{
 "name": "An agent that remembers what a person said, against the same agent with only a persona, on what that person said four years later",
 "registered": "2026-10-06",
 "why": "The life-record test (protocol-record.json) found no difference between a persona and a persona plus life events, and stays a miss. Work on individual simulation finds that a record helps when it carries what a person said about related things, not only what happened to them. This test gives the agent that memory, never the answers it is scored on, on three panels never used before.",
 "data": "GSS panels starting 2006, 2008 and 2010 (gss.norc.org Stata files, release 6), respondents interviewed at wave 1 and re-interviewed at wave 3 (panstat_3 = 1), each panel split in half by numpy default_rng(20261007) in prep_panel3w.py; only the test halves are scored (636 + 643 + 652 people). The dev halves are left for later work.",
 "arms": {
  "persona": "age, sex, race, education, region, marital status, children, work, household size, family income, party and views at wave 1",
  "memory": "the persona; every change by wave 3 in partnership, children, work, household, region and income; and the person's wave-1 answers to satjob, class, getahead, trust, helpful, fair, life and finrela. No arm contains any answer, at any wave, to the four scored questions."
 },
 "targets": "wave-3 answers to happy, satfin, finalter, health",
 "model": "gemini-3.8-flash, thinking low, 20 people per call, ask_panel3w.py (the prompt of ask_panel.py with the year of wave 3); probabilities for each answer option",
 "score": "multi-category Brier score per person and question, probabilities renormalised; mean difference persona - memory, weighting each person by their number of scored answers; bootstrap 95% interval resampling people (2000 draws, seed 7); score_panel3w.py",
 "primary": "pooled over the three test halves and all four questions: the memory arm has the lower Brier score and the 95% interval of the difference lies above zero",
 "secondary": "each question, each panel, and the relative improvement",
 "development": "none on these panels beyond a 40-person check of the pipeline (dev half of the 2006 panel, answers not scored or read). The earlier 2016-2020 test: persona 0.6122, record 0.6096 on the financial pair, interval spanning zero.",
 "spend_cap_usd": 2.0,
 "order": "persona and memory answers are written to persona-<yy>.jsonl and memory-<yy>.jsonl and committed before score_panel3w.py runs",
 "publish": "results-memory.txt whatever the outcome"
}