{
 "name": "The model as a prior: does it halve the real respondents a survey needs for its split by party and age?",
 "registered": "2026-10-05",
 "why": "No published method, and none of ours, gets a model's group shares inside a survey's own sampling noise without real respondents (results-voice-split.json, not met, stays). The use that holds is the model as a prior combined with a small real sample. Developed on the 11 questions already scored (develop-prior.py): with the model counted as 50 respondents per cell, 50 real respondents per cell gave a mean six-cell error of 3.55 points, against 4.93 for 50 real alone and 3.53 for 100 real alone.",
 "data": "General Social Survey 2024 (release 3a), weights wtssps; cells party (Democrat incl. leaners / independent / Republican incl. leaners) x age (18-44 / 45-89); truth is the full 2024 survey's weighted cell share",
 "questions": {
  "nataid": "foreign aid: too little",
  "conjudge": "confidence in the Supreme Court: a great deal",
  "confed": "confidence in the federal executive branch: a great deal",
  "immfate": "unauthorized immigrants: become citizens",
  "polhitok": "approve of police striking a citizen in some situation: yes"
 },
 "selection": "chosen from 2024 questions that no run has scored or used as an input (only respondent counts read)",
 "model": "gemini-3.8-flash, 3 asks per question with ask_prior.py (frozen here): six group shares in one JSON reply; averaged; group differences shrunk on the logit scale by 0.85 (from development) and levelled to the small sample's own national share (cell-size weights from the full survey's composition)",
 "simulation": "score_prior.py (frozen here): per question, 400 draws (seed 20261005) of n real respondents per cell, drawn with probability proportional to weight; real-only estimate = the sample's cell shares; model estimate = (50 x prior + n x sample) / (50 + n)",
 "primary": "pooled over the five questions and all draws, the mean six-cell error of the model with 50 real respondents per cell is LOWER than that of 75 real respondents per cell alone",
 "secondary": "the same against 100 real alone; per question; the share of draws inside the full survey's 95th-percentile sampling floor for each arm",
 "spend_cap_usd": 1,
 "publish": "results-prior.json whatever the outcome"
}