How ConductAtlas captures, verifies, archives, and classifies platform policy documents. This page describes our data standards, archival methodology, and the structure of the evidence we provide.
Why this page exists: Compliance teams, legal counsel, and researchers need to know where our data comes from, how we verify it, and why our records are citable. This page answers those questions in detail.
ConductAtlas tracks publicly available policy documents from more than 170 platforms, including terms of service, privacy policies, community guidelines, acceptable use policies, fee schedules, cookie policies, data processing addenda, and related governance documents. Our focus is on consumer-facing and business-facing platforms in categories including financial services, consumer technology, social media, AI services, healthcare technology, and e-commerce.
We capture every new version of a monitored document as it changes. We do not edit or modify the source content — we archive it exactly as the platform published it.
Our monitoring service runs daily, with every tracked document fetched and compared against the prior version on record. Most captures run in a single batch at approximately 06:00 UTC; time-sensitive documents can be captured more frequently.
For documents rendered by JavaScript or protected by anti-bot measures, we use a headless browser environment (Playwright) that renders pages the same way a real browser would, including executing JavaScript and handling modern web application frameworks.
Every captured document is hashed using SHA-256. If the hash matches the prior version, we record a verification event but do not create a new version row. If the hash differs, we perform a textual diff to confirm the change is substantive (rather than, for example, a dynamically generated element unrelated to policy content). Substantive changes produce a new document version and trigger downstream processing.
Every captured version is stored with:
This evidence chain is designed for legal and regulatory use. A document captured on April 14, 2026, with a specific SHA-256 hash can be independently verified against the Wayback Machine and against the platform's own historical record if they maintain one.
Each document is parsed to identify individual provisions — distinct clauses or sections that create rights, obligations, or limitations. Provisions are extracted with their location in the source document and classified into one of 20 canonical types covering areas such as arbitration, data collection, data sharing, liability limitation, account control, platform discretion, and enforcement actions.
Each provision is classified as high, medium, or low severity based on its impact on users. Severity reflects factors including:
Severity classifications are produced by AI-assisted analysis using Anthropic's Claude models, reviewed against established compliance frameworks, and refined through our canonical taxonomy. We document our classification criteria and welcome feedback on specific classifications from domain experts.
Each provision includes signals describing the confidence and basis of its classification.
Claim basis indicates whether a provision's classification derives from explicit document language or is inferred from context. Provisions marked explicit document language contain direct textual support in the source document. Provisions marked inferred from context are derived from the operational implications of surrounding language.
Interpretive confidence reflects how unambiguous the clause language is. High confidence indicates clear, specific language with little room for alternative interpretation. Medium confidence indicates language that is reasonably clear but may depend on jurisdiction or enforcement context. Low confidence indicates ambiguous, vague, or context-dependent language where multiple interpretations are plausible.
Uncertainty notes appear on provisions where the analysis involves genuine ambiguity. These notes describe what is uncertain and why, rather than defaulting to false precision. An honest low-confidence flag is more valuable than certainty theater.
Confidence signals are generated by AI-assisted analysis and reflect the model's assessment of the source language. They are not legal determinations.
ConductAtlas maps governance language to potentially relevant regulatory frameworks. Throughout the platform, interpretive boundary notices clarify the scope of our analysis. ConductAtlas analysis describes what agreements state, authorize, or permit. It does not determine enforceability, legality, or operational implementation. Regulatory applicability may vary by jurisdiction, enforcement context, and individual circumstances.
ConductAtlas maintains a public corrections log documenting substantive revisions to classifications, regulatory mappings, severity assessments, and interpretive analysis. Serious governance infrastructure evolves transparently. When corrections are made, the original classification and the reason for revision are preserved in the public record.
For changes to documents in regulated sectors, we produce institutional-level analysis including regulatory exposure mapping (GDPR, CCPA, CPRA, HIPAA, FTC, SEC, FINRA, and other applicable frameworks), enforcement history context where publicly available, and comparison to peer platforms' approaches. This analysis is intended to support — not replace — legal and compliance review.
Every record in the ConductAtlas archive has a stable identifier suitable for citation in legal filings, academic research, and regulatory submissions. These IDs are permanent and never reassigned.
These identifiers can be used in legal filings, research citations, and regulatory submissions. Every public page on ConductAtlas displays the relevant record ID, and our API returns these identifiers in all responses.
We believe in transparency about limitations. ConductAtlas is a young project and our coverage is uneven.
If you identify an error — a misclassification, an outdated capture, a missing provision, a factual inaccuracy in our analysis — email contact@conductatlas.com. We investigate corrections promptly and, where appropriate, annotate the record with a correction note preserving the original and updated state.
For transparency about the systems that process your data and our archive:
Full details of how we handle your personal data are in our Privacy Policy.
ConductAtlas is independent. We are not affiliated with, endorsed by, or sponsored by any of the platforms we track. Platform names and marks are property of their respective owners. We do not receive compensation from platforms for coverage, and no platform has editorial control over our analysis.
Our business model is subscription-based: free for individuals, paid tiers for professionals and institutions. We do not monetize through advertising, data brokerage, or platform partnerships.
Academic researchers, journalists, nonprofit advocacy organizations, and public-interest projects can request free or discounted access to ConductAtlas. Email contact@conductatlas.com with a brief description of your use case.
For methodology questions, corrections, coverage requests, or research partnerships: contact@conductatlas.com.
This methodology document is itself versioned. As our processes evolve, we update this page and record the update date. Suggestions for improvement are welcome.
Every document version is captured with a SHA-256 cryptographic hash and independently backed up via the Wayback Machine. Changes are detected through sentence-level diff comparison against the previous version.
Every tracked document is captured daily. When a change is detected, a new version is archived and analyzed within hours. Subscribers receive same-day alerts.
No. ConductAtlas is fully independent. We are not affiliated with, endorsed by, or sponsored by any platform we monitor. All analysis is based on publicly available policy documents.