research proposal (to get feedback)

[hyunjimoon]

Abstract - A concise summary of the motivation, research objectives, methods, and key results. A person unfamiliar with the topic should be able to understand the abstract.

Introduction – [≈ 2 pages] At least one paragraph should be written for a general audience, clearly enunciating why someone outside your field should care about this work. The introduction should also contain a detailed literature review that explains how your research done to date is related to previous research in the field and what your research will add to the larger body of research. Finally, the introduction should clearly state the short- and long- terms goals of the work, connecting to a broad engineering or societal problem that motivates the work.

Methods – Describe and defend your methods, including your assessment of uncertainty.

Include citations of previous applications of the method.

Results - Describe specific results, including a careful explanation of the uncertainty. Discussion - Compare and contrast the results with other studies, including citations.

Explain how you have addressed the technical questions and long-term applications mentioned in the methods and introduction.

[hyunjimoon]

General Exam Part 2 [Research Aptitude] This exam tests the following skills. First, can the student formulate a research question, set out a plan of research, and interpret the results. Second, can the student clearly present and defend this research. Third, does the student have sufficient understanding of the field to answer a broad range of questions and to comment on relevant literature. The research presented by the student can be drawn from their MST thesis, their current RA at MIT, or research conducted as part of a previous position. The research must be in the same field as the subgroup core listed in the Doctoral Program.

Part 2 of the General Exam must be completed by the end of the fourth academic term. In most cases, this exam occurs in April or May of AY2. Students should consult their research advisors when choosing the members of the evaluation committee (see requirements below). Your Part 2 general exam committee should be similar or identical to the thesis committee.

Students must submit the Part 2 Schedule Form [available at https://cee.mit.edu/resources/] before the beginning of the term in which the exam will be held, which is the end of January for exams in April/May of AY2.

The exam has three components:

1. A written document describing research completed.

2. A review of a relevant publication chosen by your advisor. The paper will be assigned one week before the presentation meeting.

3. A 30-minute oral presentation of research completed with significant questioning from committee.

Research Paper The research paper should have a maximum of 10 pages, single-spaced, with 12-pt font and 6

should review research done to date. The page limit includes figures, but not references.

Students may ask their advisor for advice in the preparation of this document. Students may also get assistance from MIT’s Writing and Communication Center, http://writing.mit.edu/wcc and the CEE Communications Lab, https://mitcommlab.mit.edu/cee/. The following elements must be included (📝above)

The student distributes the research report to their committee a minimum of one week before the presentation. The student should inquire whether each committee member prefers a pdf or hard copy and deliver the preferred format.

Review of a Relevant Publication Your advisor, in consultation with the GE2 committee, will select a single journal publication in your field. It may be a seminal paper from years ago or a new paper. It should not be longer than 20 pages and cannot be too broad, e.g., no general reviews of the field, nor should it be written by a member of the committee. You should be prepared to informally discuss the paper (no slides), focusing on a set of 2 to 4 questions that will be provided by the committee when the paper is assigned. The questions may include some of the following, or they may be more specific to the paper. Please prepare for the paper review on your own.

• What is the most important result and why is it significant?

• What is the value of the paper to the broader field?

• What are the limitations of the work and results presented?

• What is the most significant uncertainty and how could it be reduced? How do the 7

results of this paper relate to your research?

• Please show the full derivation of equation (5). Are all the conclusions justified by the results? Are the boundary conditions realistic?

• Explain in physical (chemical, biological) terms why the relationship shown in Figure 7 makes sense or does not make sense.

• How does this paper challenge the existing theory regarding ?

• Propose a new research question or hypothesis that expands on the work presented in this paper, i.e., where would you go next? Defend your choice.

Oral Presentation Meeting The student should schedule the committee meeting for 2 hours. The Chair of the Committee, who will be sent specific instructions before the meeting, runs the meeting. The student will begin by informally presenting their response to the question(s) posed by the committee regarding the paper chosen for review (see Review of a Relevant Publication above). The student should not prepare slides for this response. Necessary visuals or equations can be sketched on the black board. Committee members may ask questions for clarification or to go into further depth. After twenty to thirty minutes, the committee chair will end this discussion and instruct the student to begin their research presentation on work done to date. The student should plan a 30-minute presentation, but the actual presentation will take longer as faculty will interrupt with questions. The committee members are expected to have read the report and come prepared with questions. The committee members should push questions to the point at which the student says, “I don’t know”. The student should not be afraid of saying, “I don’t know”. It is at this point that the real scientific discourse begins, an exchange of ideas that provides a learning experience for the student. It is important to note that the research advisor is encouraged to ask questions, but he/she should not answer questions. This is a test of the student's understanding and research ability, not a test of the advisor’s research ideas. The GE oral presentation also serves as a practice for the student in preparation for their thesis proposal, which has a similar format, and is distinct in topic as it will talk about future research.

Evaluation Committee for General Exam Part 2 The evaluation committee for Part 2 is comprised of three people, including the student’s thesis advisor. The committee must have a minimum of two faculty or Senior Research Staff in the Transportation Education Committee (TEC). In many cases this group will become the Doctoral Thesis Committee. The chair of the evaluation committee must be a faculty member or Senior Research Staff in TEC and cannot be the thesis advisor. The student invites the committee members and includes their names on the Part 2 Schedule Form.

