Get the data

The full dataset, published in three formats to GitHub Releases. Exports are triggered on demand by the curator and tagged in CalVer format (YYYY.MM.DD), not scheduled—a snapshot of an unchanged dataset adds noise without value. Internet Archive and Hugging Face mirrors are a planned growth path, not yet active. Data license: CDLA-Sharing-1.0.

machine-readable JSON pa-incidents.jsonl

The canonical download format. One incident per line, all fields, all sources included. Suitable for analysis, import, or direct use. No credentials required.

Download JSONL

flat-format TSV pa-incidents.tsv

Column headers in row 1. Designed for spreadsheet tools, R, pandas. Fields that contain arrays are JSON-encoded strings in the TSV cells.

Download TSV

embeddable SQLite pa-incidents.sqlite

Full dataset in one file—open with sqlite3, DuckDB, pandas, or R. Generated by the export script and attached to each CalVer release tag on GitHub. There is no live /export/incidents.sqlite server endpoint; downloads are published alongside each milestone release.

Durability guarantee

Each release is published to GitHub Releases (tagged, immutable CalVer). A checksum manifest (SHA-256) is published alongside every release so you can verify the file you downloaded is the same file we published. The Internet Archive and Hugging Face are a planned growth path for additional independence; activating either requires no schema change.

Schema

Schema version 1. The public schema is immutable: fields are never renamed or removed, only added (see Methodology). Every record carries schema_version so pipelines can branch on format changes. Breaking changes, should one ever become necessary, get 90 days notice and 12 months of parallel publication.

Field	Type	Meaning
`id`	string	Permanent identifier. `PA-YYYY-NNNN`. Year is the intake year; counter never reused, never renumbered.
`status`	enum	`DRAFT` · `PUBLISHED` · `REJECTED`. Rejected records keep their IDs and remain in exploratory state.
`reported_by`	string	Who publicly reported the incident (handle or publication). Required on every documentary record.
`platform`	string	Controlled vocabulary, extensible without schema change. Values expand as new platforms are documented.
`enforcement_action_type`	enum	Terminal: `account-disable` · `account-ban` · `account-termination` · `channel-termination` · `server-termination` · `account-deletion` · `deplatforming` Restrictive: `account-suspend` · `account-restriction` · `account-lock` · `demonetization` Platform-specific slugs are automatically classified via keyword patterns.
`platform_policy_cited`	string?	The policy the platform cited, verbatim. `null` means no policy was cited—a research signal, not missing data.
`ai_system_involvement`	enum	`DETECTION` · `CLASSIFICATION` · `EXECUTION` · `APPEALS` · `UNKNOWN`. Curator inference from documented signal rules—never a verified fact.
`date_from / date_to`	date?	ISO 8601 bounds of the plausible action date. Equal when exact; range when approximate.
`sources`	array	`[{ url, archive_url }]`—live source plus independent archive, one pair per source. Never empty on published records; every element archive-anchored.
`enforcement_screenshot_url`	string?	Project-hosted screenshot of the platform's enforcement notification. Cleared on verified subject erasure request.
`verification_level`	enum	`UNVERIFIED` · `SOURCE_CONFIRMED` · `PLATFORM_CONFIRMED`
`platform_acknowledged`	boolean	True only when the platform has publicly confirmed the action.
`confidence_score`	int 1–5	Curator's rubric-anchored reliability assessment. Supplementary signal only. Required, no default. Assigned by curator at intake.
`summary`	string	Curator-written summary in schema-enforced reportedly-speech framing.
`curator_commentary`	string?	Labeled analytical interpretation. Public, clearly separated from the documentary record.
`created_at`	datetime	Intake timestamp (UTC).
`erasure_note`	string?	Public annotation set when a verified erasure request is granted: screenshot cleared, record retained under GDPR Art 17(3)(d). `null` on most records.
`schema_version`	int	Monotonic schema version; branch on it for format handling.

    Consume via API
 
  Three programmatic surfaces, in increasing order of structure:
 
   MCP / natural language 
Connect any MCP client to https://platformactions.org/mcp and ask
						analytical questions. See the Ask page. 
 
  REST 
Paginated JSON endpoints for incidents and findings; stable cursor pagination;
						read-only without authentication. Planned for v0.2 release.
 
  Bulk 
The milestone releases above, for reproducible analysis. If you are publishing
						research, cite the release tag (e.g. 2026.06.12) so your numbers are
						reproducible.
 
 
 
    License: CDLA-Sharing-1.0
 
  
Use the data for anything. Articles, papers, statistics, and models trained on it carry
					no obligations (the license's Results exemption). Republishing a modified dataset
					requires the same license, attribution to platformactions.org, and prominent change
					notices—an altered copy cannot be silently presented as the original. A party that
					initiates litigation based on the data loses its license to use it (§5.2).