# GENERATED by governance/data_contract.py — do not edit. # # Regenerate: python -m governance.data_contract --write # Check: python -m governance.data_contract --check # # Everything below is read from a register that already gates the build: the licence register # (docs/LICENSE_REGISTER.md), the catalogue (catalog/products.py), the published dictionary # (pipelines/publish/release.py), the freshness SLAs (quality/freshness_sla.yml) and the release # floors (quality/floors.yml). Editing this file by hand makes it disagree with the pipeline it # describes, which is the one thing a data contract must never do. dataContractSpecification: 1.1.0 id: certvas-af-fund info: title: African Company Fundamentals version: '2026.11' description: Point-in-time company fundamentals from primary filings, never exchange price feeds. owner: Certvas contact: name: Certvas email: mandase@certvas.com url: https://certvas.com/skus/af-fund.html terms: usage: "Licensed per the Certvas Data Licence Agreement. Every row carries `license_class`, the register\ \ code governing permitted use of the SOURCE it was extracted from — redistribution rights differ\ \ per source and the register is authoritative:\n - S001 GLEIF LEI Golden Copy (L1+L2) — CLEARED;\ \ basis: CC0\n - S009 African issuer annual reports / IR pages (company-published PDFs) — CLEARED_EXTRACT_ONLY;\ \ basis: Facts extracted from public disclosures\n - S012 Manually-curated public issuer identities\ \ (non-LEI) — CLEARED; basis: Public fact (name + jurisdiction + exchange ticker from public exchange\ \ listings)" limitations: "Derived from primary filings and official publications, never from exchange price feeds;\ \ Certvas holds no market-data licence and publishes no prices. Source conditions that travel with\ \ the data:\n - S001: None\n - S009: Extract facts only; no document republication; respect robots/ToS\ \ per site; keep source PDF ref in raw for provenance\n - S012: Identity facts only — no natural\ \ persons; each seed row cites its public listing; used only to give a non-LEI issuer a `dataco_id`\ \ so its S009 filings can be extracted\nAudit basis is not a column (added 2026-09-15). A vintage\ \ is one published document, and an issuer's FIRST vintage can be an unaudited full-year results announcement,\ \ with the audited statements arriving later as a restatement (a later `vintage_seq`, `restated` set\ \ on the earlier row). Known case: Oando PLC FY2023, unaudited results May 2024 (net income 74,723,154\ \ thousand naira) against audited statements November 2024 (60,277,168 thousand naira). `source_ref`\ \ names the document behind every value." billing: Certvas — African Company Fundamentals Sample (free) — FREE; Certvas — African Company Fundamentals (AF-FUND) — $129 noticePeriod: P30D models: gold_fundamentals: description: African Company Fundamentals — gold_fundamentals. type: table fields: dataco_id: type: string description: Stable canonical issuer key (LEI-anchored, or documented seed key for non-LEI issuers). period_end: type: date description: Reporting period end date. period_type: type: string description: FY / H1 / Qn — derived from the calendar against the issuer's fiscal year-end. statement: type: string description: income_statement / balance_sheet / cash_flow / per_share. line_item: type: string description: Taxonomy code, e.g. revenue, net_income, total_assets, eps_basic. value_reported: type: double description: Value in actual units (reporting scale applied; EPS left per-share). currency_reported: type: string description: ISO currency as printed in the source. value_usd: type: double description: CONVENIENCE ONLY — value_reported converted to USD. Null where no rate is cited. Reconcile and cite against value_reported + currency_reported, which are the as-reported figures; see conversion_ref for the rate used. conversion_ref: type: string description: The FX rate behind value_usd — source, rate and date. Null wherever value_usd is null. A converted figure with no citable rate is not published. extraction_confidence: type: double description: 0-1 model confidence for the extracted value. source_page: type: long source_doc_hash: type: string record_ref: type: string description: Stable id for this OBSERVATION (vintage), deterministic across rebuilds. vintage_seq: type: long description: 1 = originally filed, 2.. = each later restatement of the same key. is_as_reported: type: boolean restated: type: boolean description: A later vintage carries a DIFFERENT value for this key. Derived by comparing vintages; a re-observation of the same number is not a restatement. superseding_record_ref: type: string description: record_ref of the earliest later vintage that changed the value. Resolves within this table. observed_at: type: timestamp_tz description: When dataco captured the value. announcement_date: type: date description: AVAILABILITY date — when the filing became public. Null where no evidence tier produced a plausible date; never guessed. announcement_source: type: string description: 'Evidence tier behind announcement_date, in descending trust: disclosed (publication date stated by the issuer/exchange) / http_last_modified (the host''s Last-Modified for the filing) / pdf_creation_date (the PDF''s own metadata). Weight or exclude the weaker tiers as you see fit.' available_at: type: date description: FILTER AS-OF QUERIES ON THIS. coalesce(announcement_date, observed_at); never null. Where the announcement date is unevidenced it falls back to our own capture date, which is always LATER than true publication for a backfilled filing — so the fallback can only exclude a row from a window, never admit one before it was knowable. filing_lag_days: type: long description: announcement_date - period_end. Null where the announcement date is unevidenced — never defaulted to 0, because a lag measured against our capture date would describe when