The original plan.
The public protocol described current consumer-facing ChatGPT, Gemini, Claude and Perplexity experiences, at least three runs per exact prompt, six recommendation prompt families, and a separate evidence follow-up.
What Wave 1 actually ran.
Wave 1 used stateless API requests through Vercel AI Gateway with provider-native search or grounding. The frozen model IDs were OpenAI GPT-5.6 Sol, Anthropic Claude Sonnet 4.6, Google Gemini 3.1 Pro Preview and Perplexity Sonar Pro.
The final Gateway routes observed in the raw metadata were OpenAI via OpenAI, Claude via claudeaws, Gemini via Vertex and Perplexity via Perplexity.
The two material deviations.
First: Wave 1 measures API-model behavior under this execution environment, not the consumer applications described on the original study page.
Second: the frozen prompt file contained a separate evidence-follow-up sentence, but the automated runner did not send it as a second turn. Source findings therefore describe the citations and sources exposed during the primary grounded response.
Collection integrity.
The final Wave 1 dataset contains 72 primary observations: six prompts, four model families and three repeats. All four systems produced 18 unique provider request IDs and 18 unique answer hashes. No primary observation was duplicated during resume handling.
Other execution details.
Gemini/Vertex returned Google grounding redirect URLs alongside publisher names. Raw evidence preserves the redirect URLs; normalized domain analysis uses the publisher hostname exposed by Google when it is available.
No explicit Gateway response-cache bypass was set. Unique provider request IDs and unique answer hashes rule out simple identical-response reuse in the collected data, but they do not prove anything about internal provider-side caching.
Reporting boundary.
The public report does not rank the professionals named in the answers, does not call one model “best,” does not treat recommendation overlap as recommendation quality, and does not treat visible citations as proof of causal influence.