General Exam Part 2 Outcomes After the exchange of questions and ideas has finished, or at 1hr 40 min, whichever comes sooner, the committee chair will ask the student to leave and wait nearby. The faculty advisor will be given a few minutes to add their perspective on the student’s performance that day, on the student’s broader research ability, and any specific requirements for the student. The committee chooses one of the following outcomes, which must be communicated with the

[hyunjimoon]

CROSS-FIT🏋🏼

benchmarking dynamics of fit across fit

Strategy

value creation hypothesis: system innovation of machine that outperforms human on integrated decision support for entrepreneur’s operation can discover value for new users knowledge producers in entrepreneurship (educators and scholars) who suffer from developing and communicating their product that entrepreneurial practitioners can apply

value capture hypothesis: executing/controlling on key integrated capabilities user friendly integrative entrepreneur’s operations partner using probabilistic program and language model that formalizes and automates designing business model, valuating equity, tuning parameters, estimating rare event probability, enabling conversational inference by coherently answering questions can capture value

Product develop plan

Knowledge transfer plan

using evaluating founder fit frameworks cld

	Components	Deal-Investor Fit (2024-2025)	Product-Market Fit (2025-2026)	Probabilistic Program-Entrepreneur Fit (2026-2027)
a* = argmax E[u(e(s,a))]	utility u	Maximize alignment between deal terms and investor expectations	Optimize product features and positioning for target market segments	Maximize synergy between entrepreneur's decision-making and probabilistic program's outputs
	State s	Investor preferences, market trends, company valuation metrics	Customer needs, market segmentation, competitive landscape	Entrepreneur's decision patterns, cognitive biases, strategic thinking approaches
	Action a	Adjust investment amount, valuation cap, equity allocation	Modify product features, pricing, go-to-market strategy	Tune program interfaces, decision support algorithms, feedback mechanisms
	Environment Model e	Financial markets, investor behavior, economic indicators	Market dynamics, customer behavior, technological trends	Decision-making contexts, entrepreneurial cognition, AI-human interaction
	Backward Process (Inference) p(s|D)	Update beliefs about investor preferences and market conditions	Refine understanding of customer needs and market fit	Improve model of entrepreneur's decision-making patterns and biases
	Forward Process (Simulation)	Project outcomes of different deal structures	Predict market responses to product changes	Simulate entrepreneur's decisions with different AI support scenarios
	Simplify	Decompose complex deal structures into key components	Break down market segments and product features	Identify critical decision points for AI augmentation
	Choose	Select optimal terms and conditions for the deal	Determine target markets and core product features	Decide on level and type of AI decision support
	Reason Probabilistically	Model potential outcomes of different deal structures	Estimate market size, adoption rates, and competitive responses	Apply Bayesian networks to complex business decisions
	Calibrate	Refine deal terms based on investor feedback and market conditions	Iterate product based on user feedback and market performance	Adjust AI suggestions based on entrepreneur's decisions and outcomes
	Pivot Levels	1. Operational: Adjust specific deal terms 2. Strategic: Revise entire deal structure 3. Dynamic: Adapt overall fundraising strategy	1. Operational: Modify product features 2. Strategic: Pivot to new market segment 3. Dynamic: Transform business model	1. Operational: 2. Strategic: 3. Dynamic:
	Conversational Inference	Use LLMs to translate complex financial terms and investor expectations into probabilistic models. This can help bridge information gaps between founders and investors, simulating different stages of financing and their potential outcomes. The system can engage in nuanced dialogues, helping both parties understand and quantify risks and opportunities at each funding stage.	Leverage conversational inference to translate qualitative customer feedback and market trends into quantifiable metrics. This can help in modeling the complex relationships between product features, market needs, and competitive landscape. The system can simulate different product-market scenarios, helping entrepreneurs make data-driven decisions about product development and market positioning.	Employ conversational inference to create a dynamic, adaptive interface between the entrepreneur and the probabilistic program. This can help in translating the entrepreneur's intuitions and experiences into parameters for the probabilistic models, and conversely, translating complex model outputs into actionable insights. The system can engage in ongoing dialogue, continuously refining its understanding of the entrepreneur's decision-making process and adapting its support accordingly.

todo: what bilateral information asymmetry does this solve
josh lerner recommended Chains.pdf and i feel Financing Entrepreneurial Experimentation and financing risk and innovation by ramana and matthew are relevant

[hyunjimoon]

Module	Complexity	Field	Evaluators
1. Differentiable programming	0+1+0= 1💡(📍🌲)	Prob. program + language model	Charlie
2. Program synthesis	0+2+0=2💡(🎲🌲)	Prob. program + language model, Investment structuring	Charlie, Vikash
2. Bayes decision + game	1+1+0=2🧍‍♀️(📍🌲)	Environments for machine co-founder, Prob. program + language model	Charlie, Tom
3. Capitalize	0+2+1=3💸(🎲🕸️)	Investment structuring	Vikash
3. Innovation/entrepreneurial strategy	1+2+0=3🧍‍♀️(🎲🌲)	Machine co-founder	Tom
3. Resource rationality	1+1+1 = 3🧍‍♀️(📍🕸️)	Investment structuring, Prob. program + language model, Environments for machine co-founder	all
4. Metrics	2+2+0 =4🏢(🎲🌲)-🧍‍♀️(🎲🕸️)	Environments for machine co-founder, Innovation quality	Tom, Scott
4. Program lang. design	1+2+1= 4🧍‍♀️(🎲🕸️)	Prob. program + language model, Investment structuring	Vikash, Scott
5. Meaningful innovation measure	3+1+1=5🏢(🎲🕸️)	Innovation quality	Scott

