
header_only	4.0% (3/75)	1.4–11.1%
letterhead_signature	15.3% (11/72)	8.8–25.3%
stamped_sealed	11.0% (8/73)	5.7–20.2%
state_letterhead	10.8% (8/74)	5.6–19.9%

Weighted by layout size, about 9% of the rows in these four layouts (roughly 2.8% of the corpus) are affected. The rates are lower bounds from AI annotators: counting a row when either judge flags it gives 32/300. Most leaked names are invented; some name real public bodies or companies. Body text outside the header and closing blocks was not audited. Filter or relabel these blocks before training a model that must tag organisation names.

Row scope

Every row carries label_scope.

• full: generated under the current ownership contract. It has spans, person ids (#N) and benign look-alikes (op:, neg:, trap: slots). Use it for tagger training and for benign-versus-leak decision training. For decision training, filter label_scope == "full".
• span_only: an older row, recovered by promoting every non-PII slot on a PII label to a plain PII slot and then run through the current gate. contract_origin names its old stamp (ownership-contract-v1 or unstamped); it is null for full rows. Person ids (#N), :frag, :spoken, :obfuscated and link: forms are stripped, so every PII slot is an unbound ⟦label.subtype⟧. op:/neg:/trap: slots on non-PII labels stay, and so do the op:date schedule dates the gate writes and trap:date decoys (a reading or a dosing schedule shaped like a date). cast is null. Use it for BIOES span training only. It cannot train entity linkage, and its benign look-alikes on PII labels were promoted, so it cannot train a decision.

Why the ids are stripped: 320 documents (recovered rows, plus full rows as a control) were judged blind by Opus and Codex. 266 got an agreed verdict of defect or clean, split over the groups below, and 4 more were agreed unsure. Agreed defect rates, with counts in brackets:

Defect	Control (full)	Plain	v1	Date binding	Schedule date
Leak: a personal value as plain text	7.3% (8/109)	8.8% (6/68)	6.7% (2/30)	6.5% (2/31)	10.7% (3/28)
Hidden: a personal value behind op:/neg:/trap:	0/109	0/68	0/30	0/31	0/28
Wrong owner: a #N on the wrong person	5.5% (6/109)	23.5% (16/68)	16.7% (5/30)	12.9% (4/31)	21.4% (6/28)

"Plain" is a recovery from an unstamped row, "v1" from a v1 row, "date binding" a row with a sole-subject date binding, "schedule date" a row with a gate-written schedule date. Leaks match the control and nothing hides, so the spans hold up. Wrong owners run 13 to 24% against 5.5%, which is why the ids are gone.

span_only rows lean toward over-redaction: promotion tags catalogue codes, schedule dates and fictional sample values that the original document meant as non-personal. Until that is measured, mix at most one span_only row per full row when training a tagger.

Sister dataset to Meddies/meddies-pii (filled/labelled corpus). Generated 2026-09 to 2026-10 by the meddies-pii v3 pipeline.