{
 "name": "A voice for a question with no survey history: the split by party and age, given the national level",
 "registered": "2026-10-05",
 "why": "Answering as each person, a frontier model caricatures groups (results-gunlaw, results-battery: 16 to 20 points off; those stay as published). Asked once per question for every group's share, with its group differences shrunk by a factor learned on other questions, it becomes useful where records cannot help: a question nobody has asked before has no history to fit. This test asks whether that split is better than giving every group the national share, and how often it lands inside the survey's own sampling noise.",
 "data": "General Social Survey 2024 (release 3a), weights wtssps; cells party (Democrat incl. leaners / independent / Republican incl. leaners) x age (18-44 / 45-89)",
 "questions": {
  "natroad": "highways and bridges",
  "natcity": "solving the problems of the big cities",
  "natdrug": "dealing with drug addiction",
  "natenrgy": "developing alternative energy sources",
  "natsci": "supporting scientific research"
 },
 "answer_scored": "share answering too little among too little / about right / too much",
 "voice": "gemini-3.8-flash, the newest model available to our key, 6 asks (ask_fresh.py run twice), thinking low, the prompt of ask_fresh.py: six group shares and the all-adults share in one JSON reply; cell values averaged over the asks",
 "method": "on the logit scale, the voice's cell values are pulled toward their mean by lambda, then shifted so the cell-size-weighted mean equals the 2024 national share (the claim is the split given the national level). Lambda is chosen on the grid 0.30-1.30 (step 0.05) by mean cell error over the 22 development rows (11 questions x 2022, 2024) in dev-cells-flash.jsonl (gemini-3.8-flash), committed here; no fresh answer is used to choose it.",
 "baseline": "every cell gets the national share",
 "floor": "per question, the 95th percentile of six-cell error between observed cell shares and 400 within-cell bootstrap resamples (seed 7), as protocol-floor.json",
 "primary": "both: (a) mean cell error over the five questions lower for the voice than for the baseline; (b) the voice inside its floor on at least 3 of the 5 questions",
 "development": "with gemini-3.8-flash on the 22 development rows the method was inside the floor on 15 and better than the baseline on 18; mean error about half the baseline's",
 "disclosure": "natsci's 2024 cell sizes and floor were computed before (develop-floor.py), its shares never printed. Only respondent counts were read for the other four. The model may have seen published GSS 2024 tables in training; the question wording is given without results.",
 "spend_cap_usd": 3,
 "publish": "results-voice-split.json whatever the outcome",
 "amended": "2026-10-05, before any fresh question was asked: registered first with gemini-3.8-flash and gemini-3.1-pro-preview together (0543ae8); changed to gemini-3.8-flash alone, the newest model, at the owner's direction"
}