KRUMMHOLZ DATASET PRIVACY REVIEW - 2026-10-23
Status: PRELIMINARY. Identified unnecessary linkage and residual fine-location details were removed. No unresolved private-identifier finding was identified by this review of the revised release. This is not anonymity certification.

SCOPE
All nine previously published dataset ZIPs and their exposed documentation were inspected, including both older homepage samples and the seven archives in the October 22 preliminary release. The new 37-excerpt LLM curation pilot was also inspected. CSV and JSONL alternatives, previews and review worksheets were included, not merely primary files. The audit covered complete schemas and all released natural-language text values; 30,019 unique text values were checked. 146 archive members and standalone files were parsed. All 87 supplied internal checksum entries and all nine existing ZIP CRC checks passed.

REMEDIATION
1. Legacy Maine forestry sample: retain all 10,000 scientific observations, remove source tree/plot/previous-record links, county/survey-unit/plot/subplot/tree/condition identifiers, source CSV line numbers and the two raw/context tables. Replace source record keys with content-derived local keys and remove the old rebuild script that recreated the discarded linkage fields. Blank reviewer templates now request an internal reference rather than personal identity. Scientific measurements, units and required public-source attribution remain.
2. Legacy BANKING77 sample: replace its unredacted source-query basis with the revised 13,006-record cleaned corpus (9,933 train and 3,073 test, all 77 intents). Remove source-row mappings and exclusion logs; regenerate previews and review worksheets from cleaned text. The earlier legacy sample had 13,071 rows. The 65-row net reduction reflects the newer corpus's privacy processing, duplicate/conflicting-label exclusions and unusable-text filtering; it does not mean that 65 newly confirmed private people were found. The readable glossary is still proposed and unreviewed.
3. Preliminary BANKING77: generalize five residual fine-location spans in customer incident narratives to LOCATION placeholders. Recompute the five changed content-derived record identifiers and update every CSV, JSONL and review-copy occurrence. No records, intent labels or split assignments were removed in this additional pass. Regenerate package checksums and the public catalog's ZIP size and hash.
4. Other existing preliminary releases: no additional private-identifier finding was identified. Forestry already excludes source plot and exact location linkage; recreational fishing retains broad state/wave/area and measured activity but omits fine survey linkage; robotics contains numeric trajectories and task instructions, with no images/video; service-query releases retain screened text and labels.
5. New LLM pilot: inspect all 37 historical excerpts and the supplied provenance/reproduction material. No private contact details, account identifiers or contemporary private-person narratives were identified. Historical citation names, scientific locations, a fictional name used to illustrate logic, and source-attribution metadata remain. They are relevant public context, not newly acquired customer information.

METHOD AND CANDIDATE REVIEW
Inspect archive membership, structured schemas and alternate text copies; scan identifiers, contact/address patterns, local paths, selected sensitive-person contexts and known residual location patterns; use named-entity and capitalization candidate review for the complete natural-language text inventory. The final pass produced 52 text candidates, all read in context. They were ordinary service/currency wording, already-redacted placeholders, scientific terminology, historical public attribution/geography or fictional logical examples. One documentation candidate was a regex literal falsely matching the local-path detector. A generic question about using an identity document contained no actual number or credential. Every one of the 37 LLM passages was read in addition to automated screening.

WHAT THIS DOES NOT ESTABLISH
Automated detectors can miss identifiers and over-redact ordinary terms. Removing direct links makes identification harder but cannot establish that public text or measurements are impossible to match to their upstream sources. Historical public names and citations remain deliberately. No blanket PII-free, fully anonymous, zero-risk, production-ready or legal-compliance claim is made. Historical text may contain outdated or unsuitable assertions; the separate content/rights review and package limitations address that use risk. Keep release documentation and source licenses with any redistributed copies. Do not add personal reviewer names, customer data or confidential notes to public worksheets.

The replacements preserve the old public download paths so that new downloads receive the revised bytes. Publisher/source attribution and scientific measurements are retained where relevant. Private review candidate excerpts and old-to-new row crosswalks are not included in this public report or release.