---

expected feedback on angie's education from committee members

Critic	Perspective	Critique
Vikash	Probabilistic Programming	While Angie's education covers some aspects of probabilistic programming, there seems to be a lack of depth in program synthesis and differentiable programming (modules 1 and 2). More focus on these areas could enhance her ability to model complex systems probabilistically.
Scott	Strategy and Economics	Angie's education provides a strong foundation in microeconomics and innovation economics. However, there could be more emphasis on game theory and strategic decision-making in uncertain environments (module 3). Additionally, exploring the intersection of strategy and computational methods could be beneficial.
Charlie	Operations and Innovation Management	The education covers important aspects of optimization and innovation management. However, there could be more focus on the operational aspects of implementing innovation, particularly in rapidly changing environments. More emphasis on dynamic resource allocation and adaptive strategies (module 6) would be valuable.

[hyunjimoon]

evaluators' world model and their roles

Model	Key Focus	Core Concepts	View on PC4BE
Charlie's	Dynamic capabilities and adaptive operations	Industry evolution, value chains, clockspeed, modular vs. integral design, "nail it, scale it, sail it" framework	Tool to operationalize entrepreneurial strategy
Vikash's	Probabilistic programming as an alternative to deep learning	Human-like understanding, rational meaning construction, automated differentiation of expected values	Unifies approaches to meaning in language and thought
Scott's	Beliefs and learning in entrepreneurial decision-making	Experimentation, updating priors, information asymmetry	Framework for understanding and improving entrepreneurial strategy

#263

[hyunjimoon]

need feedback on the word choice of four products

1. is collaborating vs segmenting a representative/specific enough word choice? other candidates:

• coo vs cmo role
• internal capa vs external capa
• lowering approximation error vs statistical error
• increasing size of hypothesis space vs observation

collaborating vs segmenting's cons: jeff's question "it's almost like comparative statics, right? So you isolate a couple of things and just focus on one" -> depends on whether our study would be theoretical vs empirical (measure). any situation where agent is not functionally diversified much (or aggregated data is observed) so that the unit of observation behavior matches that with decision? three components of objective function translates into three actions (~ three causal loops?)

[hyunjimoon]

my talk slide: angie moon rational meaning construction.pdf
and below explains the measure i took to converge to this target distribution

PRIOR - how i calibrated

⭐️bayesian calibrated prior: action-oriented encoding of belief that satisfies self-consistency
Q1. can we frame this as reaching to minimal equilibrium between founder and investor?

measuring my mentors' excitement on topics guided this talk: integrating inference and decision (charlie), exchangeability (scott) simulation-data experimentation for hierarchical model (andrew), prior as action-oriented encoding of belief + state eliciting algorithm (jeff) - below ordering only has temporal meaning.

andrew gelman

Simulation-based calibration. You have a computer program to get posterior simulation draws p(theta|y). You want to check that this program is doing what it’s supposed to do. You perform this check by first drawing theta_tilde from the prior, p(theta), then drawing y_tilde from the data model, p(y|theta_tilde), then running the program to get a bunch of draws of theta from p(theta|y_tilde). Putting it together, the three things you’ve sampled come from a joint distribution, p(theta_tilde, y_tilde, theta), that is symmetric in theta_tilde and theta. Another way to put it is that posterior inferences based on the simulations of theta should have the nominal coverage, averaging over the prior. One way to do this checking would be to repeat the above process (drawing theta_tilde, then y_tilde, then inference for theta) 1000 times, and then check that the posterior 50% interval includes the true theta_tilde approximately 500 times, that the posterior 95% interval includes the true theta_tilde approximately 950 times, etc. What we actually recommend is to look at the quantile of theta_tilde from each fit and check that these quantiles are normally distributed.

• simulation-data experimentation for hierarchical model (2023.Sep)

Are you aware of any literature on simulated-data experimentation for hierarchical models that discusses some choices on which order joint distribution p(phi, alpha, y, phi_tilde, alpha_tilde) should be disassembled for testing? I just verified the usefulness of conceptualizing the process of simulation-data experimentation in an entrepreneurial context (from two entrepreneurship educators/professors) so the math need not be rigorous. You had some thoughts back in 2021 below, the period when we excitedly delved into sbc (https://statmodeling.stat.columbia.edu/2021/09/03/simulation-based-calibration-some-challenges-and-directions-for-future-research/), and I’m curious about the current state of your thinking-ahead-thoughts.
we’re often fitting hierarchical models, and then we can follow a hybrid approach. For a hierarchical model with hyperparameters phi and local parameters alpha, we can set phi_tilde to a reasonable value and draw alpha_tilde from its distribution, p(alpha|phi_tilde). If the number of groups is large, there can be enough internal replication in alpha that much can be learned about coverage from just a single simulated dataset—although we have to recognize that this can all depend strongly on phi, so it might make sense to repeat this for a few values of phi.

