{
 "name": "Lived, not learned: brothers and sisters, from lives whose partnerships make fathers, in three survey years never read",
 "registered": "2026-10-06",
 "why": "The siblings secondary of protocol-kin.json missed (engine 1.10 against the text model's 0.38) and stays a miss. Its text said the engine counts the mother's other children, but the frozen code also added the father's children with other mothers, drawn from any man of the right age, so fathers had children by too many women (46% of fathers 40-49 with two or more mothers, against 19.5% in the National Survey of Family Growth). This test fixes the rule and calibrates it to that survey, never to sibling answers, and scores it on years whose sibling answers have not been read.",
 "truth": "General Social Survey, `sibs` among respondents born in the United States (`born` = 1), weighted by wtssps, years 2000, 2010 and 2018 separately, in six cells each: women and men aged 25-34, 35-44, 45-59 (18 cells)",
 "engine": "experiment/kin/kin2.ts at this commit with weights.json: as kin.ts (United States lived from January 1950 on UN WPP 2024 fertility and life tables, parity weights from the Census CPS) plus: a birth to a woman without a partner is fathered by an unpartnered man 0-6 years older and the two become partners; partnerships end at 0.0045 a month. That rate was chosen only to match the 2017-2019 NSFG share of fathers 40-49 with children by two or more mothers (engine 0.19 in 2018, survey 0.195; 30-39: 0.12 against 0.135). Siblings = the mother's other children plus the father's children with other mothers. 200,000 lives in 1950, read at mid-year of each survey year.",
 "text model": "gemini-3.8-flash, 3 asks per year of the prompt in ask_kin2.py (the protocol-kin.json prompt with the year set), averaged; thinking low",
 "development": "on the 2022 survey only (already read): engine 0.323 (mother's side only 0.424; the registered rule 1.1), text model 0.47 (protocol-kin.json). No sibling answer from 2000, 2010 or 2018 has been read.",
 "primary": "mean absolute error over the 18 cells: engine lower than the text model",
 "secondary": "each year separately; the mother's side only; children ever had (`childs`) for the same cells, engine against the text model",
 "order": "engine and text-model answers are written to engine-kin2-<year>.json and text-kin2.json and committed before score_kin2.py reads any survey answer",
 "publish": "results-kin2.json whatever the outcome"
}