Simulating Belief: A Framework for AI Persona Generation

Introduction

For decades, persona generation has been the domain of marketers armed with demographic data. Age, location, income, education — we partition populations along these axes and call the resulting composites "personas." The approach is intuitive, scalable, and fundamentally limited. A 34-year-old software engineer in San Francisco making $180,000 tells you almost nothing about how she evaluates a pitch deck, what cognitive biases shape her decisions, or whether she trusts her gut over a spreadsheet. Demographics describe containers; they do not describe contents.

The problem becomes acute when we need personas to simulate judgment rather than merely represent a market segment. Investor personas, in particular, must produce coherent, specific evaluations of novel inputs — a pitch deck, a product demo, a go-to-market thesis — that hold up across repeated runs. Demographic vectors fail at this because they provide no structural constraint on the reasoning process. An LLM prompted with "you are a 45-year-old partner at a seed-stage fund" will generate plausible-sounding but internally inconsistent feedback. It may be risk-averse on one dimension and recklessly optimistic on another, or it may simply regress to the mean of its training distribution, producing the generic "founder-market fit" commentary that entrepreneurs have learned to ignore.

We've been exploring a model of belief, not identity, as the foundation for AI persona generation. The framework encodes decision-making dispositions as continuous values in a bounded vector space, then maps these vectors onto coherent investor archetypes — something we're exploring at Gatekeep. We cover the formal structure of the belief vector, early validation against real venture capital feedback, and broader implications for synthetic user research.

The Belief Vector Architecture

We define a belief vector as a point in [0,1]¹², where each dimension captures a distinct cognitive or behavioural disposition. Unlike demographic attributes, which are categorical and static, belief dimensions are continuous and composable. Two personas can agree on eight dimensions, diverge sharply on two, and produce recognisably different evaluations — while sharing identical demographic profiles. This property is critical: it means belief vectors can generate variation within a demographic cohort, which is where most of the interesting variance actually lies.

The twelve dimensions are:risk tolerance — willingness to accept variance in outcomes;time preference — discount rate applied to future returns;technical depth — the level of abstraction at which a persona reasons about technology;domain expertise — prior knowledge in a specific vertical;decision speed — latency preference between stimulus and judgment;social proof sensitivity — weight assigned to signals of other actors' behaviour;contrarian tendency — predisposition to invert consensus views;loss aversion — asymmetric weighting of downside versus upside;narrative susceptibility — responsiveness to storytelling versus quantitative evidence;information processing style — position on the intuitive-to-analytic spectrum;trust modality — whether trust is earned through competence, warmth, or social proximity; andvalue alignment — the moral and ideological frame through which decisions are filtered.

These dimensions are not merely descriptive labels. Each maps directly and measurably onto an investor archetype. A seed-stage generalist, for instance, scores high on risk tolerance (0.8–0.95), high on contrarian tendency (0.7–0.9), moderate on technical depth (0.4–0.6), and low on loss aversion (0.1–0.3). A late-stage institutional LP inverts nearly every dimension: low risk tolerance, high loss aversion, high social proof sensitivity, and moderate-to-high narrative susceptibility. The belief vector does not assign these archetypes — it defines them. An archetype is a region in the 12-dimensional space.

Beyond the static profile, the Nichesim architecture introduces a dynamic cognitive model that governs how a persona's beliefs evolve in response to new information. This layer comprises five parameters.prior_belief (0–1) sets the persona's starting position on a given topic before exposure to any evidence — a persona sceptical of cryptocurrency might begin at 0.25, while one enthusiastic about AI safety might start at 0.80.convincers are an ordered list of trigger conditions that increase belief: domain-specific predicates such as "sees a live demo," "reads a peer-reviewed study," or "hears a testimonial from a trusted source." When satisfied during simulation, the corresponding belief increments.dealbreakers are the inverse — triggers that decrease belief: "product misses a stated deadline," "founder has a prior failed venture in the same space," "revenue projections lack cohort retention data."evidence_preference selects the dominant evidence class for a given belief domain: "data" (quantitative metrics, benchmarks, statistical significance), "story" (narrative framing, case studies, founder journey), or "authority" (institutional signals, credentials, social proof).update_style determines how belief changes when a trigger fires: "jump" for immediate discrete shifts, "gradual" for smooth interpolation over multiple messages, or "stubborn" for resistance to change requiring repeated triggers. Together, these parameters transform the static belief vector into a stateful cognitive model. The 12 dimensions describe what a persona is; the dynamic layer captures how a persona becomes — and, critically, how it can be changed through sustained interaction.