I agree that these are interesting questions. The only literature I know that's related to this at all is:
(a) this paper (not on simulation experimenting but on predictive checks: https://www.sciencedirect.com/science/article/pii/S0378375802003038)
(b) the Cook et al. (2006) paper where we discuss different sorts of checks of hierarchical modeling
(c) more recent work from the BLS team (Terrance Savitsky et al.) doing simulation-based calibration, where they get good results using posterior (rather than prior) simulation, and I think the reason it's working is that they're doing this in the context of hierarchical modeling so that their posterior simulation is conditional on hyperparameters and is thus almost a prior simulation (to the extent that the hyperparameters are well estimated from data.

after this conversation, andrew wrote: Understanding posterior recalibration for a simple example

Andrew frames this as simulation-based experimentation.

Simulation-based experimentation. Data analysis can be expensive in time and effort, and this can lead to us thinking that if a project took a lot of work then it has to be good. To state that belief is to mock it, yet it persists. How can we avoid what might be called the “fallacy of effort”? We recommend simulation-based experimentation, which requires the following steps: (1) create a fake world, (2) simulate parameters and data from this world, (3) analyze the simulated data and get inference for the underlying parameters, (4) compare those inferences to the parameter values simulated in step 2. This can be done systematically in a Bayesian context (Modrák et al., 2024) but in practice informal checking can work just fine, in that problems will often show up in a simple simulation. Creating a fake world is not easy—if analyzing a dataset is like playing Sim City, simulating fake data is like writing Sim City—but this effort can be well worth it, not just for the benefit of uncovering problems and the increased confidence arising from successful recovery, but also because constructing a simulation experiment is a way to clarify our thinking. Indeed, we often recommend simulating fake data before embarking on any real-world data collection process, to get a sense of what can realistically be learned from a proposed design.

charlie fine

(tbc)

scott stern

(tbc)

jeff dotson

(tbc)

POSTERIOR

Feedback in order of actionability

utah

Jay Barney

• he liked my two by two table and recommended framing it tech-understanding and patent valuing. strategy ppl (e.g. andrea pignatoro) had issues with the word choice "meaning" which is from "from world to world" paper by josh tenenbaum's school so i'm considering changing it.

• when i asked where to start from, he recommended recent review paper on patent quality and value: Ananthraman23_Divide and conquer Relating patent quality and value in a conceptual.pdf
• he shared why he got into bayesian (average effect doesn't make any sense for strategy prescription perspective). he was very proud of his SMJ paper with jeff (and revolutionary potential of it) and expressed sadness that strategy field is not following Bayes more - "maybe 15 years more people will know the value of that paper more?" two causes were Bayes has much to learn and illusion of rigor from identification strategy were his two causes of this trend.
• he was very skeptical on using econometric measures (identification; 2SLS, IV) to study strategy. I shared MackeyDotson24_Bayesian Stats in Mangement.pdf on essential heterogeneity.

Teppo Felin (forward looking causal logic)

• (before the talk) his positive reaction on transportation/mobility analogy shaped the final slide. high uncertainty in the states in between but low uncertainty on the origin and destination. he resonated with my skepticism on causal logic. he seems to one of the few people who can provide real examples and mechanism to develop the diagram below:

• (after the talk) he really liked transportation metaphor (building on the materialized subpath) and agreed with my push back on arnaldo's comment on adding up subjective probability originated from different agents.

Todd Zenger

• he said "great talk" but thought the talk was a little abstract and requested to send him written paper

Timo Ehrig

Timo thought bayes rule's entropy always decrease but when i showed bayes self-consistency equation, he was very surprised. he named this as "learning rule with self consistency" which i liked. my intuition is path dependent/irreversibility happens as people make decisions that are not self-consistent. if they do (simple mechanism to check the coverage i.e. when i apply my inference on the simulated data, does the posterior 50% interval includes true theta_tilde 50% times (if time permits they can keep testing 25%, 75% - this incremental test specificity increase can be designed by setting test quantities)). i feel this calibration can mitigate decrease of entropy and path dependent situations. this may be implemented as maintatining local balance between the observation-state flow and state-action flow.

Bocconi

Alfonso Gambardella

• I questioned Alfonso's dichotomy framing "belief vs likelihood (you can have belief about likelihood), theory vs alternative theory (you can always make higher dimension mixture model)" and my frustration on distinguishing state and action. he resonated with both and recommended reading camuffo24-theory-driven-strategic-management-decisions.pdf then having more conversation.

From

Arnaldo Camuffo

advised institution would be interesting in matching (founder and investor). i disagreed. algorithm from founder (😈maxwell sorting demon in moving system)'s perspective would be more meaningful and more cumulative

Mike Ryall

• interested in LLM experiment setting and asked robustness check on the result. context and content validity
• honestly i wonder whether the test for comparing synthetic agent and real person is necessary. two main goal is to algorithm to infer the family tree of investor which is both generic and customized. it is generic as algorithm being applied is the same but is customized as the classified (resulting tree) would be different. in other words, investor family tree is calibrated prior.

friends

Liz Calder shared she liked seeing visuals (not having to read words)

MJ advised me to think about how to pitch. To start with try Bayesian ENT ppl (economist background; scott stern and Joshua gans) and Arnaldo Camuffo.

