# Edition 01 data dictionary **Cutoff:** September 12, 2026, America/New_York. **Currency:** USD throughout. These are reviewed case extracts, not a comprehensive grants or relationship database. ## Files - `sources.json`: 13 relied-on primary-source records; publisher, original URL, document date (nullable), precise locator, supported claim and limit, exact retrieval time, method, request body, HTTP status, SHA-256 and provenance hash. - `entities.json`: canonical institutional names and relevant identifiers. The original WIV label is retained. A missing identifier is not a resolved match. - `relationships.json` / `.csv`: 11 affirmative documented relationships; each has direction, a typed claim, source IDs, valid-time semantics, observation date, financial-observation references and specific grant relevance. - `amounts.json` / `.csv`: 10 source-specific observations. String values retain decimal precision. They are **not additive**; several describe the same award or relationship using different measures and dates. - `subaward-rows.csv`: all seven EcoHealth rows returned at this cutoff, retaining original recipient labels, API row IDs, action dates and amounts. - `nih-award-years.csv`: all seven returned EcoHealth award-year records, preserving year, application ID, budget and project dates, amount and `is_active` field. - `search-boundaries.json`: five scoped nonfindings/limits; these are not affirmative edges, risk scores or certificates of absence. CSV list fields (`source_ids`, `amount_ids`) use `|` between values. Empty date cells represent missing/unspecified bounds; they do not mean “continues forever.” The JSON version uses `null` for unknown dates. Some dates have month precision (`2021-07`, `2023-01`) because the source specifies only a month. The source register's `hash_basis` states how to reproduce its `sha256`: parsed API responses use canonical UTF-8 JSON (sorted keys, compact separators, `ensure_ascii=False`); HTML/PDF sources use the exact saved bytes. Original HTTP body hashes remain in retrieval receipts; a reserialized JSON wrapper naturally has a different file hash. No timestamp is included in the canonical response hash. ## Financial measures - **NIH award-year amount:** the selected fiscal-year record and its budget period. - **USAspending total obligation:** the retrieved prime aggregate and that source's performance period; not a payment total. - **OIG historical award amount:** Table 5/6 snapshot in the audit's FY2014–FY2021 window, not the current USAspending obligation. - **OIG historical amount spent:** Table 5 EcoHealth spending as of July 2021, not a WIV-only expenditure. - **Sum of reported subaward rows:** arithmetic over the explicitly listed API rows; dates are reported action dates, not verified bank-transfer dates. - **OIG Table 7 subaward amount:** the original table's measure. The period is the federal award project period. The footnote selects subawards with expenditures by July 2021; it does not turn this into a paid-cash total. The main comparison is **A06 − A05 = $111**, unresolved. The School of Public Health comparison is **A08 − A07 = $3.90**, also unresolved. Do not add A05/A06 or A07/A08. Do not pass the full prime amount through every related institution. There is **no amount observation of $0 for DKU**. ## Relationship types and dates `prime_award` and `reported_subaward` are the only grant-funding edge types here. `institutional_affiliation`, `prc_public_institutional_status`, `ccp_institutional_governance`, `founding_partnership` and `historical_board_representation` have separate meanings; none moves grant dollars. `institutional_support` records Duke's statement of Kunshan city investment in DKU. That statement is financially worded, but supplies neither amount nor a link to either NIH award. It must not be treated as a U.S.-grant transfer. Read `date_kind` before interpreting `valid_from` and `valid_to`: - `source_record_performance_period` / `nih_budget_period`: the specified source interval, not necessarily every other source's interval for that award. - `reported_action_date_range`: first/last action in the returned subset, not a complete legal subaward term or cash-payment interval. - `dated_event` / `document_as_of`: evidence at that specific event/document date. Equal from/to dates do not mean the institution existed for only one day. - `work_year`: the year the institutional report describes; it was published later. - `documented_by_and_observed`: the relationship was documented by the from-date and observed again on the undated institutional page at cutoff. This **does not establish a legal commencement date or prove every intervening day's status**. `observed_at` is the edition's observation date; source records carry exact UTC retrieval times. `document_date` is not interchangeable with the retrieval time or a URL's migration folder. USAspending source dates marked as last-modified retain that distinct meaning in the source note. ## Identity and nonfindings EcoHealth and Duke joins use the award number and matched UEI. WIV is an analyst-reviewed **institutional** match against the same-award OIG table and its own dated institutional publication. The USAspending list provides no Chinese registration identifier. Its “capital construction” suffix is retained, not silently normalized into a new legal finding or construction-purpose claim. `wu-sph`, `wuhan-university`, `wiv`, and `dku` are not interchangeable. No funding edge is created from the School of Public Health to DKU simply because Wuhan University is a DKU partner. No corporate ownership percentage or PLA edge is asserted. WIV Party-governance sources support institutional context, not grant-specific direction. Negative or incomplete searches live in `search-boundaries.json`, not as zero-valued funding edges. No model-derived risk score is generated. ## Rights and revision The downloads provide factual extractions and source attribution. Original institutional publications can have their own copyrights. Local full-source receipts are research evidence, not a licensed institutional mirror. Keep dates, definitions, original labels and source IDs when reusing the extracted facts. Change a conclusion only with new reviewed evidence. Preserve superseded values, note the affected source/claim, and issue a dated revision. The one-shot refresh script does not automatically publish a new edition.