MatrAIx: Simulating the World with 8.3 Billion Persona Agents

System, dataset, and validation paper introducing an end-to-end infrastructure for testing AI systems and digital products with heterogeneous simulated users. Its three pieces are Persona 8B, a common categorical representation of people; the MatrAIx Playground, which runs persona-conditioned agents in Survey, AI Chatbot, Web, and native App environments; and MatrAIx Applications, a library of versioned evaluation tasks. The important conceptual inversion is that the product is the system under test. Personas are sampled as evaluators whose requests, interactions, and judgments vary, while the target product and task remain controlled.

Population construction

Persona 8B claims approximately 8.3 billion records over 1,290 categorical dimensions spanning background, psychology, capability, behavior, and lifestyle. The number names a generated internal population, not the public artifact. The released coreset contains 999,847 records: 599,847 “human-grounded” records and 400,000 synthetic ones.

Synthetic records are sampled in topological order from a directed acyclic graph. The visualized graph has 1,308 nodes and 6,999 edges: 1,290 emitted attributes plus 18 latent roots. Source-informed conditional factors shift local distributions, while separate compatibility masks forbid logically invalid combinations. Edges are added when a source reports a conditional; source and schema reviews inspect them, and an LLM-assisted coverage pass proposes omissions for human acceptance or rejection. This is more auditable than independently sampling 1,290 marginals, but it remains a designed model of dependence rather than an estimated joint distribution. The released mix is calibrated only toward age bracket, region, gender identity, and urbanicity marginals, not representative joint coverage over all dimensions.

Human-grounded records map six sources into the same partial schema: Wikipedia biographies (323,438), Amazon review histories (97,915), the Stack Overflow Developer Survey (113,120), the General Social Survey (63,532), PRISM Alignment (1,487), and 355 consented volunteer responses. Coded surveys use deterministic crosswalks; Wikipedia, Amazon, and PRISM free text use constrained LLM extraction. Unsupported attributes remain null. Direct identifiers are removed, and the volunteer instrument collects no name, contact detail, or account identifier. “Human-grounded” therefore identifies provenance, not truth, completeness, representativeness, or consent to every downstream use.

Evaluation as an executable contract

Each study binds a population query and realized cohort to a versioned task, persona-agent model, environment, seed, outcome measures, and verifier. Survey trials return structured answers; Chatbot trials preserve conversations and service calls; Web and App trials preserve actions, screenshots or state, and final submissions. Programmatic verifiers check observable outcomes, while human or LLM judges handle interpretive properties. Reports aggregate by task, cohort, and subgroup but retain links to trial evidence. This separation is the paper’s strongest systems contribution: product outcome, persona fidelity, and execution failure do not collapse into one score.

The application library contains 1,010 specifications across more than 25 domains — 621 Survey, 371 Chatbot, 12 Web, and 6 App — but availability is not execution. The paper reports 18,189 trials across only eight representative tasks, two per environment, using GPT 5.5, Claude Opus 4.8, and Claude Haiku 4.5 as persona models. Survey, Chatbot, and Web conditions generally use about 1,000 personas per model; the two expensive native-App tasks use only 24 and 20. The manuscript explicitly distinguishes available, implemented, executed, and empirically reported tasks, a valuable defense against reading catalog size as evidence breadth.

What the validation establishes

The controlled adherence study tests ten behavioral attributes in four environments, with five positive and five negative personas per cell. An Opus 4.8 LLM judge finds the declared behavior expressed or correctly suppressed in 366 of 400 trials (91.5%); 33 of 40 cells meet the paper’s strong-cell threshold. The figure makes the environment gradient visible: Survey 96%, Chatbot 92%, Web 95%, and App 83%, with App strong on only 6 of 10 attributes. The appendix’s same-protocol GPT-5.6-sol run reaches 317 of 400 (79.2%) under the same judge, with an adapter exception in Chat. Coding-style attributes transfer better than soft traits, while requests to become verbose, impolite, or otherwise less polished often lose to the acting model’s prior.

Extraction quality is evaluated separately. GPT 5.5 and Claude Opus 4.8 judge 1,000 extracted personas; six humans each rate a source-matched subset of 100. The human mean is 4.135/5, with 97.2% of human-rater comparisons within one point. Claude and GPT are within one point of the six-rater human mean in 93.8% and 79.2% of comparisons. This supports the extraction pipeline on that subset; it does not validate behavioral resemblance to people, population representativeness, or product predictions.

Application results themselves show why that distinction matters. One OpenBB task recovers the same trust-group ordering across all three models, but the meal-planning figure’s largest apparent subgroup differences do not survive multiple-comparison correction. Elsewhere, an identical cohort’s paid-plan share spans 23.2% to 93.9% across persona models, and median pairwise Cohen’s κ over 88 joinable fields is approximately zero. Persona-model choice is a first-order experimental factor, not an interchangeable implementation detail.

Assessment

• Durable: product-as-system-under-test framing; explicit manifests and reproducible cohorts; typed task, telemetry, and verifier contracts; and the insistence that catalog, execution, persona fidelity, and product outcome are different quantities.
• Era-bound: the 8.3-billion headline, frontier-model roster, task catalog, and cost-driven environment balance. The public, inspectable artifact is the approximately one-million-record coreset.
• Weaknesses: the top-line adherence rate is produced by one LLM judge and measures compliance with declared style, not human likeness; the four calibration marginals cannot establish joint population fidelity; most task specifications were not executed; App conclusions rest on tiny cohorts; and model providers evaluated in the paper also supported the work, although the conflict is disclosed.
• Responsible-use boundary: the paper appropriately calls the personas simulation instruments rather than people, warns against impersonation, targeting, and consequential population claims, and requires human studies before real-world decisions. It also identifies an untested shared-backbone bias: a persona model may prefer output from the same model family. Direct comparison with longitudinal human behavior — disclosure, correction, refusal, and abandonment — remains future work.

In liu2026‘s deployment-facing language, MatrAIx supplies a way to close evaluation on user-shaped product interactions rather than on an agent benchmark. It is not itself a self-evolving system, and its simulated feedback should remain hypothesis-generating until matched against real users.