Simulatte Research Pitch Deck Teardown: A Technical

Simulatte Research: A Technical Deep Dive into Cognitive Persona Architecture

The document provided by Simulatte Research is titled "Cognitive Persona Architecture: A Multi-Engine Framework for Simulating Venture Capital Decision-Making." Dated March 2026, this is a highly technical whitepaper that serves as a foundational document for a simulation platform. It eschews the standard startup pitch deck tropes—there are no slides on market growth or competitive landscapes. Instead, it focuses entirely on the engineering of cognitive bias in large language models (LLMs).

Slides 1-3: The Thesis and the Problem

The deck opens with a clear problem statement on Slide 5 : "LLM Persona Homogeneity." The authors argue that standard LLMs, when asked to role-play as VCs, tend toward balanced, analytical evaluations. They correctly identify that real VCs are "cognitively constrained, emotionally biased, and archetypically distinct."

The core thesis presented on Slide 5 is that authentic simulation requires "Pre-LLM Cognitive Filtering." Rather than asking an LLM to ignore information via a prompt, the system deterministically removes that information from the input. This is a significant architectural choice that moves the "intelligence" of the system upstream from the generative model.

Slides 4-12: Persona Construction and Data Sourcing

Slide 7 introduces the 9-dimensional persona vector. This vector is expanded into a 4-layer schema: evaluation signals, individual markers, firm signals, and derived signals. This slide also introduces 11 "Decision Archetypes," such as the "convictionbuyer" (who amplifies founder story signals by 1.5x) and the "dataoptimizer" (who suppresses narrative strength by 0.6x).

On Slide 9 , the deck explains how these personas are built. The system uses a multi-source discovery phase across 18 archetypes, scraping blog posts, podcasts, conference talks, Twitter/X, and news interviews. The use of Playwright and BeautifulSoup for extraction is mentioned, grounding the technical claims in specific tooling. Slide 11 details "Layer 2: Individual Markers," which include temperament (e.g., "directbutwarm") and notable phrases (e.g., "Mutants can look ugly and dwarfed").

Slide 13 provides a worked example of "Investor K," an anonymized seed-stage investor. The pipeline extracted 23 high-confidence beliefs from 47 sources, resulting in a persona with a high "founderfocus" (0.85) and high risk tolerance (0.80).

Slides 13-19: The Three-Engine Architecture

The meat of the technical proposal is found in the description of the three engines. Slide 15 details the "Perception Engine," which uses a 4-step pipeline: Extract, Hard Dropout, 9-Stage Distortion, and Attention Allocation. The "Hard Dropout" is particularly notable; it removes signals like "teampedigree" or "marketsize" entirely if they conflict with the investor's core beliefs.

Slide 17 describes the "Cognition Engine," which uses a "7-Priority Pre-Decision State Machine." This is a heuristic-driven model where a "VETO" (Priority 1) can force a "PASS" regardless of other signals. This mirrors the reality of venture capital where a single deal-breaker often outweighs a dozen positive indicators. The formula for "Polarity-Aware Belief Activation" is also provided, showing how signals are weighted against match quality.

Slide 19 covers the "Decision Engine," which imposes seven hard constraints on the LLM's output. These include a "Max 3 Focus Areas" rule to force prioritization and a list of "11 generic VC phrases" that are banned to prevent the model from falling back into generic "balanced analysis."

Slides 20-23: Validation and Comparative Results

Slide 21 returns to the "Investor K" example, showing the results of the Perception Engine. In this simulation, the investor's reality contained "NO traction data, NO unit economics, and NO market size information." This demonstrates the system's ability to force the LLM to ignore critical data points that a specific investor archetype would deem irrelevant.

Slide 23 is the most critical slide for a potential investor or partner. It presents a comparative evaluation of the CPA system against GPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro. The CPA system outperformed vanilla models across all six dimensions, most notably in "Internal Tension" (94% vs. 22% for Claude) and "Information Omission" (76% vs. 18% for Claude). This provides quantitative backing for the claim that prompt-based role-playing is insufficient for high-fidelity simulation.

Slides 24-33: Conversation Dynamics and Humanization

The deck moves beyond static evaluation into multi-turn conversation dynamics. Slide 25 introduces a "Conviction Volatility Model" and a "Topic Memory State Machine." The latter ensures that topics in a "resolved" state are never revisited, preventing the repetitive loops common in LLM interactions.

Slide 27 details "Signal Exploration Control," which enforces a "probe -> deepen -> shift" pattern. This is designed to mirror the cadence of a real pitch meeting. Slide 29 shows how temperament modulates tone across early, mid, and late phases of a meeting. For example, a "patient" investor moves from "warm curiosity" to "gentle probing" to "supportive framing of concerns."

Slide 31 and Slide 33 focus on the "Humanization Layer" and "Behavior-Driven Negotiation." The system uses a readiness assessment (e.g., "Topic depth >= 2 deeply explored topics") to trigger deal discussions, rather than a simple turn counter. This adds a layer of organic realism to the simulation.

Slides 34-39: Quality Control and Limitations

The final section of the deck covers the system's internal quality metrics. Slide 35 introduces a "Data Quality Score" (0-100) based on ten weighted dimensions, such as "Vector Completeness" and "Archetype Coherence." Slide 37 outlines mechanisms for "Persona Identity Drift Prevention," including a "Role Drift Check" where the model explicitly asks itself if it is drifting toward being a "helpful advisor" rather than a "VC evaluator."

Finally, Slide 39 acknowledges current limitations. The system currently relies on keyword-based extraction rather than semantic understanding and uses static persona vectors that don't account for how an investor's thesis might evolve over time. The "Future Directions" section suggests moving toward semantic extraction and multi-persona interaction modeling (simulating an entire IC meeting).

What Works in This Deck

Technical Depth: The deck provides actual formulas and state machine logic. This isn't just a "we use AI" pitch; it's an architectural blueprint. · Specific Problem Identification: By focusing on "LLM Persona Homogeneity," the founders have identified a specific, technical weakness in current generative AI and proposed a structural solution. · Validation Data: The comparative table on Slide 23 is essential. It moves the conversation from "this sounds cool" to "this is measurably better than the status quo." · Worked Examples: The "Investor K" thread throughout the deck makes the abstract architecture concrete and easy to follow.

What Is Missing

The Business Case: There is no mention of how this will make money. Is it a SaaS tool for founders? A data product for LPs? A training tool for junior VCs? · Market Size: The deck assumes the value of VC simulation is self-evident, but it provides no data on the potential market for such a tool. · The Team: There is no team slide. In a technical deck like this, knowing the background of the researchers (e.g., PhDs in Cognitive Science vs. LLM Engineers) is vital. · The Ask: The deck ends with a conclusion but no call to action. There is no information on how much capital is being raised or what the milestones for the next 12-18 months are.

Founder Takeaways

Architecture Over Prompts: If you are building an AI-native application, show that your value add is structural. Simulatte Research demonstrates that their value is in the pre-processing and state management, not just a clever system prompt. · Use Quantitative Comparisons: If you claim your tool is better than using a raw LLM, you must provide a table like the one on Slide 23. Define your metrics (e.g., "Cognitive Consistency") and show the delta. · Solve for 'Drift': Any long-form AI interaction suffers from context drift. Showing that you have built-in "Self-Check" mechanisms (Slide 37) to keep the AI on track builds significant trust with technical investors.