Aticus agreed consideration of time/pace is the biggest gap between theory and practice in ENT. learning happens on
Aticus<>Angie agenda123.pdf

---

⭐️would like to elicit more thoughts on irreversibility and path dependency

[hyunjimoon]

my heart is converging twd "exchangeability in entrepreneurship" as thesis theme.

infer 🗣️2🧠, 🧠2🤜 (understanding, judging) with ⚙️🗺️, ⚙️📐 (finding path, calibration plot)given 🗄️🗣️🧠🤜 (investor's belief, perception, action) described in Data4DM/decode-venturing@7115373#commitcomment-151790051

todo: connecting with ✂️dimension reduction via rejection sampling from #234 (reply in thread)

[hyunjimoon]

👥->🔥type123 diagram using andrew's sbc theorem, arnaldo's aom talk alfonso's paypal encryption example cld

[hyunjimoon]

💠Applying Thomas and Wanda's 4-leadership capability model for choosing Angie's thesis theme: ⚖️exchangeability in entrepreneurship

👁️visioning

Among topics of emerging bayesian entrepreneurship (BE) and theory-based view (TBV), Angie with focus on "parsimonious persuasion under uncertainty" using simulation-based calibration. (thesis title: exchangeability in entrepreneurship)

🫳sense-making

deep understanding of three samples

• my entrepreneurial experience (which I have full information on intentionality)
• Kyungsoo (MIT MBA on construction + AI, part of designX, who promised to share all his progress)
• Vikash (IT) or Kiyeon (health + AI)

👥relating

👥with 🧭TBV/strategy audience

• Parsimony as architectural propertise of theories (low belief necessary for breakthrough (big update in probabilities)orthogonal attributes (bigger update (fewer, attributes, parsimony)) -> modeling time series from polynomial vs fourier model class from josh's class. AGAIN AGT problem. existing models are all 🌍👀(external observer detached from action) not 👁️👀🧠🤜(internal actor in action-oriented situation). problem with josh's approach is, model class complexities cannot be measured nor compared by an agent (esp.ex-ante). entrepreneur cannot know the complexity of all potential business models given theory1 and ~theory1 (theory1 = AV with "vision only" technology) as they update with marginal likelihood ratio: P(y|theory1)/P(y|~theory1) where y = self-driving future, theta = implementation (e.g. strategy, business model)

• exchangeability as mechanism of changing mind -> instead of searching for signal that best rejects one's theory, we extrapolate theory using hierarchical modeling (goal integration; above ⛏️mining evaluation metric with exchangeability). two types of surprises:

1. what we believed to be distinguishable turns out to be indistinguishable 2) 🔬 what we believed to be indistinguishable turns out to be distinguishable.

👥 with 💸BE/economists audience

• agree with joshua's "believing in exploiting can be rationalized by choosing lowbar experiment", disagree with joshua's assumed intention of entrepreneurs looking for signals to reject one's theory. theory/belief is given and learning doesn't happen on that level.
• what affects "profitable abstraction" (lowered sampling cost, model parsimony)

todo: connect similar 💡ideas on false positive and false negative cost ratio

$1: \lambda$

$L(a, \theta)=1{\theta_1&gt;\theta_2, a_1&lt;a_2}+1{\theta_1&lt;\theta_2, a_1&gt;a_2}- \ \lambda(1{\theta_1&lt;\theta_2, a_1&lt;a_2}+1{\theta_1&gt;\theta_2, a_1&gt;a_2})$
from Andrews25_lambda_rank.pdf

$1: \lambda_0, \lambda_1$

Entrepreneurial-optimal experiment: $\arg \max _{{\lambda_1, \lambda_0} \in E} \lambda_1 \mu_E V-(\lambda_1 \mu_E+(1-\lambda_0)(1-.$ $..\mu_E)) C$
Investor-optimal experiment: $\arg \max _{{\lambda_1, \lambda_0} \in E} \lambda_1 \mu_O V-(\lambda_1 \mu_O+(1-\lambda_0)(1-\mu_O)) C$

from Gans24_exp_choice_lowhighbar.pdf

$\gamma$

Graphical test for discrete uniformity and its applications in goodness-of-fit evaluation and multiple sample comparison

from Teemu Paul Aki 22 Graphical test for discrete uniformity.pdf

👥 with 📊Stats./CS audience

three excerpts from andrew's blog:

calibration does not satisfy everyone as a means of ensuring good decisions
from How far can exchangeability get us toward agreeing on individual probability?

Angie asked: Could an algorithm that helps individuals determine the optimal test quantity for their specific situations address this issue? If there’s potential demand for such an algorithm, I’d be eager to explore this idea further. I’m also curious whether the workflow focused on test quantity, as outlined in the poster below (Fig.2), aligns with practical needs:
https://github.com/Data4DM/stanify/blob/main/vignette/FiddamanMoon23_BayesSD.pdf

aspects of “nonexchangeability” might be better viewed as structure that can be modeled.
from understanding exchangeability instatistical modeling a 🎃thanksgiving themed post

Angie pitched exploring random with intention to @jeff-dotson and @tomfid #23 (comment)

in 10 years, statistical theory will be mostly about gaining a deep understanding of simulation as profitable abstraction of counter-factually repeatable phenomenon.
from Somethings do not seem to spread easily – the role of simulation in statistical practice and perhaps theory.

