--- pretty_name: Ontario Lobbying & MPP Disclosure Graph (oico-dd structured projection) language: - en tags: - government - lobbying - transparency - accountability - canada - ontario - public-records - entity-resolution size_categories: - 100K **Read this before using the data:** every record here describes **lawful, > public filings**. Registration to lobby is a legal, regulated activity. > Nothing in this dataset is evidence of wrongdoing, contact, outcome, or > influence, and it must not be presented as such. See > [Limitations](#limitations). ## The discipline lives in the data This dataset was built so that its epistemic limits are expressible **from columns, not just prose**: - Every resolved or derived reference (`person_id`, `org_id`, `riding_id`, `govbody_id`) travels with a `confidence_tier` / `resolution_tier` and a `match_method` **on the same row**. - A `NULL` reference means *not resolved* — the verbatim source string is still on the row. Under-resolution is exported honestly; no column is filled by guesswork. - Every table carries a `source` column naming its upstream dataset, so a single parquet file separated from this card still states where its rows came from. - Identity is never a name. Raw name strings live in `name_aliases` observations; persons and organizations are resolution-layer entities with an audit trail (`canonical_name_source`, `match_method`, `reviewed`, `corroborations`). ### Confidence tiers | tier | meaning | safe use | |---|---|---| | `confirmed` | exact-string identity match across sources, OR human-merged **with** a second corroborating signal (see `corroborations`) | may be used in analysis without further qualifier | | `probable` | human-reviewed or normalized-equal name match, **no** second signal | qualify ("apparently the same person/organization") | | `possible` | fuzzy candidate, **not** human-reviewed | research-only; never in an external claim | | `ambiguous` | (held_by only) a by-election parliament with multiple holders whose hand-over date is not resolved; **all** holders are emitted, each marked ambiguous | filter out for any per-person claim, or handle the overlap explicitly | | `role_only` | (former_offices only) the declaration names a *class* of body ("a minister's office"), not a specific one; joinable at role level only | never attach to a specific ministry | **Under-merging is by design.** Two records that *might* be the same actor are kept separate unless the merge survives review. An under-merged graph understates connectivity; an over-merged one fabricates connections. When the source data could not support a link, the link is absent — that absence is information, not an error. ## Provenance All sources are public. Point-in-time capture; see `manifest.json` (`capture_windows`, `git_commit`, `source_db.sha256`, per-table row counts and SHA-256) for the exact snapshot identity of this export. | source | capture | |---|---| | Ontario Lobbyist Registry (lobbyist.oico.on.ca) | full active-set crawl, May–June 2026; parsed into the source database 2026-06-12 | | MPP Public Disclosure Statements (PDS) | fetched 2026-05-24; filings spanning 2005–2025 | | OLA roster (ola.org) | roster export generated 2026-05-26 (export id and source commit in `manifest.json`), imported 2026-06-11 | | OLA member pages (minister portfolios) | fetched 2026-06-12; per-row `source_url` and `source_sha256` | **The lobbyist capture is the registry's *current active set*.** De-registered historical filings are not in the registry's public search and are therefore absent here. Any count over past periods derived from this data (e.g. "registrations first filed during the 42nd Parliament") is a **lower bound**: only filings still active at capture time appear. The registry's grid shows a status column, but it was not persisted; rather than fabricate one, no `status` column is exported — all exported registrations were in the active result set at capture. This export is a derived, regenerable projection of a SQLite mirror, which is itself parsed from an archived raw-HTML capture. No fact is authored in this export that is not in the source database; `registrations.raw_html_path` and `registration_versions.raw_html_path` point into the (unpublished) raw archive for full traceability. **Deliberately not republished** (although present in the public source): registrant email, phone, fax, street address, and postal code — for all parties (registrants, clients, agents, additional parties). City, province, and country are kept. The full redaction list is `manifest.json → redacted_fields`. ## Methodology ### Entity resolution (persons, organizations) Names observed across sources (registrant names, firm names, client names, funder