Anamnesis: An Open-Source Platform for Large-Scale Backstory-Conditioned Survey Simulation

Society, culture, and collective behavior2026arXivApproved editorial review

Authors: Song-Ze Yu, Joseph Suh, Serina Chang, David M. Chan

Keywords: Survey simulation, Virtual personas, Narrative backstories, Probabilistic demographic matching, Human alignment, Multimodal surveys, Research platform

Source: Open primary source (opens in a new tab)

4
Authors
12
Findings
21
Limitations
6
Evidence

Editorial summary

English

Anamnesis is an arXiv v1 system paper about an open-source web platform for building surveys and running them with virtual participants conditioned on synthetic life narratives. It does not create replicas of real people or show that an LLM has a human personality. The application combines single-choice, multi-select, open-response, ranking, image and audio questions; Supabase storage; a RabbitMQ queue; and model-calling workers. Anthology mode prepends a full narrative and accumulates the same virtual persona's earlier answers. The zero-shot control uses a short demographic description. When output is noncompliant, another LLM can convert it to an option letter, adding a second model dependency that can affect results.

The reported database contains 34,907 synthetic English backstories. Their demographic attributes are neither observed nor self-reported: another LLM estimates probability distributions. Top-K sums selected category probabilities within each dimension, multiplies across dimensions and takes the highest scores; balanced mode creates quotas for demographic combinations and runs Hungarian assignment over the 50 highest-scoring candidates per cell. This controls selection relative to the model's own predictions, but does not establish correct identities, calibrated probabilities or census representativeness. Multiplication assumes independence and can distort intersectional combinations. In logprobs mode, if the code recognizes no option token, it replaces the failure with a uniform distribution that can look like valid uncertainty. Demographically controllable therefore does not mean statistically representative population.

The first case reproduces part of Anthology using LLaMA-3.1-8B, 20 selected questions from three American Trends Panel waves, and thousands of virtual respondents. It uses Wasserstein distance between response distributions and Frobenius norm between correlation matrices; lower is better. Max-weight Anamnesis beats BIO and QA in all six table cells: Wave 34 is 0.160/0.837 versus the best baseline's 0.235/1.481; Wave 92 is 0.251/1.603 versus 0.346/1.719; and Wave 99 is 0.148/1.026 versus 0.180/1.229. Greedy improves five of six comparisons with the best baseline, but its 1.352 Wave 99 Frobenius value is worse than BIO's 1.229. Every Anamnesis configuration remains clearly behind the Human row (0.057/0.418; 0.091/0.411; 0.081/0.327). The paper provides no intervals, run-to-run variation, multiple-model evaluation, subgroup/item analysis, complete selection rule or contamination test.

The second case uses 49 cartoon-caption contests, Gemini 2.5 Flash at temperature 1.0 and 20 choices per contest. Anthology identifies the human winner in 29/49 contests (59.2%; Wilson interval 45.2–71.8) and zero-shot in 25/49 (51.0%; 37.5–64.4): four additional contests with overlapping intervals. The paper says it uses exact McNemar testing but omits its p-value and discordant-pair table. Vote share for the human winner rises from 52.0% to 59.8%, a 7.8-point gain (paired bootstrap interval 3.2–12.8; exact sign-flip p=0.0024). This supports greater aggregate agreement with that benchmark, not general humor understanding or individual fidelity.

The application code is genuinely public and AGPL, but the evidence is not reproducible end to end. Main does not contain the exact enriched 34,907-person snapshot, ATP/caption inputs, run configurations, seeds, model outputs or scripts that regenerate the tables. Public datasets contain 11,364 Anthology and 41,053 Alterity backstories as text only, without the demographic probabilities used. The frontend builds, but the audit found 27 lint errors, 60/128 failing frontend tests in the checked environment, a broken pytest collection and two further failures among the remaining 256 tests; npm audit found three high-severity production vulnerabilities. More seriously, SECURITY DEFINER SQL functions that read decrypted API keys or mutate work do not revoke EXECUTE from PUBLIC or check ownership, and media functions verify authentication but not object ownership. These issues were not exercised against production and no data was accessed. The public site is clear, responsive and uses landmarks and a skip link, but its animated heading exposes partial phrases, mobile navigation hides most links without a menu, and registration offers no privacy or terms links. Overall, Anamnesis is a useful open prototype with a real interface and relative gains in two bounded cases; it does not validate representative synthetic participants or a ready replacement for human studies.

