{
 "name": "A voice tuned on women's life records, against age alone and on the level of births, in a survey never opened",
 "registered": "2026-10-10",
 "why": "Two earlier births results stay as they are: on GSS panel women the small voice ranked like a regression (AUC 0.646) but was not shown to beat age alone; asked directly, voiced lives put births three times too high (5.24 against 1.62 per woman). This test tunes a voice on NSFG life records and asks both things of it on NSFG 2022-2023, never opened here.",
 "truth": "NSFG 2022-2023 female respondent and pregnancy public-use files (CDC): women 18-44 at the start of a three-year window ending the year before interview (window start = interview year - 3), outcome a live birth (OUTCOME 1) ending in the window; facts at the window start: age, children born before it, years since the last birth, education on one scale (births_prep.py), race and Hispanic origin (HISPRACE2), religion (RELIGION), metropolitan residence (METRO). Unweighted.",
 "voice": "experiment/births/births_voice.py at this commit: Qwen/Qwen3-4B-Base revision 906bfd4b4dc7f14ee4320094d8b41684abff8539 in 4 bits with a low-rank adapter (r 8, alpha 32, q and v projections), 2 epochs, learning rate 1e-4, micro-batch 8, accumulation 8, seed 0, trained on all NSFG 2017-2019 women (w2017e.csv, 4,757, from births_prep.py) with cross-entropy on the observed outcome; probability from the A/B option scores. Run on the desktop through run.ps1 with GPU memory capped at fraction 0.42. It reads a copy of the 2022-2023 facts with the outcome set to 0.",
 "level": "The voice's mean probability is carried to the 2022-2023 window years by the national general fertility rate (NCHS final data, births per 1,000 women 15-44), averaged by interview year: 2022-2023 interviews, windows 2019-2021 and 2020-2022 (2019 58.3, 2020 55.7, 2021 56.3, 2022 56.0; NCHS data briefs 442 and 477), mean 56.38; 2017-2019 interviews, windows 2014-2016, 2015-2017, 2016-2018 (2014 62.9, 2015 62.5, 2016 62.0, 2017 60.3, 2018 59.1), mean 61.51; ratio 0.917, fixed in births_score.py.",
 "baselines": "age alone: logistic regression on age and its square; the regression on all facts (births_dev.py design); both fitted on NSFG 2017-2019 and scored on 2022-2023",
 "development": "NSFG 2017-2019 (already read) and 2011-2013 (new, development only). Halves of 2017-2019 (2,379 scored): voice AUC 0.735, age 0.654, regression 0.723; voice minus age +0.058 to +0.101; voice mean 0.200 against observed 0.202 (0.186-0.219). Across waves (voice tuned on 2011-2013, 4,380 women; scored on all 2017-2019, 4,757): voice AUC 0.734, age 0.653, regression 0.719; voice minus age +0.069 to +0.096; observed 0.195 (0.184-0.207); voice mean 0.204, carried by the same rule (ratio 0.951) 0.194. No 2022-2023 record has been opened.",
 "primary": "both: (1) the bootstrap 95% interval (2,000 resamples, seed 20261010) of AUC(voice) - AUC(age alone) lies above zero; (2) the voice's carried mean lies inside the bootstrap 95% interval of the observed birth share",
 "secondary": "voice against the regression on all facts; the voice's mean before carrying; the carried means of age alone and of the regression",
 "order": "births_prep.py builds the 2022-2023 table only after this commit; the voice's scores are committed as voice-2022.jsonl before births_score.py reads the outcome; scored once",
 "publish": "results-births.json whatever the outcome"
}