🪵 BayesSD meeting Log

[hyunjimoon]

@tomfid
I will log our trials during cooking meeting here!

For Sep.W4:

• I wonder what each dimensions of prior_pred means (especially 4) 20 is length of timeaxis.
  

• 🎼 Master B major to control F major: simulation of, by, for Box-Flow #25

• If "Using net flows does solve the problem of autocorrelation of the measurements" this is the reason we need to invent a structure in generator, could statistical model structure that estimates parameter assuming first order autocorrelation embedded time series (gaussian process; Mike Betancourt's robust gaussian process estimation) be a viable solution?

[tomfid]

Not sure what I'm looking at here.

[hyunjimoon]

Sorry, I will use @ button more prudently, but the above is prior_pred from this line.

[hyunjimoon]

@tomfid mentioned kicking off cookbook so I drafted the title and its theme "Constructing Bayes x SD joint space". Construction of product measure is in fact based on rectangles.

Definition of Product Distribution Function. If $F_1, \cdots, F_d$ are distribution functions on $\mathbf{R}$, then $F(x) \equiv F\left(x_1, \cdots, x_d\right):=F_1\left(x_1\right) \cdots F_d\left(x_d\right), \quad x \in \mathbf{R}^d$ is a distribution function on $\mathbf{R}^d$, called the product distribution function of $F_1, \cdots, F_d$. If each $F_1, \cdots, F_d$ is a probability distribution function, then so is the product $F: \mathbf{R}^d \rightarrow$ $[0,1]$. For a given distribution function $F: \mathbf{R}^d \rightarrow \mathbf{R}$, we define $\nu((a, b]):=\Delta_{b-a} F(a)$ on the class $\mathcal{G}$ of rectangles of the form $(a, b]$ (From Ioannis karatzas' lecture note).

Finished ContinuousCode folder reconstructing its structure (I wish our book to revolve around codes) and described in
https://github.com/Data4DM/BayesSD/blob/master/ContinuousCode/def(ContinuousCode).md .

[hyunjimoon]

Oct W1,2 @tomfid
- how to implement prior_pred in vensim

• Inventory management casestudy #17 (comment)
• completed draws2data for inventory management model
• introducing python's labeled multidimensionxarray:

# Xarray is multi-dim labeled array which would allow temporal and cross-sectional comparison
# It use `coordinates` and `data variables`; the latter is equivalent to pandas' column name
# arviz (Bayes computation visualization library) used netCDF data structure as it need
# labels (posterior, prior, prior predictive, loglik) for each of the parameters each of which has 1,000 samples

• 📦 Zipping casual loop as lookup or inter-dependent fraction from coflow structure #38 (comment)
• developing our discussion on B-major vs F-major:

which can be found in: marginnote3app://note/DC95299A-D9FF-43F5-A84D-FE143701D782 (may only open in my comp) - hierarchical and dynamic aggregation: ""best cluster" is not necessarily the same as "clustering" - discussion on measuring linger time and wealth for (who-knows-how-much-inherent-heterogeneity) red and blue ppl: "After an exponentially-distributed time they leave. This is basically the classic first-order well mixed system". As long as we measure quantities red and blue differs (f(theta1, theta2); not f(theta)) it derails from poisson process with exponential delay and becomes compound and/or conditional poisson process. - three diagrams with four questions; strategy https://github.com//discussions/7#discussioncomment-3859648 - https://github.com//discussions/21#discussioncomment-3834648

[hyunjimoon]

@tomfid the above is from one of Nathaniel's lecture which could be helpful initial point of the lingo mapping table. I started in wiki but I became very certain in one day that wiki may not be a good place (doesn't have access in mobile app).

[tomfid]

I like it! I'd like to build the same thing for SD/Bayes.

[hyunjimoon]

10/21 agenda @tomfid

• stanify demo
• when do yo use timeshift in vensim?
• revisiting Tom's comment on poisson process (daily customer)
• could cumtime function in stanify make timestep dynamic?

[hyunjimoon]

Oct W4

1. I am certain I wish to do research on win-win strategy for platform business (Ed and Johan) using System data science (Hazhir and Tom).
2. how to break HMC but not break other samplers, degeneracy (colinearity, hierarchy) bad parameterization example in SBC
3. process noise as part of data structure? Stan cannot solve SDE; record in Can we include variation from noise into variation from prior? #70
4. advise on paper or literature; structure Jair suggested but I don't wish to constrain this to readsdr as it lack auto-translation
5. vensim plots; Tom has used seaborn

• existing plotting code?
• issue opening Vengine on Windows. We are getting this error: “reinstall vensim (error 977)”
• resolved by older version of Vengine, and so I wondered if this is an issue with the newer version of Vengine.

