{
 "name": "a battery of five measured splits: lives fitted to records, and a frontier-model voice, against the 2024 General Social Survey",
 "registered": "2026-10-05",
 "why": "One question can be picked to be easy. Five fresh questions, all reported, test whether lives fitted to records reproduce measured population splits in general, and whether a current frontier-model voice does. Earlier verdicts (welfare, gun permits) stand as published.",
 "honesty": "Before registering, only valid-answer counts per year were read for these questions; no 2024 answer distribution was read. Development on 2016-2021 vs 2022 (develop-records.py) found no model change that helped reliably, so the records model is exactly the one registered in protocol-gunlaw.json. On 2022 it was within 1.71 points for CAPPUN and 5.41 for GUNLAW; this is disclosed because it shows a single question can land either side of the bar.",
 "questions": {
  "cappun": "Do you favor or oppose the death penalty for persons convicted of murder? (1 favor, 2 oppose); outcome: share favor",
  "grass": "Do you think the use of marijuana should be made legal or not? (1 legal, 2 not legal); outcome: share legal",
  "natenvir": "Are we spending too much, too little, or about the right amount on improving and protecting the environment? (1 too little, 2 about right, 3 too much); outcome: share too little",
  "natheal": "... on improving and protecting the nation's health? outcome: share too little",
  "natrace": "... on improving the conditions of Blacks? outcome: share too little"
 },
 "population": "for each question, every 2024 respondent with a valid answer and valid facts, described exactly as in protocol-gunlaw.json",
 "methods": "carry_forward, regression (logistic on the facts, fitted on 2016-2022), voice (gemini-3.8-flash, the prompts of protocol-gunlaw.json with each question's wording and options; low thinking; default temperature) and anchored (regression log-odds plus b x voice, b fitted on 500 voiced 2022 respondents per question), all exactly as in protocol-gunlaw.json",
 "budget": "voice spend for the whole battery capped at USD 12",
 "outcomes": {
  "primary": "per question and method: six-cell MAE (party group x age group 18-44 / 45+) in percentage points with 2,000-resample bootstrap intervals, and the protocol-gunlaw.json match rule (MAE <= 3.0 AND each party group's age gap inside its 95% interval)",
  "battery": "per method, the number of the five questions on which it matches the measured split, and its mean MAE over the five",
  "decision": "the claim 'lives fitted to records reproduce measured population splits' is supported if the regression matches on at least 4 of 5; the voice is credited as a stand-in for a population only if it matches on at least 4 of 5. Reported whatever happens."
 }
}