wtf is orchestra?

Orchestra is a public archive that repeatedly asks several AI systems the same set of questions across different topics, then keeps every response, good or bad, as an honest record instead of a polished summary. Each cycle sends eight questions covering areas like markets, research, technology, energy, geopolitics, business, demographics, and supply chains to five different AI systems, currently Grok, Perplexity, Gemini, DeepSeek, and Kimi. Every capture is preserved, then the same set of responses is handed to two separate synthesis processes so their readings can be compared side by side, and any tracked claim is followed over time until it is confirmed, proven wrong, fades away, or stays unresolved. The project's author, Jaron, says the idea started from wanting to build something like P.A.M., the information gathering robot from the game Fallout 4: a system that could gather scattered information and detect patterns while openly preserving uncertainty instead of presenting a guess as a fact. The same approach turned out to be useful across many different subject areas, which is how Orchestra grew into its current broader scope. The repository is organized into three layers. The acquisition layer opens fresh AI sessions, sends the prompts, and stores the raw responses without making any judgment about what they mean. The convergence layer synthesizes the results twice, declares what source material each synthesis read, compares the two readings, and updates tracked signals. The consumption layer, which currently includes financial, research, content, and monitoring uses plus a planned public website, only receives the finished, traceable artifacts. As of the README, the first full cycle from July 18 2026 collected 40 responses across the five sources, with 38 usable, 2 refused, and 45 total attempts once retries are counted. Orchestra states three specific promises to anyone reading its output: every claim can be traced back to its original prompts, captures, and model versions, disagreement between sources is kept visible rather than averaged away, and failures like refusals or unavailable services stay part of the visible record rather than being hidden. It does not promise that any given reading is correct, only that it is produced honestly from the available material. The project is released under the MIT License.