Blog · Research · Oct 21, 2026 · 5 min
A population lived month by month in the United States from 1950 matched how many children Americans have had and how many brothers and sisters they report, closer than a frontier model answering from what it has read. The father's side, first counted through partnerships, missed, and became the next test.
How many children have you had? How many brothers and sisters? These are two of the simplest questions a survey can ask, and two of the hardest for a model to answer for a whole population. The answer to each depends on decades of births, deaths and partnerships that happened to other people: your mother, your father, the people they were with before and after.
Living Worlds asks whether a population that has lived those decades, month by month, can give the answers real people give. A frontier language model can answer too, from what it has read. The tests below put the two side by side on survey answers neither had seen.
The population lived in the United States from 1950, on each year's official fertility and mortality. No General Social Survey answer was used to build it. In the first test its only fitted weights came from Census counts of children among women aged 40 to 44. It was then scored on the General Social Survey, the long-running national survey of American adults run by NORC. The 2024 round was collected from April to December 2024, face to face and on the web.
Every test was registered before any answer from the scored years was read, and scored once. "Never read" in a title means the sibling answers from those years had not been opened by anyone working on the engine when the rule was written. That is what turns a match into evidence: the population could not have been tuned towards answers nobody had seen.
Lived, not learned: children ever had scored the population on six cells of the 2024 survey: US-born women and men aged 25 to 59. It matched how many children they had had within 0.17 on average. A frontier model answering from what it has read missed by 0.36. Met.
The test also counted siblings, and there it missed: 1.10 against 0.38. That miss became the next test rather than a footnote.
Births are moving, which is why this needs to hold on fresh years. Final US birth data for 2024 showed 3,628,934 births and a 1% fall in the fertility rate, reversing the 1% rise in the provisional estimate once updated Census population figures came in.
Siblings come through both parents, and the father's side depends on who partners with whom, and for how long. Siblings, from partnerships that make fathers was scored on three survey years, 2000, 2010 and 2018, never read for siblings.
On children ever had, the lived population came closer than a text model (0.12 against 0.15). On siblings it came closer in 2010 and 2018, but pooled it missed by 0.49 against the model's 0.41, overcounting people then aged 35 to 44 in 2000. Not met. Counting the mother's side alone missed by 0.36, which pointed to the next step.
Siblings on the mother's side, in four years never read counted each mother's other children and compared them with the brothers and sisters Americans report, full, half, step and adopted. The engine was unchanged, and no sibling answer from 2004, 2008, 2012 or 2016 had been read.
Across 24 cells it came within 0.28 on average, against 0.51 for a frontier model, and was closer in every year: 0.21, 0.25, 0.28 and 0.38 against 0.43, 0.51, 0.53 and 0.55. Children ever had came within 0.10 against 0.16. Met. Adding the father's side gave 0.47, still ahead of the model, and that side became the next test.
Siblings on both sides, in three years never read counted brothers and sisters through both the mother and the father, across Americans aged 25 to 59, on 18 cells from 2002, 2006 and 2014. It came within 0.31 on average, against 0.45 for a frontier model and 0.55 for the earlier engine, and was closer than the model in 2002 and 2006. Met.
It was built without a single sibling answer. On a check it never saw, 14% of its fathers had children with more than one woman in 2014, against 14.6% in the Census survey that year.
A family is a record of lives that met. The number of siblings a person has depends on their mother's partnerships, their father's, the births in each and who survived. A model that has read about families can describe them; a population that has lived them has to produce the count. Across four tests on years nobody had read, the lived population came closer to what Americans report than a frontier model answering from text, on children in every test and on siblings once the counting was set right.
The order matters as much as the results. The sibling miss in the first test, and the father's-side miss in the second, were published as recorded and each set the next registered test. The final engine meets on both sides without having seen a sibling answer.
These are averages across age-and-sex cells of one country's survey; they say nothing yet about individual families, or about countries other than the United States. In the both-sides test the lived population was closer than the model in two of the three years, not all three. The share of fathers with children by more than one woman matched in one year, 2014. And the father's side was first counted through partnerships and missed, so the tests that met reflect the engine as it now stands, after that miss.
Living Worlds' idea is lived, not learned: most of what makes us human was never written down, it was lived. Its populations are born, partner, have children and grow old on each year's official rates, and every life keeps its record. Counting children and siblings from those records, and checking them once against answers nobody had read, is how the world shows the records hold.
Blog
Language models answering as real people, scored once on survey answers nobody had read: a frontier model met 0 of 6, prompted voices overstated births threefold, and carrying the last shares forward beat a small model. Lives fitted to records came closer, a model's estimate counted as a prior meant fewer people to ask, and subgroup splits landed inside the survey's own noise in the United States and in 30 of 31 African countries.
Living Worlds' life expectancy forecasts, frozen twenty years back and scored once on years no run had seen: met for women and for men on the latest national figures, not met against 2024 national figures or across 155 countries, with honest intervals for women and a stacked UN model that did not help.
Census-built populations, agents grounded in interviews and language models speaking for whole groups have all advanced in 2025-26. The evidence now shows clearly which of them match real people, and the dividing line is what each simulated person is built from.