A genuinely quiet Monday, and a paper arguing that the question list is the measurement.

Checked against the real repository this morning rather than a local copy: ten remote branches, none named for perplexity, sonar or agent; `main` still at commit 20b1172 dated 22 August 2026; `src/adapters.ts` line 14 still reads `perplexity: "sonar"` and line 91 still calls `https://api.perplexity.ai/chat/completions`. Perplexity's changelog is unchanged since the last check, still four September entries (custom MCP connectors, Perplexity OAuth for MCP, Gemini 3.8 Flash, GLM 5.3 Flash), none of which moves the date. The deprecation is independently confirmed in a pull request opened by a Perplexity employee, andrewmadson-pplx, against models.dev on 20 August, which marks `sonar`, `sonar-pro`, `sonar-reasoning-pro` and `sonar-deep-research` deprecated and was deliberately held as a draft because those models "remain functional until the sunset date."

This is a critical survey with a framework, not an experiment. The author is explicit about that limit: "No new experiments are reported; its general empirical validity remains to be assessed." Its argument has three load-bearing parts. First, "The prompt corpus selects the situations evaluated, while weights determine their relative importance. Together they define an 'answer market' that need not represent actual user demand." Second, "Prompt wording can alter retrieval, competing sources, and generated answers," so the questions change the answers rather than merely observing them. Third, on scoring: "If a language model performs this task, its instruction can change the score assigned to an unchanged answer." The paper also separates two things the category routinely conflates, noting that "A citation alone does not establish a source's contribution," and proposing a with-and-without comparison in a controlled documentary context instead.