names, additional-party names, PDS MPP names) are clustered into entities only by: exact string identity; normalized identity (whitespace/diacritic/suffix normalization); or human-reviewed fuzzy matches, with a written review note. A merge is `confirmed` only with a second, non-name signal (shared address, co-registration, riding history — exported in `corroborations`). Unreviewed fuzzy candidates are **not merged and not exported as links**. ### Riding resolution (the lobbyist↔MPP bridge) Lobbyist filings never name an MPP — they target the *Office of the Member for <riding>*. PDS and the OLA roster name persons + ridings. The join therefore runs through a canonical `ridings` table; observed string variants (three different dash characters across sources, mojibake-era forms, renames) are `riding_aliases`. Exact and dash-collapse matches are `confirmed`; genuine rename candidates were human-reviewed and are never auto-merged. A registration is connectable to an MPP **only** via riding + term window: `registration_targets.riding_id` → `held_by` (riding, parliament, term dates) → `persons`. Whether to require the registration's `initial_filing_date` to fall inside the term is the analyst's choice — do so for any per-person claim. ### The person spine (OLA roster) and the one unavoidable name-bridge `held_by` comes from the OLA roster, keyed on a stable `member_slug` — never on a name. The single name-match in the pipeline is the bridge between PDS-resolved MPP entities and roster slugs (`persons.roster_link_method` / `roster_link_tier`): exact name → `confirmed`, normalized-only → `probable`, ambiguous or missing → **not forced** (the person remains in the data with `roster_linked = false`; ~70 PDS-era MPPs are such retained "bridge islands"). ### By-election ambiguity Where a parliament had multiple holders of one riding (by-elections), `held_by` contains **all** holders, each with `ambiguous = true` and `confidence_tier = 'ambiguous'`, sharing the parliament bounds. This is the honest shape: the source does not date the hand-over. Per-person analysis must filter to `confirmed` or model the overlap. ## The one assertion beyond the sources: minister portfolios The OICO former-public-office checklist is *generic*: "Minister of Cabinet" names no ministry. For the small set of ex-ministers who themselves filed consultant-lobbyist registrations, `minister_portfolios` pins the named person to the specific ministry/ministries they led, **sourced from their OLA member page** (the Legislative Assembly's own record; per-row `source_url` + `source_sha256`). This is the only place the dataset asserts a fact not found in the OICO source, and it is scope-limited and tiered: - The person↔OLA bridge (`minister_portfolio_bridge`) uses the same discipline as the MPP bridge: clean unique name match or nothing. Ambiguous matches and deputy ministers (appointed officials, no OLA page) are exported with `bridge_status` `ambiguous` / `no_match` and **no** portfolio rows. - A portfolio attaches to a `government_bodies` node **only** when the derived ministry name equals an existing lobbying-target ministry byte-for-byte after normalization (`body_match_kind = 'exact_current_node'`). Cross-era ministry renames are never forced: the best fuzzy candidate is recorded as review-only (`body_candidate_id`, `body_candidate_sim`) with a NULL `confidence_tier`. ## Limitations **Read before making any claim from this data.** 1. **A registration is registered intent, not contact.** It says "on or after this date, this lobbyist may contact public office holders about these topics." It is not evidence a meeting occurred, that anyone was contacted, or that anything was influenced. Subject matters and techniques are self-classified by the registrant. 2. **Registration ≠ effect, and absence ≠ inactivity.** A registration in this data records intent to lobby; it does not indicate the lobbying succeeded or that any outcome followed. Conversely, the absence of a registration does not mean no lobbying occurred — unregistered lobbying exists and is itself a separate compliance violation, invisible to a registry-based dataset. Both directions limit what can be inferred from presence or absence here. 3. **Co-occurrence is not causation.** "Lobbying targeted riding X while MPP Y held it and MPP Y disclosed an interest in Z" is three facts about public filings, not a narrative. This dataset must not be used to allege wrongdoing; lobbying and disclosure are the *compliance* mechanisms working as designed. 4. **Active-set lower bound.** See [Provenance](#provenance): historical counts are lower bounds; de-registered filings are absent. 5. **PDS lists interests, not amounts.