Methodology

How records enter this database, what they claim, what they don't, and the ethical posture behind every editorial decision.

Methodology

Every record describes a user-reported event, not a confirmed platform action—and this is enforced at the schema level, not as a style suggestion. The summary field is validated to contain reported-speech language; reported_by cannot be null; a record with an empty sources array cannot be published; platform_acknowledged defaults to false and is set true only on public platform confirmation.

This framing is not hedging. It is accurate. The database documents what people publicly reported—that is the fact being recorded, and it is independently verifiable through the archived sources attached to every record. Records asserting platform intent are not published; analytical inferences about patterns live in the clearly labeled curator-interpretation layer and in findings, never in the documentary record.

All v0.1 records are manually curated from public OSINT through a structured curation workflow—a guided agent skill that is the sole interface for creating records. A record physically cannot be created with an unarchived source. There is no admin portal; methodology lives in one enforced workflow, not in curator memory.

Berkeley Protocol alignment

Collection and verification practices are consistent with the Berkeley Protocol on Digital Open Source Investigations (UC Berkeley Human Rights Center / UN OHCHR, 2022). The intake workflow maps directly onto the Protocol's phases:

  1. 1
    Preliminary assessment

    Scope and credibility gate before any source is collected.

  2. 2
    Collection and preservation

    Every source archived independently at intake, before publication.

  3. 3
    Verification—authentication strand

    Is this source what it appears to be? Account history, notification UI consistency, signs of image manipulation.

  4. 4
    Verification—corroboration strand

    Do independent sources agree? Drives the verification level on every record.

  5. 5
    Chain of custody

    Who captured each source, when, and how—documented on every record.

§ Scope

Platform Actions documents platform enforcement actions where AI system involvement is reported or reasonably inferred : account disablings, account suspensions, content removals, appeal denials, shadowbanning, and reach restrictions attributed to the affected user or by available evidence to automated or AI-assisted platform systems.

An important constraint exists: platforms do not disclose whether a specific enforcement action was executed by an AI system, a human, or both. "AI-driven" cannot be confirmed from outside. What can be documented is what the affected user reported, what the enforcement notification said, the speed and scale of the action, and what is publicly known about the platform's moderation architecture. The ai_system_involvement field records the curator's inference from that evidence—it is an assessment, never a verified fact.

The dataset is Meta-focused because that is where the evidence is richest. The schema is platform-agnostic by design: no Meta-specific fields exist, and adding TikTok, X, or YouTube records requires no schema changes.

What we collect—and what we don't

Collected
  • Enforcement notice screenshots
  • Platform policy cited in the notice
  • Reporter's public account URL
  • Archive.today / Wayback snapshots of each source
  • Curator's written basis for AI involvement inference
Not collected
  • Post content or images that triggered enforcement
  • Email addresses, phone numbers
  • Home addresses
  • Financial data, government IDs
  • Anything the reporter did not make public

Categorical exclusions

Four categories of enforcement cases are excluded at intake regardless of record quality or source availability. These are structural scope decisions, not case-by-case calls:

  • CSAM and child safety policy citations. Any enforcement action where the platform's cited policy involves child sexual abuse material, child sexual exploitation, or child safety is excluded. The exclusion applies based on the policy cited—not on whether the underlying content actually violated that policy, and not on whether the platform later reversed the enforcement. Enforcement citing these policies is legally mandated under NCMEC reporting obligations and equivalent international law, which places these cases outside the category of discretionary platform judgment this dataset studies.
  • Adult content and sex worker accounts. Accounts primarily engaged in adult content creation or sex work are excluded. Enforcement in this population is largely legally compelled under post-FOSTA-SESTA (2018) frameworks rather than discretionary, making cross-case comparison unreliable—platform behavior cannot be isolated from legal context.
  • Violent extremism. Accounts primarily associated with violent extremism, terrorism, or incitement to political violence are excluded. Enforcement here is largely legally compelled through GIFCT hash-sharing, government referrals, and counterterrorism obligations. This exclusion does not apply to accounts that are themselves victims of over-enforcement due to proximity to extremism discourse—a journalist covering extremism whose account is wrongly flagged is in scope.
  • Documented platform ToS violations. Accounts suspended specifically for holding multiple accounts in violation of a platform's single-account rule are excluded. Enforcement for documented structural ToS violations of this type is not discretionary platform behavior and does not reflect AI misclassification.