Español

Anamnesis es un artículo de sistema, todavía en arXiv v1, sobre una plataforma web de código abierto para construir encuestas y ejecutarlas con participantes virtuales condicionados por historias de vida sintéticas. No crea réplicas de personas reales ni demuestra que un LLM posea una personalidad humana. La aplicación integra preguntas de opción única, selección múltiple, respuesta abierta, ranking, imagen y audio; un backend Supabase; una cola RabbitMQ; y workers que consultan a un modelo. El modo Anthology antepone una narración completa y acumula las respuestas anteriores de la misma persona virtual. El control zero-shot usa una descripción demográfica breve. Si la salida no cumple el formato, otro LLM puede convertirla en una letra, introduciendo una segunda dependencia de modelo que puede afectar el resultado.

La base descrita contiene 34.907 historias sintéticas en inglés. Sus atributos demográficos no son observados ni autodeclarados: otro LLM estima distribuciones de probabilidad. Top-K suma las categorías elegidas por dimensión, multiplica entre dimensiones y selecciona las puntuaciones más altas; el modo balanceado crea cuotas por combinaciones y aplica asignación húngara sobre los 50 mejores candidatos de cada celda. Esto controla la muestra respecto a predicciones del propio modelo, pero no garantiza identidades correctas, probabilidades calibradas ni representatividad censal. La multiplicación supone independencia y puede distorsionar combinaciones interseccionales. En modo logprobs, si el código no reconoce ningún token de opción, sustituye el fallo por una distribución uniforme que puede parecer incertidumbre válida. Por tanto, demográficamente controlable no significa población estadísticamente representativa.

El primer caso reproduce parte de Anthology con LLaMA-3.1-8B, 20 preguntas seleccionadas de tres olas del American Trends Panel y miles de personas virtuales. Mide distancia de Wasserstein entre distribuciones y norma de Frobenius entre matrices de correlación; menor es mejor. La variante de peso máximo mejora a BIO y QA en las seis celdas: en la ola 34 obtiene 0,160/0,837 frente al mejor baseline 0,235/1,481; en la 92, 0,251/1,603 frente a 0,346/1,719; y en la 99, 0,148/1,026 frente a 0,180/1,229. Greedy mejora cinco de seis comparaciones contra el mejor baseline, pero su Frobenius 1,352 en la ola 99 es peor que 1,229 de BIO. Todas las variantes siguen claramente por detrás de la fila humana (0,057/0,418; 0,091/0,411; 0,081/0,327). No se informan intervalos, variación entre ejecuciones, múltiples modelos, análisis por subgrupo o pregunta, criterio completo de selección ni prueba de contaminación.

El segundo caso usa 49 concursos de viñetas, Gemini 2.5 Flash a temperatura 1,0 y 20 elecciones por concurso. Anthology acierta el ganador humano en 29/49 casos (59,2%; IC Wilson 45,2–71,8) y zero-shot en 25/49 (51,0%; 37,5–64,4): cuatro concursos más, con intervalos solapados. El artículo dice aplicar McNemar exacto pero no publica su p-valor ni la tabla de pares discordantes. La cuota de voto al ganador humano sube de 52,0% a 59,8%, +7,8 puntos (IC bootstrap 3,2–12,8; prueba de signos exacta p=0,0024). Esto muestra mayor acuerdo agregado con ese benchmark, no comprensión general del humor ni fidelidad individual.