Encoding a persona as a cognitive model requires translating qualitative profiles into both the 12-dimension belief vector and the 5-parameter dynamic model. The process uses a structured elicitation pipeline: a base LLM call produces an initial estimate from a persona description (covering both static dimensions and dynamic parameters), a refinement pass adjusts for coherence constraints (e.g., high contrarian tendency and high social proof sensitivity cannot both be near 1 simultaneously, and an "authority" evidence preference paired with a "stubborn" update style produces the slowest possible opinion shifts), and a final validation step checks that the resulting model produces consistent outputs across a battery of test prompts. The encoding pipeline looks roughly like this:

// Belief vector (static cognitive profile)
interface BeliefVector {
  riskTolerance: number;          // 0–1
  timePreference: number;         // 0–1 (higher = sooner)
  technicalDepth: number;         // 0–1
  domainExpertise: number;        // 0–1
  decisionSpeed: number;          // 0–1 (higher = faster)
  socialProofSensitivity: number; // 0–1
  contrarianTendency: number;     // 0–1
  lossAversion: number;           // 0–1
  narrativeSusceptibility: number;// 0–1
  infoProcessing: number;         // 0–1 (higher = analytic)
  trustModality: [number, number, number]; // competence, warmth, proximity
  valueAlignment: number[];       // embedding vector
}

// Dynamic belief-update model (Nichesim cognitive layer)
type EvidenceType = "data" | "story" | "authority";
type UpdateStyle = "jump" | "gradual" | "stubborn";

interface CognitiveModel extends BeliefVector {
  priorBelief: number;              // 0–1 starting position
  convincers: string[];             // triggers that increase belief
  dealbreakers: string[];           // triggers that decrease belief
  evidencePreference: EvidenceType; // dominant evidence class
  updateStyle: UpdateStyle;         // rate-of-change profile
}

function encodeArchetype(description: string): CognitiveModel {
  const raw = llmCall(ENCODE_PROMPT, { description });
  const constrained = applyCoherenceConstraints(raw);
  return validateOutputs(constrained);
}

The Gatekeep Implementation

Gatekeep is an experiment in whether AI investor personas can stress-test startup pitches. Founders submit their pitch materials — deck, memo, financial model, demo link — and receive structured feedback across twelve dimensions from virtual investors calibrated to specific funds, stages, and thesis areas. Each persona is backed by a belief vector, and early data suggests the quality of the output depends on how faithfully that vector constrains the LLM's reasoning.

The scoring pipeline runs as follows. First, the pitch input is decomposed into structured claims: market sizing assertions, competitive positioning statements, unit economics projections, team competency signals, and narrative framing elements. Second, a belief vector extraction step maps each persona's archetype specification onto the 12-dimension coordinate system described above. Third, a persona-matching algorithm selects the persona whose belief vector best corresponds to the stated investor profile. Finally, each dimension of the belief vector gates a specific evaluation lens — risk tolerance governs the persona's response to the financial projections, technical depth shapes its assessment of the technology moat, narrative susceptibility determines how much weight the persona gives to the storytelling versus the traction data, and so on.