** MPP disclosures are interest-presence prose (with limited exceptions, no dollar values). "MPP X holds an interest in Y" is supportable; "MPP X owns $N" is not. Cross-year changes reflect disclosure changes, not necessarily economic ones. Disclosure free text is exported verbatim (`*_text` columns); structured interest→organization edges were deliberately **not** extracted (an NLP pass with its own error rate — deferred). 6. **The targets grain trap.** `registration_targets` is RAW-mention grain: the same (registration, target, subject_type) repeats across lobbying activities within one filing. **Always `DISTINCT` over `registration_no` (and `subject_type` if relevant) before counting.** `target_row_count` in `government_bodies` is raw-mention count, not registration count. 7. **HELD_BY is parliament-level.** Term dates come from the OLA roster at parliament granularity; by-election parliaments are multi-holder and tier-marked `ambiguous` (see Methodology). 8. **Name-bridge tiers are load-bearing.** Any cross-source person claim must filter the relevant tier columns to `confirmed` (and should state the tier either way). `probable` links are name-only matches. 9. **Self-reported sources.** The Integrity Commissioner reviews filings for completeness, not truthfulness. 10. **Known warts, kept honestly.** A small number of mojibake-era name twins survive as separate entities (e.g. one registration carrying both `Matthew Dubé` and `Matthew Dubé` as filer aliases resolved to two entities, ≈13 registrations affected); merging them automatically would violate the review discipline, so they remain split (under-merge by design). A handful of riding string variants likewise remain uncollapsed. 11. **Resolution decisions are auditable, not infallible.** Match methods, review flags, and corroboration signals are exported so a third party can check any call. ## Intended use Research, journalism, civic transparency, and accountability analysis of Ontario's lobbying and disclosure regimes. Appropriate uses include aggregate analysis of lobbying activity, registered-interest mapping, and tooling that surfaces public filings with their confidence labels intact. **Out of scope / inappropriate:** presenting any row as evidence of misconduct; dropping tier columns and presenting `probable`/`possible` links as fact; using redacted-category contact information from other sources to re-identify or contact individuals; claims about dollar values of MPP interests. All individuals appear in their public, regulated capacities (registered lobbyists, filing MPPs, declared clients/officers of organizations) in records those individuals or their organizations filed under Ontario law. ## Licensing & attribution Not yet determined — this build is staged privately. Source records are public filings published by the Office of the Integrity Commissioner of Ontario and the Legislative Assembly of Ontario; an explicit license and attribution statement will be set before any public release. ## Schema Stable string ids shared across tables (same vocabulary as the project's graph projection): `person:slug:{member_slug}` / `person:ent:{entity_id}`, `org:{entity_id}`, `gov:{body_kind}:{normalized_name}`; ridings use integer `riding_id`; registrations use the registry's own `registration_no`. Dates are ISO `YYYY-MM-DD` strings (NULL where the source omits them); timestamps are ISO-8601 UTC. Free-text fields are predominantly English; a small fraction of filings (<1% of rows, mostly francophone organizations' `goal_text` / `client_business`) contain French prose, kept verbatim. `*_norm` columns are deterministic lowercase/whitespace normalizations used as join keys upstream — they are not display names. ### Core entities **`persons`** — one row per resolved person (lobbyists, MPPs, ex-office holders). Union of resolution-layer person entities and OLA spine persons who held a seat; merged where bridged. Historical roster-only MPPs (no seat row in scope, no captured filings) are intentionally out of scope (count in `manifest.json → notes`). Columns: `person_id`, `entity_id`, `member_slug`, `canonical_name`, `canonical_name_source`, `roster_name`, `is_mpp`, `is_lobbyist`, `roster_linked`, `roster_link_method`, `roster_link_tier`, `parliaments` (JSON array string), `confidence_tier`, `source`. **`organizations`** — one row per resolved organization (firms, clients, parents, beneficiaries, coalition members, funders). `org_roles` is a JSON array string of observed roles. Columns: `org_id`, `entity_id`, `canonical_name`, `canonical_name_source`, `org_roles`, `confidence_tier`, `source`. **`name_aliases`** — one row per verbatim name string observed in a source and attached to a person/org; the full audit trail of the identity layer. Columns: `alias_id`, `node_id`, `node_kind`, `raw_name`, `normalized_name`, `alias_source` (which source field the string was observed in), `source_ref` (e.g. the registration_no), `match_method` (`exact`|`normalized`|`fuzzy`| `manual`), `match_confidence`, `reviewed`, `review_note`, `confidence_tier` (of the parent entity), `source`. **`corroborations`** — one row per second-signal corroboration backing a `confirmed` merge: `corroboration_id`, `node_id`, `entity_id`, `signal_kind` (`shared_address`|`co_registration`|`riding_history`|`other`), `signal_value`, `source_ref`, `source`. ### Lobbyist registry **`registrations`** — one row per registration (current state, active set). Grain: `registration_no` (`CL`/`OL`/`PP` prefix = consultant / in-house organization / in-house persons-and-partnerships). `filed_on_date` is the last amendment timestamp. Columns: `registration_no`, `registration_type`, `lobbyist_number`, `initial_filing_date`, `filed_on_date`, `registrant_first_name`, `registrant_middle_initial`, `registrant_last_name`, `firm_name`, `city`, `prov_state`, `country`, `fetched_at`, `raw_html_path`, `source`. **`registration_filers`** — one row per resolved (registration, filer); `filer_kind` `person` (the registrant individual) or `firm` (the registrant organization) — both file. Columns: `registration_no`, `filer_kind`, `node_id`, `resolution_tier`, `match_method`, `source`. **`registration_versions`** — changefeed: one row per observed (registration_no, last_amendment_date). Columns: `registration_no`, `last_amendment_date`, `registration_type`, `raw_html_path`, `observed_at`, `source`. **`registration_clients`** — one row per client line (CL only). The resolved `client_org_id` joins on the verbatim name within the same registration; NULL = unresolved. Columns: `client_row_id`, `registration_no`, `client_name`, `client_norm`, `client_business`, `city`, `prov_state`, `country`, `client_org_id`, `resolution_tier`, `match_method`, `source`. **`registration_lobbyists`** — one row per named lobbyist line; `role` is `consultant` | `in_house_current` | `in_house_former` | `pp`. These names are NOT entity-resolved (deliberately — see `former_offices`). Columns: `lobbyist_row_id`, `registration_no`, `lobbyist_name`, `lobbyist_norm`, `title`, `commencement_date`, `completion_date`, `role`, `source`. **`registration_subjects`** — one row per subject line; `subject_kind` `subject_matter` (controlled topic vocabulary) or `lobbying_activity` (typed: Bill, Regulation, Policy/Program, …) with free-text `goal_text`. Columns: `subject_row_id`, `registration_no`, `subject_kind`, `subject_type`, `subject_text`, `subject_norm`, `goal_text`, `source`. **`registration_targets`** — ⚠ RAW-mention grain (limitation 6): one row per target mention per lobbying activity. `target_kind` ∈ `mpp` (riding-office; resolves to `riding_id`), `ministry`, `ministers_office`, `agency` (resolve to `govbody_id`), `other_institution` (kept verbatim, no resolved ref). `resolution_tier` is set only where a riding resolution is asserted. Columns: `target_row_id`, `registration_no`, `target_kind`, `target_name`, `target_norm`, `subject_type`, `riding_id`, `riding_match_method`, `govbody_id`, `resolution_tier`, `source`. **`government_bodies`** — one row per distinct (body_kind, normalized name) government body observed as a lobbying target, plus bodies named only by former-office declarations (`created_for_former_office = true`, e.g. the dissolved Ontario Power Authority). Columns: `govbody_id`, `body_kind`, `name`, `normalized_name`, `target_row_count` (raw mentions, not registrations), `created_for_former_office`, `source`. **`government_funding`** / **`government_funding_detail`** — the registry's "is the client/organization government funded?" answer, and one row per declared funder line with verbatim and parsed amounts (`amount_value` NULL where unparseable). Detail columns: `funding_row_id`, `registration_no`, `funder_name`, `funder_norm`, `amount_raw`, `amount_value`, `funder_org_id`, `resolution_tier`, `match_method`, `source`. **`related_parties`** — one row per additional-party line behind a CL registration; `party_kind` ∈ `parent_company`, `other_beneficiary`, `coalition_member`, `private_funder`. Columns: `party_row_id`, `registration_no`, `party_kind`, `name`, `name_norm`, `city`, `prov_state`, `country`, `org_id`, `resolution_tier`, `match_method`, `source`. **`registration_agents`** — "retained by agent" lines (CL only): `agent_row_id`, `registration_no`, `agent_name`, `agent_norm`, `city`, `prov_state`, `country`, `source`. **`communication_techniques`** — one row per declared technique line: `technique_row_id`, `registration_no`, `technique_text`, `technique_norm`, `source`. ### Revolving