El código de aplicación es realmente público y AGPL, pero la evidencia no es reproducible de extremo a extremo. La rama principal no incluye la instantánea enriquecida de 34.907 personas, datos ATP/viñetas, configuraciones, semillas, salidas ni scripts que recompongan las tablas. Los datasets públicos contienen 11.364 historias Anthology y 41.053 Alterity, solo texto, no las probabilidades usadas. El frontend compila, pero la auditoría encontró 27 errores de lint, 60/128 pruebas frontend fallidas en el entorno comprobado, una colección pytest rota y dos fallos adicionales entre las 256 pruebas restantes; npm audit detectó tres vulnerabilidades altas de producción. Más grave: funciones SQL SECURITY DEFINER que leen claves API descifradas o cambian tareas no revocan EXECUTE a PUBLIC ni comprueban propiedad, y las funciones de medios verifican autenticación pero no pertenencia del objeto. No se probaron estos fallos contra producción ni se accedió a datos. La web es clara, responsive y usa landmarks y enlace de salto, pero el titular animado expone frases incompletas, la navegación móvil oculta casi todos los enlaces sin menú y el registro carece de política de privacidad o términos. En conjunto, Anamnesis es un prototipo abierto útil con una interfaz real y mejoras relativas en dos casos acotados; no valida participantes sintéticos representativos ni una alternativa lista para sustituir estudios humanos.

Research question

Can an open-source web platform operationalize narrative-history conditioning, probabilistic demographic sampling, and distributed survey execution, and reproduce two aggregate patterns of human data better than demographic prompts?

Method

System article with two cases. ATP runs LLaMA-3.1-8B on 20 selected questions from waves 34, 92, and 99 and compares BIO, QA, Anamnesis with maximum weight, Anamnesis greedy, and a human reference using Wasserstein and Frobenius. The multimodal case uses 49 vignette contests, Gemini 2.5 Flash, temperature 1.0, and 20 choices per contest; it compares majority accuracy with Wilson intervals and vote share with paired bootstrap and a sign test.

Sample: Two bounded evaluations: 20 questions from three ATP waves run with thousands of virtual persons and compared with human distributions; and 49 vignette contests with 20 synthetic choices per condition. The operational base is described as 34,907 generated stories, not as observed human persons.

Findings

  • Anamnesis implements construction, selection, distributed execution, tracking, and analysis of textual and multimodal surveys.
  • The stories are synthetic and their demographics are probabilities inferred by an LLM, not observed labels.
  • The maximum-weight ATP variant improves over BIO and QA in all six combinations of wave and metric.
  • Greedy improves five of six comparisons against the best baseline; in Frobenius for wave 99 it falls behind BIO.
  • All Anamnesis metrics remain behind the human reference.
  • In vignettes, Anthology correctly predicts 29/49 winners and zero-shot 25/49; the Wilson intervals overlap.
  • The mean vote share of the human winner rises 7.8 points, from 52.0% to 59.8%, with CI 3.2-12.8 and p=0.0024.
  • The application code is public under AGPL and allows inspecting frontend, database, queue, and workers.
  • The data and scripts needed to reproduce the two evaluations are not on the main branch.
  • The public story datasets do not contain the demographic probabilities of the 34,907-case base.
  • The website is visually clear, but communicates representativeness more strongly than its empirical limits warrant.
  • The static audit identifies serious authorization flaws that must be corrected before handling user data or keys.

Limitations

  • It is a nine-page arXiv v1 preprint, not a peer-reviewed archival publication.
  • The simulation is limited by the diversity and biases of the LLM-generated story set.
  • Inferred demographics are not calibrated or validated against observed labels.
  • Multiplying probabilities by dimension assumes independence and may fail for intersectional identities.
  • The logprobs fallback converts absence of valid tokens into a uniform distribution.
  • ATP uses only 20 selected questions and does not fully document its selection rule.
  • ATP matching and metric details are delegated to the earlier Anthology article.
  • No intervals, seeds, variation across runs, or multiple models are shown for ATP.
  • There is no analysis by question, subgroup, stereotype, calibration, or fairness.
  • The vignette evaluation contains only 49 items and one multimodal model.
  • The McNemar p-value and the table of discordant pairs are not published.
  • The contest winner is a collective preference, not a universal truth about humor.
  • The stories and simulations are exclusively in English.
  • Virtual persons are static and multimodal models may omit human nuances.
  • The enriched snapshot of 34,907 stories is not published, nor is its exact selection.
  • Inputs, outputs, configurations, seeds, or analytical scripts for the cases are not published.
  • The frontend and worker suites are not green and the lint contains 27 errors.
  • Production dependencies present three high-severity npm findings.
  • SECURITY DEFINER and media functions lack sufficient execution or ownership controls.
  • The public site offers no privacy policy, terms, retention, or account deletion.
  • Mobile navigation hides links without a menu and the animated H1 produces partial text.

