{
 "name": "Siblings on both sides, from partnerships that follow the record, in three survey years never read",
 "registered": "2026-10-10",
 "why": "protocol-kin2.json (both sides) missed, 0.49 against the text model's 0.41, and stays a miss; protocol-kin3.json met the mother's side alone. kin2.ts let a woman give birth at the same rate whatever her partnership and ended partnerships at 5.3% a year in every year from 1950, so the fathers of people born in the 1950s and 1960s had children by far too many women: about 0.4 siblings were added through the father in every cohort. In the record, 4-5% of births in the 1950s were outside marriage and about one marriage in a hundred ended each year. People also report far fewer half-siblings through the father than through the mother (NLSY: 6% and 9% against one in five and one in four with any half-sibling by 18; Amorim and Tach 2019).",
 "truth": "General Social Survey, `sibs` (0-60) among respondents born in the United States (`born` = 1), weighted by wtssps, years 2002, 2006 and 2014, in six cells each: women and men aged 25-34, 35-44, 45-59 (18 cells)",
 "engine": "experiment/kin/kin4.ts at this commit with weights4.json and params4.json, 80,000 lives in 1900, seed 20261010, read at mid-year of each survey year. Lives from 1900 on UN WPP 2024 fertility and life tables (before 1950, the 1950 age pattern scaled to each year's total fertility). Each year the share of births to women without a partner follows NCHS births outside marriage less the share to cohabiting couples; other births go to partnered women and the partner is the father; 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 a rate that follows each year's divorces per 1,000 married women, higher for younger women. Fitted by calibrate-kin4.ts with no sibling answers: two eras of parity weights to the Census CPS children ever born of women 40-44 (1976, 1980, 1985, 1990; 2000, 2010, 2016, 2022); the scale of endings, the rise with cohabitation (fitted to zero) and rho, the chance a person counts one of their father's children with another woman (0.231), to the half-siblings people report by 18 in NLSY79 and NLSY97 (any: 20% and 25.5%; father's side: 6% and 9%). Siblings = the mother's other children + rho x the father's children with other women + stepShare x step-siblings, stepShare 0 (chosen on development years: 0 gave 0.322, counting step-siblings at rho 0.330, all step-siblings 0.535)",
 "text model": "gemini-3.8-flash, 3 asks per year of the prompt in ask_kin4.py (the protocol-kin.json prompt with the year set), averaged; thinking low",
 "earlier engine": "kin2.ts as registered in protocol-kin2.json (commit 9a7b8b6) with weights.json, 200,000 lives in 1950; answers engine-kin2-2002.json, engine-kin2-2006.json, engine-kin2-2014.json, committed with this protocol",
 "development": "only survey years already read (2000, 2004, 2008, 2010, 2012, 2016, 2018, 2022, 2024; 54 cells): engine 0.320 (by year 0.215, 0.217, 0.248, 0.345, 0.276, 0.364, 0.490, 0.401, 0.321); the text model 0.456 and kin2.ts about 0.46 on the same years (earlier results). No sibling answer from 2002, 2006 or 2014 has been read.",
 "primary": "mean absolute error over the 18 cells: engine lower than the text model and lower than kin2.ts",
 "secondary": "each year separately; the mother's side alone (`maternal`)",
 "order": "engine answers (engine-kin4-<year>.json) and text-model answers (text-kin4.json) are committed before score_kin4.py reads any survey answer",
 "publish": "results-kin4.json whatever the outcome"
}