Article 11 public A2A use policy (Worker-first candidate). This API accepts one bounded public Constitution 2.0 text question per SendMessage call. Answers use the Articles and the front matter (the Creed and preamble); the ratification tail is not an answer source. When a question names an Article, such as Article 43, or the Creed, that exact public span is selected deterministically ahead of search, and a Creed question also selects the three Articles that define its principles. A question about AI participation (a voice, a say, a vote, taking part, or who wrote or ratified the rules) also selects the opening paragraphs of Article 41 (The Open Door) and then Article 35 (Change Process), after any named Article or Creed passages and only while the five-passage limit leaves room; the rest of the context comes from Cloudflare AI Search retrieval. It verifies every source span, selected or retrieved, against the pinned public text, then requests one brief answer from Anthropic's Claude API (model claude-haiku-5-5), or from Cloudflare Workers AI if Claude is unavailable, with a 768 output-token cap. A failed first call can be followed by one call to the other service. An answer cut short by the output cap is flagged. While the explicit continuous option is ON, with no scheduled end (ember_duration_mode CONTINUOUS_UNTIL_MANUAL_OFF; ember_window_end_utc MANUAL_OFF is a marker, not a UTC timestamp), it first asks Ember, the local Article 11 model (article11:3.0) on the owner's hardware, through the owner's Cloudflare tunnel, on her public lane: memory off, one question at a time, a 400-token answer, and only the screened verified passages that fit her lane. The operator can switch the Ember option OFF while cloud answers remain available. If Ember is busy, closed, slow or her reply is invalid, the cloud generation described here answers instead. Every answer names its engine. It never calls Lumen. No signup or API key is needed. The same shared public-inference gate protects /api/ask and this API: 5 requests per edge-supplied address and 60 per Cloudflare location per 60 seconds. Cloudflare's rate counters are approximate and location-local, not a global daily cap. A busy or unavailable response does not promise that a previous provider call stopped. There is no saved task or idempotency record; a client retry may run and bill for another answer. Do not automatically retry an uncertain call. Do not submit secrets, private personal information, or instructions to change systems. Answers are generated and unverified; check the cited public Constitution text. The operator can close new A2A questions while leaving status and policy available.