---

Detail:
1.
Tool-wise: Stan

• simulation-based checks: attached paper allows real data and test-quantity (e.g. lp__) consideration in SBC
• stanify: couples managerial with coupled statistics and computation which Stan-enabled. draws2data2draws directly applies my SBC paper above. Three stanfiles translated from inventory management dynamics diagram below (diff.equations as function block here) and I wish to push this to hierarchical settings. Compared to stan-team effort above, I am lonely doing this alone!

Target-wise: affordable analytics
will meet Geoffrey Parker, author of platform revolution, next week to propose how platform can prevent inequality from non-affordable analytics problem.

Problem:
- What justifies this paper?
- Who's the target (audience) of this paper?
- Benefits/drawbacks of Stan & Vensim/Stan

Hypothesis:
How readsdr could address such a problem?
Efficiency
Mitigation of errors

Analysis:
How many case studies & which?
Describe Bayesian workflow, where each step is supported by a function from readsdr
Model building
Synthetic data?
Exploratory optimisation?
Prior predictive checking
Sampling (Hamiltonian Monte Carlo & other algorithm?)
Posterior predictive checking & other expectations
Simulation-based calibration

Donella's wisdom:

Places to Intervene in a System (in increasing order of effectiveness)

12. Numbers: Constants and parameters such as subsidies, taxes, and standards

11. Buffers: The sizes of stabilizing stocks relative to their flows

10. Stock-and-Flow Structures: Physical systems and their nodes of intersection

9. Delays: The lengths of time relative to the rates of system changes

8. Balancing Feedback Loops: The strength of the feedbacks relative to the impacts they are trying to correct

7. Reinforcing Feedback Loops: The strength of the gain of driving loops

6. Information Flows: The structure of who does and does not have access to information

5. Rules: Incentives, punishments, constraints

4. Self-Organization: The power to add, change, or evolve system structure -> self-governance

3. Goals: The purpose of the system

2. Paradigms: The mind-set out of which the system—its goals, structure, rules, delays, parameters—arises

1. Transcending Paradigms

Guidelines for Living in a World of Systems

1. Get the beat of the system?

2. Expose your mental models to the light of day?

3. Honor, respect, and distribute information. 

4. Use language with care and enrich it with systems concepts.

5. Pay attention to what is important, not just what is quantifiable.

6. Make feedback policies for feedback systems.

7. Go for the good of the whole.

13. Defy the disciplines.

14. Expand the boundary of caring.

15. Don’t erode the goal of goodness.

Geoffrey's wisdom from platform revolution -> design data-sharing api (collaborative sharing e.g. maintenance)

## 6. monetization (capturing value created from network effects)
- well-managed platform create excess value in four ways: access to value creation, access to the market, access to
tools, and curation. Monetization is about capturing a portion of the excess value created.
- Techniques for monetizing a platform include charging a transaction fee, charging users for enhanced access, charging
third-party producers for access to a community, and charging a subscription fee for enhanced curation.
- One of the most crucial monetization choices is deciding whom to charge, since the difference in roles played by various
platform users means that charging them can have widely differing network effects.
- Given the complexity of the monetization challenge, platform managers should take potential monetization strategies into
account in every decision they make regarding platform design.

## 8. self-governance: increase value and enhance growth
- prevent and mitigate market failures caused by information asymmetry, externalities, monopoly power, and risk
- The basic tools for platform governance include laws, norms, architecture, and markets. 
- Each must be designed and implemented with care in order to encourage platform participants to engage in 
positive behaviors, incentivize good interactions, and discourage bad interactions.
- Self-governance is also crucial to effective platform management. Well-run platforms govern their own activities following the principles of transparency and participation.

[hyunjimoon]

I wish to use this Fri 3pm meeting as one milestone to couple our efforts; this blueprint is one initial point :)

The goal is to seek optimization + visualization support for dynamic modelers (especially pure Vensim users) who I think are the greatest market segment (@tomfid is there any statistics on user segmentation of each tool? for instance, from the following three cases, what would be the ratio of each three? + how much % of people are we leaving out?)