The key design decision was to make the vector generative rather than merely classificatory. A naive approach would use the belief vector to select a pre-written template, but templates cannot anticipate the combinatorial space of possible pitches. Instead, each dimension functions as a soft constraint on the LLM's output distribution. A persona with risk tolerance 0.92 will, across any input, produce risk evaluations that are structurally different from a persona with risk tolerance 0.18 — not just different in tone, but different in which aspects of the pitch are foregrounded, which questions are asked, and which objections are raised. The result is feedback that reads as coming from a specific, consistent mind rather than a generic oracle.

Belief-as-State: Why Self-Grading Failed

Early prototypes of the belief-update system took the most direct approach: have personas evaluate themselves. After generating each response, the system appended a meta-prompt — "After saying this, how convinced are you (0–100)?" — and recorded the self-reported belief score. The approach was intuitive, computationally cheap, and conceptually elegant. It was also incorrect in ways that took three design iterations to fully diagnose.

Three distinct problems emerged. First, calibration failure: LLMs are systematically poor at numerical self-assessment. When asked to rate their own conviction on a 0–100 scale, models exhibit strong anchor bias (clustering responses near the midpoint or a primed value), scale confusion (treating 0–100 as a coarse Likert scale with effectively three buckets: low, medium, high), and a documented tendency to produce scores that correlate with linguistic surface features rather than the underlying reasoning content. A persona generating a sceptical rebuttal produces conviction scores of 72 on one run and 34 on the next — not because its belief changed, but because the act of self-assessment introduces stochastic noise uncorrelated with the persona's actual stance. Second, cognitive interference: the meta-prompt adds a secondary task to each generation step, degrading primary message quality. The persona is simultaneously asked to produce coherent investor feedback and introspect on its own mental state — objectives that pull in opposing directions, with the introspection task consuming context-window budget and attention that would otherwise go to evaluation quality. Third, and most consequential, meta-awareness collapse: when personas are instructed to report on their own beliefs, they begin performing belief rather than holding it. The discourse becomes self-conscious and theatrical, with personas explicitly narrating their shifts ("I was sceptical earlier, but now I am warming up to this approach"). Real humans do not think this way; they update beliefs without comment, and the commentary itself distorts the simulation. The meta-aware persona ceases to be a useful model of an investor and becomes a model of an investor who knows they are being watched.

The current architecture replaces self-grading with batched analytical assessment. Message generation runs at full fidelity with zero self-reflection: the persona responds naturally according to its belief vector and dynamic model, producing output indistinguishable from a non-self-aware agent. Every 10 messages, a separate LLM call ingests the buffered conversation transcript, extracts the belief-update triggers that fired (matching convincers and dealbreakers against message content), and analytically recomputes each belief dimension. The scoring prompt is structured as a third-party observation task rather than an introspective one: "Given the following conversation, what has shifted in this persona's beliefs? Provide numerical updates with justification." The result is a system where 92% of belief shifts match the sentiment direction of the triggering messages, compared to 61% under the self-grading approach. The persona acts; the system observes. This separation of generation and evaluation is, we argue, the correct architectural boundary for any multi-agent belief simulation.

Empirical Validation

We benchmarked Gatekeep's persona outputs against a corpus of real VC feedback gathered from rejections, diligence notes, and partner meeting summaries. Across 47 pitch-pairs — matching the same deck evaluated by both a Gatekeep persona and a real investor sharing the same stated thesis — the persona outputs achieved 78% thematic overlap with human feedback as measured by blinded evaluator agreement. More importantly, the personas surfaced objections that the human evaluators missed in 14% of cases, typically on dimensions where the persona's belief vector introduced a deliberate contrarian skew. These are not hallucinations; they are structured disagreement programmed into the architecture.

