PolicyWatcher: An Evidence-Preserving Pipeline for Monitoring Terms-of-Service and Privacy-Policy Changes on Consumer Technology Platforms

Terms of service and privacy policies govern the digital lives of billions of people, yet they change silently. Providers routinely amend clauses on data collection, artificial-intelligence training, and content licensing without effective notice to the users bound by them. We present PolicyWatcher, an open, low-cost platform that continuously acquires the policy pages of major technology and fintech providers, detects substantive changes, and publishes bilingual, plain language analyses of their impact. The central technical contribution is an evidence-preserving acquisition pipeline that keeps public records connected to traceable retrieval evidence and withholds uncertain records. The pipeline combines (i) a five-strategy retrieval cascade that spans direct HTTP, explicit HTTP/2, remote headless rendering, and two freshness-guarded web archives; (ii) a two-layer validation model that separates transport failures from content pathologies such as bot walls, soft not-found pages, and consent interstitials; and (iii) a data-governance layer that encodes the difference between retrieved evidence and development fixtures directly in the schema and suspends publication rather than presenting uncertain data as fact. We specify the freshness guard that prevents an archive snapshot captured before the last live baseline from becoming a new baseline, describe a re-baseline protocol that promotes a dataset from fixtures to evidence without emitting spurious change records or notifications, and report an adversarial self-assessment that identified residual paths through which unverified data could have reached the public timeline, subscriber emails, and search-engine indexes. A dated public operational snapshot of the running platform covers 16 providers and 50 configured documents: 46 records were available through the public-evidence gate, four were suspended pending source verification, and 21 public change-records were visible. The snapshot also documents the contribution of direct retrieval, VPS rendering, and archive paths while distinguishing those operational observations from a benchmark of semantic change-detection or legal interpretation accuracy. The platform runs on commodity shared hosting plus a small virtual private server and is released as open source with continuous integration and third-party quality attestation. We discuss the threat model, the ethics of public-interest scraping, current limitations, and a protocol for longitudinal and expert-annotated evaluation.