What the study does not establish

  • That the 34,907 synthetic stories are representative persons of a census
  • That an inferred demographic probability is an observed or self-declared identity
  • That Hungarian assignment guarantees real demographic representativeness
  • That the sample is statistically representative of a target population
  • That every Anamnesis configuration outperforms every baseline on every metric
  • That Anamnesis reaches the human reference on any ATP wave
  • That improving aggregate metrics implies fidelity to specific individuals
  • That 49 contests demonstrate general understanding of human humor
  • That synthetic participants can replace human subjects in research
  • That the results generalize to other models, languages, cultures, or sensitive domains
  • That the evaluations can be reproduced end-to-end from the main branch
  • That the public datasets contain the exact enriched base
  • That the code has green gates for tests, lint, dependencies, and authorization
  • That the public workflow has a documented privacy or ethics framework
  • That publishing the application code alone makes the evidence reproducible

Traceability

Scope: Full text

Version: arXiv:2607.10628v1; 9-page preprint; repository commit 31a0d7df44966e8618b82ff0492c1c397832cdbc; live platform and Hugging Face source datasets checked 2026-07-16

Consulted source: https://arxiv.org/abs/2607.10628v1

Review: Codex 9-page full-text visual, arXiv metadata, platform architecture, GitHub code/test/CI/dependency/security, live desktop/mobile UI and Hugging Face dataset audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • LLaMA-3.1-8B
  • Gemini 2.5 Flash
  • Modelos configurables mediante OpenRouter o vLLM
  • LLM parser de respaldo configurable

Instruments and metrics

  • Historias narrativas sintéticas Anthology/Alterity
  • Condicionamiento secuencial con acumulación de contexto
  • Prompt demográfico zero-shot
  • Ranking top-K por probabilidades inferidas
  • Asignación húngara con cuotas demográficas
  • Distancia de Wasserstein
  • Norma de Frobenius entre matrices de correlación
  • Intervalo de Wilson
  • Prueba exacta de McNemar
  • Bootstrap pareado
  • Prueba exacta de signos

Data used

  • Base declarada: 34.907 historias sintéticas en inglés con distribuciones demográficas inferidas; instantánea exacta no publicada
  • American Trends Panel olas 34 (n=2.537), 92 (n=10.916) y 99 (n=10.260); 20 preguntas seleccionadas
  • New Yorker Caption Contest Benchmarks: 49 concursos con imagen
  • Hugging Face anthology_backstory: 11.364 filas; alterity_backstory: 41.053 filas

Evidence and location

  • Architecture, algorithms, cases, results, appendices, and limitations: arXiv:2607.10628v1 PDF, 9 pages; every page rendered and visually inspected
  • Date, version, authors, and official categories: Official arXiv record and export API for 2607.10628v1
  • Implementation, tests, CI, dependencies, SQL, security, and absent artifacts: DavidMChan/Anamnesis at 31a0d7df44966e8618b82ff0492c1c397832cdbc
  • UI/UX, responsive navigation, alt text, and public transparency: simulate.group live site and production bundle checked 2026-07-16
  • Size, columns, language, and license of source datasets: Hugging Face Dataset Viewer metadata for anthology_backstory and alterity_backstory checked 2026-07-16
  • Consolidated audit of validity, evaluation, code, security, and UX: reports/verification/article-275-anamnesis-platform-demographic-validity-evaluation-code-security-and-ux-audit.json