Persona coherence — whether the same persona produces consistent evaluations across repeated runs — was measured by running each persona against a fixed battery of 20 pitches five times and computing pairwise agreement scores on the resulting verdicts. Mean intra-persona agreement was 0.91 (Cohen's kappa, p < 0.001), confirming that belief vectors produce stable behaviour. Edge cases emerge when two dimensions conflict sharply — for instance, a persona configured with both high narrative susceptibility and high analytical processing produces oscillatory outputs where the persona "wavers" between story-driven and data-driven evaluations. We treat these conflict zones as feature, not bug: they correspond to real cognitive tensions observed in human evaluators, and exposing them to founders is, in practice, more useful than a perfectly stable but unrealistically coherent persona.

Cognitive Models in Production

The cognitive architecture described here powers two production systems. Gatekeep deploys investor-tuned belief vectors to simulate VC feedback on startup pitches: the static dimensions (risk tolerance, technical depth, decision speed) generate the substantive evaluation, while the dynamic model governs how each persona's position evolves if a founder iterates and re-submits — a "jump" update style means the persona overweights the latest deck version, while "stubborn" means the persona anchors on its initial impression. Nichesim deploys the same architecture at community scale: thousands of interacting personas, each with a full cognitive model, simulate how a product launch propagates through a market. Early adopters update first (high update_style = "jump"), sceptics shift only after repeated exposure (high update_style = "stubborn"), and dealbreakers gate whether a persona drops the product entirely when specific failure conditions are met. The same five-parameter dynamic model that governs a single investor's evolving opinion on a pitch deck scales seamlessly to modelling adoption cascades, backlash dynamics, and network effects across entire synthetic communities.

The encoding pipeline for both systems follows a consistent three-stage process. Stage one produces a raw LLM vector estimate from a natural-language persona description — a few paragraphs covering background, dispositions, and domain attitudes across all 17 parameters (12 static dimensions plus 5 dynamic). Stage two applies coherence constraints: dimensions that create known conflict zones are flagged and adjusted. The tension between high narrative susceptibility and high analytical processing, for instance, produces constructive oscillatory outputs where the persona genuinely wavers between story-driven and data-driven assessments — a dynamic we preserve intentionally rather than suppress, because it mirrors the productive cognitive dissonance observed in real evaluators. Stage three validates the model against a battery of calibrated test prompts and measures inter-run consistency. These conflict zones, where orthogonal dimensions create unstable belief dynamics, are not bugs in the architecture. They correspond to real cognitive tensions that make human evaluators interesting. A persona that rigidly adheres to a single processing frame produces predictable but shallow feedback; a persona that wavers produces the kind of nuanced, self-contradictory evaluation that founders actually encounter in pitch meetings. The variance is the value.

Broader Implications

While Gatekeep applies belief vectors to investor personas, the architecture generalises to any domain where judgment simulation is the objective. A belief vector can encode a procurement officer, a regulatory reviewer, a hiring manager, or a community moderator. The dimensions may change — a procurement officer cares more about vendor reliability than contrarian tendency — but the structural insight holds: decision-making dispositions, not demographic labels, are the right primitive for synthetic personas. This is the animating principle behind Nichesim, our platform for simulating community-level reactions to product launches, where thousands of interacting belief vectors model the emergent dynamics of adoption, backlash, and network effects.

The broader trajectory points toward synthetic user research as a complement — and in some contexts, a replacement — for traditional focus groups and survey panels. The cost argument is compelling (see Chua, "The Economics of Simulated Attention," 2025), but the deeper advantage is methodological. Build a community of personas with calibrated belief vectors, and you can test not just one product but every variant, not just one message but every frame, not just one launch strategy but every rollout sequence. The combinatorial explosion that makes traditional research expensive becomes, in simulation, merely computational. The challenge ahead lies not in generating more personas but in validating them — a topic we address in "Coherence, Not Accuracy: Evaluating LLM Personas at Scale" (2026), which formalises coherence as the correct evaluation target for synthetic personas and proposes a gradient-descent scoring method for persona quality assessment (see also Park et al., 2023, on generative agent architectures; Shao et al., 2024, on personality-consistent LLM behaviour; and the broader literature on belief-desire-intention models in multi-agent systems, Rao & Georgeff, 1995).