KYC Drift currently holds 44 fact records about eight hosted-AI services. All 44 come from a provider’s terms, privacy policy, or help material and point to archived evidence. There are no controlled interface-run observations yet. This page therefore reads providers’ published positions; it does not report how the systems behaved in a fresh registration or controlled test.
What the pilot can support today#
The comparison has eight services and 17 dimensions, or 136 possible service-by-dimension fields. Forty-five fields contain at least one current scoped record; 91 contain no current scoped record. These counts describe the present inventory, not a completeness percentage or a claim that every scope was studied. Each record applies only to its displayed region, plan, platform, signup method, and other stated conditions.
Forty-four evidence-linked records are enough to expose concrete differences in published rules, but not enough to order the services on a single privacy scale. Every contrast below stands on its own dimension. You may care about phone requirements and not downstream routing, or the reverse. The pilot supplies dated evidence for those decisions without turning unlike choices into a score.
Three independent contrasts#
Provider documents describe different identifiers at account setup. Anthropic’s help material states that SMS phone verification applies to all new Claude users, cannot be skipped, rejects VoIP and several other number types, and limits each phone number to one account. DeepSeek’s privacy policy says that account setup may collect an email address and/or telephone number. This is a contrast between two official statements, not a claim that either registration path was reproduced here. The Claude record and DeepSeek record show the exact scope and archived text; the pilot has no interface run for either path.
“Delete” refers to different retention rules. Venice states that it does not retain prompts or outputs. Claude and ChatGPT state time-bounded removal for deleted data, subject to their stated exceptions. Mistral says chats remain until the user deletes them and describes longer retention for some account and identity data. Gemini states that some human-reviewed conversations can remain for up to three years even after activity deletion. DeepSeek allows chat-history deletion while describing retention more broadly as lasting as long as necessary. These are separate provider statements, not a finding that any one system enforced its policy.
Routers divide responsibility between two providers. OpenRouter and Venice both state that prompts can reach downstream model providers. OpenRouter says retention and training depend on the selected downstream provider and offers controls for choosing providers with particular data policies. Venice says it imposes zero-data-retention conditions downstream. The relevant question is not whether one architecture is better; it is which party receives the prompt, which written commitment applies, and whether that commitment has been independently tested. In this pilot, enforcement remains untested for both.
What the pilot cannot establish#
The dataset does not measure network anonymity, payment-intermediary exposure, or legal-process access. It does not test the product interface, confirm regional availability, or audit backend enforcement. A value labeled provider-stated means the cited document contained the claim when its archived body was checked. It does not mean CypherpunkGuide verified the provider’s system.
The 91 fields with no current scoped record are equally important. Not assessed means that the dataset has no current record for that service, dimension, and displayed scope; it never means the provider has no requirement or practice. The service pages expose these gaps so sparse documentation cannot look like a favorable result.
A monthly re-check is the operating target, not a live service-level promise. It becomes a committed cadence only after the capture, validation, diff, adjudication, and release pipeline completes two consecutive production-equivalent cycles with every gate GREEN and an observable heartbeat. Until then, a fact’s checked-through date is derived conservatively from its current supporting evidence streams. The pilot-wide checked-through date is the oldest such fact date, so a partial refresh cannot make the entire dataset look newer.
Corrections, disputes, and missing services: the methodology explains the rules. Write to editor@cypherpunkguide.com. If our review confirms an error, we publish the correction. We normally aim to acknowledge a dispute within 14 days, but that is an operating target rather than a guaranteed response time.