door **`former_offices`** — one row per former-public-office declaration line. `office_source = 'cl_consultant'`: the statutory checklist declared by the consultant who filed — entity-resolved, so `person_id` is set; `role` is a controlled vocabulary (the seniority class); `body_govbody_id` is set ONLY where the checklist names a specific body (OPG/OPA/IESO), tier `confirmed` — otherwise the row is tier `role_only` and must not be attached to a specific ministry. `office_source = 'employee'`: per-employee OL/PP grid lines — **not** entity-resolved, `person_id` NULL (attributing them to the firm's registrant would fabricate identity). Columns: `former_office_row_id`, `registration_no`, `person_id`, `employee_name`, `office_text`, `office_norm`, `office_source`, `role`, `body_govbody_id`, `body_kind`, `resolution_tier`, `resolution_method`, `source`. **`minister_portfolio_bridge`** — one row per distinct ex-minister / ex-deputy-minister person among CL registrants, recording whether they bridge to an OLA member slug: `bridge_id`, `person_display`, `person_norm`, `role_class`, `lobbyist_number`, `registration_nos` (JSON array string), `member_slug`, `bridge_status` (`bridged`|`ambiguous`|`no_match`), `bridge_method`, `confidence_tier`, `bridge_note`, `source`. **`minister_portfolios`** — one row per ministerial portfolio parsed from a bridged person's OLA member page (see [the one assertion beyond source](#the-one-assertion-beyond-the-sources-minister-portfolios)): `portfolio_id`, `member_slug`, `person_display`, `role_class`, `parliament`, `role_title_raw`, `term_start`, `term_end`, `ministry_candidate`, `ministry_norm`, `body_govbody_id` (set only when `body_match_kind = 'exact_current_node'`), `body_match_kind`, `body_candidate_id`, `body_candidate_sim` (review-only fuzzy candidate), `confidence_tier`, `source_url`, `source_sha256`, `source`. ### The riding bridge & temporal spine **`ridings`** — one row per canonical electoral district: `riding_id`, `canonical_name`, `normalized_name`, `notes`, `source`. **`riding_aliases`** — one row per observed riding string variant: `riding_alias_id`, `riding_id`, `raw_name`, `normalized_name`, `alias_source` (`mpp_disclosure`|`lobbyist_target`), `source_ref`, `match_method`, `reviewed`, `review_note`, `source`. **`held_by`** — one row per (riding, person, parliament) holding; by-election parliaments emit multiple rows marked `ambiguous` (limitation 7). `term_end` NULL = ongoing. Columns: `held_by_id`, `riding_id`, `riding_raw`, `member_slug`, `person_id`, `parliament_number`, `term_start`, `term_end`, `term_start_raw`, `term_end_raw`, `ambiguous`, `confidence_tier`, `source`. ### MPP disclosures **`disclosures`** — one row per disclosure FILING EVENT, grain (mpp, year, as_of_date) — an MPP can file twice in a year (by-election + year-end), and both rows coexist. Interest panels are verbatim free text (limitation 5). Columns: `disclosure_id`, `person_id`, `mpp_id`, `mpp_name`, `riding_raw`, `year`, `as_of`, `as_of_date`, `income_text`, `assets_text`, `liabilities_text`, `gifts_text`, `offices_text`, `spouse_income_text`, `spouse_assets_text`, `spouse_liabilities_text`, `person_resolution_tier`, `fetched_at`, `source`. ## Example: the anchor query "Registered lobbying intent aimed at the office MPP X held, in-term, confirmed-tier only" — directly from the parquet: ```python import pandas as pd persons = pd.read_parquet("data/persons.parquet") held_by = pd.read_parquet("data/held_by.parquet") targets = pd.read_parquet("data/registration_targets.parquet") regs = pd.read_parquet("data/registrations.parquet") pid = persons.loc[persons.canonical_name == "Michael Mantha", "person_id"].iloc[0] terms = held_by[(held_by.person_id == pid) & (held_by.confidence_tier == "confirmed")] hits = ( targets[targets.resolution_tier == "confirmed"] .merge(terms[["riding_id", "term_start", "term_end"]], on="riding_id") .merge(regs[["registration_no", "initial_filing_date"]], on="registration_no") ) in_term = hits[ (hits.initial_filing_date >= hits.term_start) & (hits.term_end.isna() | (hits.initial_filing_date <= hits.term_end)) ] n = in_term.registration_no.nunique() # DISTINCT — limitation 6 ``` The result is a count of *registrations declaring intent to lobby that office during that tenure* — a lower bound (limitation 4), not a contact log (limitation 1). ## Reproducibility Generated by `oico-dd export hf` from the project's resolved SQLite mirror — deterministic: re-running over the same database reproduces byte-identical parquet files. `manifest.json` records the export timestamp, source-DB SHA-256, generating git commit, capture windows, redacted fields, and per-table row counts, byte sizes, SHA-256 digests, and grain statements.