Certvas backfilled, not how fast the issuer reports. Per-market distributions ship in gold_filing_lag_stats. source_ref: type: string description: Source document hash + page, traceable to the filing. license_class: type: string description: License Register code governing permitted use of the source. valid_from: type: date valid_to: type: date quality: &id001 - type: custom engine: certvas description: 'arithmetic cross-validation pass rate: min 0.8. Enforced by exports/release_gate.py — below its floor a release auto-HOLDs and does not publish.' specification: arithmetic_pass_rate min 0.8 - type: custom engine: certvas description: 'dual-model agreement rate: min 0.95. Enforced by exports/release_gate.py — below its floor a release auto-HOLDs and does not publish.' specification: dual_model_agreement_rate min 0.95 - type: custom engine: certvas description: 'minimum extraction confidence: min 0.85. Enforced by exports/release_gate.py — below its floor a release auto-HOLDs and does not publish.' specification: extraction_confidence_min min 0.85 - type: custom engine: certvas description: 'Soda check failures: max 0. Enforced by exports/release_gate.py — below its floor a release auto-HOLDs and does not publish.' specification: soda_failures max 0 - type: custom engine: certvas description: 'Pandera schema failures: max 0. Enforced by exports/release_gate.py — below its floor a release auto-HOLDs and does not publish.' specification: pandera_failures max 0 - type: custom engine: certvas description: 'Arelle XBRL validation errors: max 0. Enforced by exports/release_gate.py — below its floor a release auto-HOLDs and does not publish.' specification: arelle_errors max 0 gold_fundamentals_as_reported: description: African Company Fundamentals — gold_fundamentals_as_reported. type: table fields: dataco_id: type: string description: Stable canonical issuer key (LEI-anchored, or documented seed key for non-LEI issuers). period_end: type: date description: Reporting period end date. period_type: type: string description: FY / H1 / Qn — derived from the calendar against the issuer's fiscal year-end. statement: type: string description: income_statement / balance_sheet / cash_flow / per_share. line_item: type: string description: Taxonomy code, e.g. revenue, net_income, total_assets, eps_basic. value_reported: type: double description: Value in actual units (reporting scale applied; EPS left per-share). currency_reported: type: string description: ISO currency as printed in the source. value_usd: type: double description: CONVENIENCE ONLY — value_reported converted to USD. Null where no rate is cited. Reconcile and cite against value_reported + currency_reported, which are the as-reported figures; see conversion_ref for the rate used. conversion_ref: type: string description: The FX rate behind value_usd — source, rate and date. Null wherever value_usd is null. A converted figure with no citable rate is not published. extraction_confidence: type: double description: 0-1 model confidence for the extracted value. source_page: type: long source_doc_hash: type: string record_ref: type: string description: Stable id for this OBSERVATION (vintage), deterministic across rebuilds. was_later_restated: type: boolean superseding_record_ref: type: string description: record_ref of the earliest later vintage that changed the value. Resolves within this table. observed_at: type: timestamp_tz description: When dataco captured the value. announcement_date: type: date description: AVAILABILITY date — when the filing became public. Null where no evidence tier produced a plausible date; never guessed. announcement_source: type: string description: 'Evidence tier behind announcement_date, in descending trust: disclosed (publication date stated by the issuer/exchange) / http_last_modified (the host''s Last-Modified for the filing) / pdf_creation_date (the PDF''s own metadata). Weight or exclude the weaker tiers as you see fit.' available_at: type: date description: FILTER AS-OF QUERIES ON THIS. coalesce(announcement_date, observed_at); never null. Where the announcement date is unevidenced it falls back to our own capture date, which is always LATER than true publication for a backfilled filing — so the fallback can only exclude a row from a window, never admit one before it was knowable. filing_lag_days: type: long description: announcement_date - period_end. Null where the announcement date is unevidenced — never defaulted to 0, because a lag measured against our capture date would describe when Certvas backfilled, not how fast the issuer reports. Per-market distributions ship in gold_filing_lag_stats. source_ref: type: string description: Source document hash + page, traceable to the filing. license_class: type: string description: License Register code governing permitted use of the source. quality: *id001 servicelevels: freshness: description: "Maximum age of the newest observation, per table, from quality/freshness_sla.yml. Suspensions\ \ are stated rather than hidden:\n - gold_fundamentals: 2 days" timestampField: observed_at availability: description: 'No uptime commitment. Certvas is a sole trader and the API is best-effort: automated checks alert the founder when it fails, but there is no on-call rota. Data is delivered as versioned files, which do not depend on the API being up.' support: description: No contractual response time. One person, one time zone (Africa/Gaborone, UTC+2). mandase@certvas.com is read daily on working days. backup: description: No stated recovery objective. Raw source documents are immutable and archived; a published pack is reproducible from them, but no RTO or RPO is promised. retention: description: No deletion schedule. Published versions are permanent by design — a point-in-time dataset whose history can be withdrawn is not point-in-time. Superseded versions stay resolvable via their DOI.