If an original poster deletes their post, the live source link dies automatically — deletion by the subject propagates without the project lifting a finger. The independent archive remains. A platform-generated enforcement notice is the only tolerated PII exception: it authenticates the notice itself.

§ Verification

Every record carries one of three verification levels, assigned from the corroboration strand of the intake workflow.

PLATFORM CONFIRMED

The platform has publicly acknowledged the enforcement action. Rare by design — platforms almost never confirm individual actions.

HOW IT DISPLAYS

Solid green chip on the record and in the incident table. The strongest signal in the dataset.

SOURCE CONFIRMED

Two independent sources report the same incident, or the primary source includes a platform enforcement notification screenshot.

HOW IT DISPLAYS

Green outline chip. The record is corroborated but not platform-acknowledged.

UNVERIFIED

Source URL and archive URL exist and are accessible. Source has not been independently corroborated.

HOW IT DISPLAYS

Grey outline chip. The record is displayed normally and archived—not hidden or downranked.

A supplementary 1–5 confidence_score records the curator's rubric-anchored assessment of overall record reliability (1 = authentication concerns flagged; 5 = enforcement notification screenshot plus independent corroboration or platform confirmation). It is a research signal, not a filter—the rubric condition applied is logged on every record.

Inferring AI involvement

The ai_system_involvement field uses values DETECTION, CLASSIFICATION, EXECUTION, APPEALS, UNKNOWN. It is assigned based on observed signal hierarchy: an enforcement notification that explicitly states automation (strongest); instantaneous action with no prior warning and no human appeal path (strong); consistency with known automated enforcement patterns on that platform (moderate); and public reporting or transparency reports indicating AI use for that enforcement type (moderate). With no signal, the value is UNKNOWN—AI involvement is never inferred from the project's scope alone. The inference basis is logged on every record.

§ Chain of Custody

Every source on every record carries two URLs: the live source and an independent archive captured proactively at intake—before the source has any chance of being removed. X/Twitter sources are archived via archive.today (the Wayback Machine does not crawl X); other sources via the Wayback Machine's Save Page Now. No record publishes with an unarchived source. chain_of_custody logs who captured each source, when, and by what method.

The dataset is published to GitHub Releases on each milestone release—tagged CalVer snapshots, immutable once published—so that it survives Cloudflare account termination. Internet Archive and Hugging Face mirroring are a planned growth path, not yet active; activating either requires no schema change (see ADR-0005).

§ Legal & Privacy

LEGAL BASIS—PUBLIC-INTEREST RESEARCH

Platform enforcement records are research. Enforcement decisions affect speech, livelihoods, and access to community. This archive operates under First Amendment protection, the public-interest archiving exemption in GDPR Article 89(1), and the principle that facts—including the fact that a named platform took a named enforcement action on a named date—are not removable. See the full legal notice.

This is a principled position, stated openly—the same legal basis relied on by academic research databases, investigative journalism archives, and human rights documentation projects. The practical consequences:

§ Licensing

Code

MIT. The application, templates, and tooling are free to reuse without restriction beyond attribution.

Data

CDLA-Sharing-1.0 (Linux Foundation). Anyone may use the dataset for any purpose. Articles, papers, statistics, and trained models derived from it carry no license obligations—the choice to share derivatives is yours. What CDLA-Sharing requires is deliberate: anyone who republishes a modified dataset must keep it under the same license, attribute its source, and may not certify the altered copy as the original. That is the chain-of-custody guarantee, extended to the dataset itself.