{
 "name": "Lived, not learned: brothers and sisters on the mother's side, in four survey years never read",
 "registered": "2026-10-07",
 "why": "Both earlier sibling tests missed and stay misses (protocol-kin.json 1.10 against 0.38; protocol-kin2.json 0.49 against 0.41). In the second, the engine's count of the mother's other children was closer than the text model (0.355 against 0.410), while the father's children with other mothers were too many. Development since (kin3-dev.ts, kin-analytic.ts) found the father's side still too high once rates before 1950, two eras of parity weights and divorce-following separations were added. This test asks only for what the engine gets right so far: the mother's other children, from kin2.ts unchanged.",
 "truth": "General Social Survey `sibs` among respondents born in the United States (`born` = 1), weighted by wtssps, in 2004, 2008, 2012 and 2016 separately, six cells each: women and men aged 25-34, 35-44, 45-59 (24 cells). The question counts full, half, step and adopted siblings; the engine counts the mother's other children (full and maternal half siblings), so the comparison favours neither in a known direction",
 "engine": "experiment/kin/kin2.ts at this commit (as registered in protocol-kin2.json, 9a7b8b6) with weights.json, 200,000 lives in 1950, read at mid-year of each survey year; the primary uses `sibsMaternal`",
 "text model": "gemini-3.8-flash, 3 asks per year of the prompt in ask_kin3.py (the protocol-kin.json prompt with the year set), averaged; thinking low",
 "development": "only years already read: 2000, 2010 and 2018 pooled (protocol-kin2.json results): mother's side 0.355, engine with the father's side 0.491, text model 0.410; 2022: mother's side 0.424, text model 0.47. No sibling answer from 2004, 2008, 2012 or 2016 has been read.",
 "primary": "mean absolute error over the 24 cells: the engine's mother's-side count lower than the text model",
 "secondary": "each year; the engine with the father's side; children ever had (`childs`), engine against the text model",
 "order": "engine and text-model answers are written to engine-kin3-<year>.json and text-kin3.json and committed before score_kin3.py reads any survey answer",
 "publish": "results-kin3.json whatever the outcome"
}