Angie concluded simulation model = prior distribution

	Simulation Model	Prior Distribution
🧠2🤜	profitable	action oriented
👀2🧠	abstraction of	encoding of
👁️	counterfactually repeatable phenomenon	belief

👥 with ⚙️system dynamics/engineering audience

⚙️inventing

infer 🗣️2🧠, 🧠2🤜 (understanding, judging) with ⚙️🗺️, ⚙️📐 (finding path, calibration plot)given 🗄️🗣️🧠🤜 (investor's belief, perception, action) described in Data4DM/decode-venturing@7115373#commitcomment-151790051

[hyunjimoon]

1.🫀 thesis theme

Angie's life goal: living ergodic life
Angie's desire: understanding how to update belief given desire and belief

Angie's thesis explores "exchangeability in entrepreneurship" through the lens of parsimonious persuasion under uncertainty, using simulation-based calibration methods. She grounds her research in three concrete entrepreneurial experiences: her own venture (with direct access to intentionality data), Kyungsoo's MIT construction+AI startup, and either Vikash or Kiyeon's tech ventures. Her work bridges three distinct academic communities: engaging with strategy scholars on theory parsimony and model complexity, economists on experimental design optimization, and statisticians on calibration and exchangeability principles. The research aims to develop tools that transform investor beliefs, perceptions, and actions into calibrated judgments through path-finding algorithms and calibration plots.

Framework Component	Key Elements	Implementation Approach
👁️ Visioning	Parsimonious persuasion under uncertainty	Simulation-based calibration within Bayesian entrepreneurship
🫳 Sense-making	Three cases of entrepreneurs who is intentional and can do probabilistic thinking	Angie, Kyungsoo (MIT MBA (construction + AI), Vikash (IT) or Kiyeon (integrated health care + AI)
👥 Relating	Three academic audiences	TBV/Strategy: Theory parsimony BE/Economists: Experimental design Stats/CS: Calibration methods
⚙️ Inventing	Translation tools	Algorithms for path-finding and calibration plots to map investor beliefs to judgments

summary of #252 (comment)

paper implementations

Venturing into Complexity: Entrepreneurship, Hierarchical Models, and Pockets of Reducibility

Entrepreneurial ventures are critically important drivers of economic growth, yet their outcomes are notably hard to predict. Drawing on insights from computational irreducibility, this paper proposes that the path of any individual startup must be “run in full” before its ultimate success or failure can be truly known. At the same time, we emphasize that not every aspect of entrepreneurship is beyond explanation: “pockets of reducibility” emerge when comparing many ventures using hierarchical modeling. These models highlight how general patterns and partial predictability can be extracted across firms, even if each individual firm’s journey remains irreducibly complex. By exploring these dual realities, we seek to inform both theory and practice, showing how strategic guidance can be developed from cross-firm analysis without resorting to overly deterministic models of new venture creation.

Keywords: Entrepreneurship; Hierarchical Models; Computational Irreducibility; Pockets of Reducibility; Theory Building

Decoding Venture Capital Decisions: A Hierarchical Bayesian and LLM-Augmented Approach

Venture capitalists (VCs) make high-stakes decisions in uncertain environments, relying on limited, observable characteristics of startups to infer unobservable qualities such as team capability and business model validity. Despite the importance of these evaluations, little is known about how investors map observable characteristics onto these perceptual dimensions or how this mapping varies by expertise and investment focus. This paper introduces a novel framework integrating hierarchical Bayesian modeling with large language model (LLM) augmentation to investigate how early-stage investors interpret deal characteristics into structured beliefs and decisions. Through a combination of investor archetype modeling and experimental design, we uncover systematic differences in how VCs evaluate identical startups, revealing how rational meaning construction and lay theory formation drive valuation disparities and decision-making inefficiencies. Our findings contribute to theories of entrepreneurial finance and provide actionable insights for both investors and entrepreneurs.

Keywords: Venture Capital; Hierarchical Bayesian Models; Large Language Models; Investment Evaluation; Decision Making

Q1. if you were me, what'd you be mindful when casting this into transp. Research Paper for exam (ten pages) #252 (comment)?

---

2. evaluation committee

Feasibility Type	Evaluator	Assessment Focus
Technical	Vikash & Mosche	Theoretical correctness of calibration (Vikash) for entrepreneurial choice (Mosche)
Operational	Charlie	Practical deployability of theory
Desirability	Scott	Market fit of theory

• scott and mosche are both aware.
• will share situation with vikash on overlapping class schedule. be listener for his class tmr. his implicit condition of being a committee was taking josh and his class. if vikash said no, perhaps should develop thesis more to invite him as phd committee.

Q2. how to reach out to finalize? can you help me?
Q3. for reading list, may I propose my specific needs? all wanted to help🍀🙏.

3. three classes 2025S

1. demand analysis (Mosche Ben-Akiva)
   1202_s25_Syllabus_V0.pdf

• 12-Mar Use of Instrumental Variables, Simultaneous Equation Models
• 23-Apr Simulation-based Estimation
• 25-Apr Mixture Models is relevant due to de-finetti theorem
• 31-Mar Aggregate Forecasting, Microsimulation, Iterative Proportional Fitting is relevant for segmentation alg.
• 5-May Practical Issues in SP Design and Combining Revealed and Stated Preferences

2.a vikash/josh's 9.S916 (Tues, 5-7p)
2.b josh lerner/shai bernstein's Theoretical and Empirical Perspectives on Entrepreneurship and Innovation (Tues, 5:30-8:30p)

Q4. 👀2🧠understanding on relevance of angie's theme with josh/shai's class #277 (comment)

Q5. 🧠2🤜judgement on which class to be physically present this semester?

first but data for this can be generated using LLM augmentation

1. randomness = complexity = information (measured by entropy) are all subjective

2. Exposing oneself to randomness with intention affects encoding of belief. Mechanism is choice of test quantities at profitable abstraction level.

3. Last chapter of the book 🪙Ten ideas about chance: “coherence of belief constrains inductive skepticism”

[hyunjimoon]

raw material after chat with charlie on tuesday

1. summarizing >50 mail thread so far
   reflection.pdf

• todo: scrap all mails (agenda today, agenda yesterday) and project it on the theory hyperplane formed by chat transcript with charlie and jeff (which were my verbalized desire on understanding belief update given belief and desire)

---

Gap between theory and practice in entrepreneurship is all about non-ergodicity and time.

1. cognitive science (problem: mudane job, averaged out p ~ unif (.5,1))

• One and done "under reasonable assumptions about the time costs of sampling, making many quick but locally suboptimal decisions based on very few samples may be the globally optimal strategy over long periods"

which is not true in entrepreneurship as we deal disequilibrium

2. entrepreneurship

• pace INNOVATION ECOSYSTEMS AND THE PACE OF SUBSTITUTION: RE-EXAMINING TECHNOLOGY S-CURVES

4. agency

• unit of intervention is different from unit of observation (inspection paradox, subjective (due to different desire and belief) probability)

• information is not free (chp.9)

"cannot explain origin of heuristics by considering their contingent utility"

remedies: make current state markovian as much as possible

3. different types of thinking

• falsifiers

computationally irreducible because of contextual complexity. this increases prediction accuracy

1. entrepreneurs are sorting agent with intention
2. ranking algorithm

[hyunjimoon]

This paper introduces a novel framework for entrepreneurial decision-making grounded in probability theory's exchangeability principles. We propose that effective entrepreneurial cognition emerges from two complementary mechanisms: internal exchangeability, which enforces flow balance between understanding and evaluation at the information processing stage, and external exchangeability, which enables pattern detection through experimental feedback. Our model represents this through a tripartite cognitive process where "undigested observations" transform into "unprocessed information" and "unrealized intentions" through dynamically adjusted fractions of projecting, sorting, and designing effort. This dual exchangeability framework provides both theoretical insights into how entrepreneurs maintain cognitive consistency while detecting market patterns, and practical guidance for balancing exploration with exploitation under uncertainty. Through mathematical formalization and empirical validation, we demonstrate how this approach can improve entrepreneurial decision quality while bridging the gap between probability theory and practical venture building.

---

The diagram shows a progression from raw data, through the formation of usable information, to the (still probabilistic) intention awaiting realization. 👀 undigested observation represents unorganized inputs waiting to be interpreted. unprocessed information is the yet‐to‐be‐organized signal containing potential structure. 🫀 unrealized intention is the emerging plan or desire that has not yet been translated into action, capturing the possibility that remains latent until acted upon.
Each "fraction" represents a dynamic allocation of cognitive resources that must be continuously calibrated throughout the entrepreneurial journey. The system maintains two critical balances: an internal flow balance at "unprocessed information" and an external experimental balance through feedback loops. By allocating different proportions of effort to projecting frac (converting undigested observations into understanding), sorting frac (evaluating unprocessed information to shape unrealized intentions), and designing frac (experimenting with unrealized intentions to generate new observations), entrepreneurs create an adaptive learning system. Like MCMC sampling, proper balance ensures both consistent mental models through detailed balance at information processing (internal exchangeability) and effective exploration of market possibilities through experimentation (external exchangeability). This dual balance determines how effectively entrepreneurs construct rational meaning construction before choosing their action.

---

how to measure this? ΔS = -k∫(E - Ec)Θ(t)dt

[hyunjimoon]

committee

i was aiming for some tech push

calibration

1. forming Angie-Charlie prior: p(🧠🫀theta')

• A expressed desire for exchangeability and S's weak verification
• C expressed four step structure to formulate this: 1.Define exchangeability, 2.Survey textbook presentations, 3.Review paper applications, 4.Identify target entrepreneurship problem

2. product-oriented encoding of belief: p(📝y|🧠🫀theta')

• A encoded her belief (200-page PDF) on mapping customer in innovation ecosystem and probabilistic inference technology in C's format

3. apply inferring algorithm (called scott) to 📝y: p_A(🧠🫀theta''|y)

• A verified Angie-Chalire-Scott is well-calibrated as S's inferred (theta'')the motivation of the document (Y) as Angie-Charlie have intended (theta'): "🚨"
• S propsed "🚨🚨" as research question to answer
• This decision-making angle wasn't explicitly written
• The interpretation revealed an implicit motivation in the work
  Understanding how exchangeability in sampling and inference problems extends to entrepreneurial decision-making, particularly in the sequence of testing ideas versus testing strategies

[hyunjimoon]

2017:
discovered poison: agent in complex system can't build perfect model

antidote1: continuous improvement

2024:
discovered poison: agent in complex system can't know evaluation measure

antidote2: cognitive equilibrium (understanding inflow = judging outflow)

[hyunjimoon]

three step process of writing a visionary paper, using auto(experimental choice model cld)

💪(⚙️(➡️ , 📦))capa(process(in, product))

easier for me to think

💪Capability	➡️ IN	⚙️ PROCESS	📦 OUT	📐Module (subsection)
👁️ Visioning (Purpose)	- Keywords and core literature - Research gaps in existing literature - Contradictions in current theories	1. Review literature systematically 2. Map intellectual landscape 3. Identify knowledge gaps 4. Formulate hypotheses from gaps	- Comprehensive literature map - Set of identified research gaps - Testable hypotheses - Direction for investigation	Module 1: PaperExplorer - Literature review - Gap identification - Hypothesis formulation
🤜 Inventing (Process)	- Formulated hypotheses - Methodological approach - Available evidence - Theoretical frameworks	1. Construct logical arguments 2. Test against available evidence 3. Compare theoretical perspectives 4. Analyze strengths and limitations 5. Articulate contribution	- Set of tested arguments - Evidence map connecting to claims - Theoretical comparisons - Articulated unique contributions	Module 2: ArgumentModeler - Argument construction - Evidence testing - Perspective comparison - Strength analysis
🕸️ Sensemaking (Perspective)	- Evaluated arguments - Significance assessments - Broader research context	1. Evaluate competing arguments 2. Assess significance to field 3. Select optimal framing approach 4. Integrate perspectives	- Prioritized arguments - Clear significance statement - Optimal theoretical framing - Coherent perspective	Module 3: PaperSelector (Part 1) - Argument evaluation - Significance assessment - Framing selection
👥 Relating (People)	- Audience needs - Selected framing - Justified significance	1. Create structured writing framework 2. Justify intellectual contribution 3. Adapt to audience expectations	- Complete writing plan - Compelling justification - Reader-oriented structure - Final paper draft	Module 3: PaperSelector (Part 2) - Writing framework - Contribution justification - Paper draft generation

📐(⚙️(➡️ , 📦)) module(process(in, product))

easier for computer to implement

📐Module	💪Capability	⚙️ PROCESS	💻Function	➡️ IN	📦OUT
MODULE 1: PaperExplorer Purpose: Explore literature and form causal hypotheses based on identified research gaps	👁️ Visioning	Surveys existing knowledge to identify purpose and direction	review_literature()	keywords, core_literature	Updated literature_map for further analysis
	👁️ Visioning	Visualizes relationships in the research landscape	map_intellectual_terrain()	literature_map	Enhanced literature map with conceptual structure
	👁️ Visioning	Pinpoints missing pieces or conflicts in existing knowledge	identify_gaps()	literature_map	List of identified_gaps
	👁️ Visioning	Turns gaps into testable statements	formulate_hypotheses()	identified_gaps	List of hypotheses
	👥 Relating	Understands what potential readers/editors expect	assess_audience_needs()	target_audience context	Dictionary of audience_needs
MODULE 2: ArgumentModeler Purpose: Build and test multiple arguments to evaluate hypothesized relationships	🤜 Inventing	Creates a structured logical path from premises to conclusion	construct_argument()	hypothesis, methodology	New argument added to arguments list
	🤜 Inventing	Evaluates how well the argument stands up to real evidence	test_against_evidence()	current argument, available data	Updated evidence_map
	🤜 Inventing	Looks at arguments through different theoretical lenses	compare_theoretical_perspectives()	arguments from multiple angles	theoretical_comparisons structure
	🤜 Inventing	Evaluates logic & coherence; identifies logical leaps	analyze_argument_strengths()	constructed arguments, tested evidence	Internal rating/assessment
	🤜 Inventing	Summarizes how each argument uniquely advances the field	articulate_contribution()	set of refined arguments	contribution_statements array
MODULE 3: PaperSelector Purpose: Select and justify arguments based on significance and audience relevance	🕸️ Sensemaking	Evaluates trade-offs, theoretical/empirical depth, novelty	evaluate_arguments()	completed arguments	evaluated_args dictionary
	🕸️ Sensemaking	Gauges importance (theory, practice, novelty, audience interest)	assess_significance()	evaluated_args	significance_report
	🕸️ Sensemaking	Decides how best to 'frame' the paper's storyline for clarity/impact	select_optimal_framing()	significance_report, broader context	optimal_framing information
	👥 Relating	Outlines sections, assigns argument flow	create_writing_framework()	chosen framing, best arguments	writing_plan structure
	👥 Relating	Provides a rationale for why readers should care	justify_intellectual_contribution()	final framing & significance	Summary statement for introduction & discussion
	Integrated	Combines all elements into a coherent draft text	generate_paper_draft()	All prior module outputs	Complete paper draft

[hyunjimoon]

I will ask Hart for help in making genre of academic paper I'd like to write

My two measure of importance: 1) voluntary diffusion (by me) , 2) chain reaction of diffusion (from others to the other)

Fine22_OM4Entrep.pdf
Gans et al 19_found_es.pdf
GelmanEtal2020-BayesianWorkflow.pdf
Almaatouq24_beyond20.pdf
Vul14_One and Done.pdf

[hyunjimoon]

table2paper2table

using inverse paper to table mapping cld
and restructuring cld