To be very careful but honest, removing the use of optimization algorithm without diagnosis is one option. I've heard good comments on Stan's design decision of not offering Metropolis Hastings option.

@jandraor could you please prepare the summary of rvar doc by Matthew (and SD-connection) as we discussed last week?

I am meeting with two Stan devs before - will share “arviz’s wrapper or multi-index for posterior group can streamline 2,3” process on Fri.

• meaning of from Multimodality reported in hierarchical prey-predator #89
  "I also added diffusion among the regions. This is potentially pretty interesting, because it means there are two ways to get similar behavior among the 4 regions:
  low variety in the vital rate parameters, and similar initial conditions, so the regions are on similar but independent trajectories
  significant migration among the regions - it doesn't take much to couple the regional trajectories fairly strongly
  Hopefully this creates an added opportunity for multimodality, where the variety and migration parameters are confounded."

[hyunjimoon]

#95

[tomfid]

A couple additional items from my view:

• getting Stan versions of pred/prey with 4-6+ parameters working
• priorities for further Vensim models

[hyunjimoon]

For NovW4

• Formulation of parameter upperbound #108

[hyunjimoon]

@tomfid
I wrote down the recipe for our seasonal cuisine. If you get the output of any of the following steps before next Thursday, please do share! It would be so meaningful to share this NEW data structure (using only vensim!) with seminar students. No pressure though, but I am tempting you with the beautiful plot for eight school model as below :)

step 1. vensim -> inferencedata structure .nc

• organize vensim's Multimodality reported in hierarchical prey-predator #89 output into .nc
• arviz example code from dict or dataframe to .nc
• .nc stanify produced

step 2. hierarchical in vensim -> inferencedata structure (adding coordinate):

• once we reach inferencedata data format, you could directly run .plot_trace() to get the following (e.g. for eight-school hierarchical model):

# test code for plot
import arviz as az
idata = az.load_arviz_data("centered_eight")
az.plot_trace(idata, var_names = "theta", coords = {"school": ["Choate", "Deerfield"],}, kind = "rank_bars", figsize= (20, 15), rank_kwargs={"bar_kwargs": {"alpha": 0.2}})

step 3. sbc_inferencedata structure sbc_nc

From step3, sbc and its plot are state-of-the-art, so let's write history together! Since this is new for arviz as well, we (arviz developers) are focused on designing sbc plots for the hierarchical model.

• arviz example code from stan output to .sbc_nc
• .sbc_nc stanify produced

[hyunjimoon]

updated the location with stanify's .nc!

[hyunjimoon]

12W1 agenda (Tom, Jair, Angie)

• three questions from scale, prevent outlier, remove process noise for speedup #124
• visualization with rvar, sbc @jandraor

[hyunjimoon]

12W3 (Tom, Jair, Angie)

• 2. Typify parameters in syntax-semantic with colors, units, bounds #7 (reply in thread)
• plot design for the hierarchical model, which variable to compare other than sigma (https://discourse.mc-stan.org/t/best-practices-for-simulation-based-calibration-with-hierarchical-models/19630)
• https://github.com/stefanradev93/BayesFlow/blob/master/docs/source/tutorial_notebooks/Linear%20ODE%20system.ipynb
• spatial and temporal hierarchy #120 + Pinocchio productization project here with better rendering here

• can you share model from inventory and hybrid modeling in Getting Ed's agreement on need for sparsity-inducing prior #88

[hyunjimoon]

• hazhir's policy optimization alternating mode example
• a model that compares top-down vs bottom up regulation for angel and devil system(naming on Reinforce and Balance-intervention?)
• trust_factory

[hyunjimoon]

• connecting stan function with enroads (possible)?

[hyunjimoon]

• difference btw capacity and capital from 🎼 Master B major to control F major: simulation of, by, for Box-Flow #25 (comment)

[hyunjimoon]

with Tom and Jair

• can spatial hierarchy in spatial and temporal hierarchy #120 treated as multi-echelon setting with I) backlogs and ii) order-fulfillment logic? can I, ii live without the other?

[hyunjimoon]

Dec20 Angie, Hazhir
a. process noise's effect on inventory model

• two - three estimated parameters is boundary for the nonhierarchical model
• I am using time_saveper = .5, but DT is not passed to stancode.

rather than using (flow-flowStock)/tau with tau=DT, perhaps it would make sense to use tau = something small, but fixed and >2*DT. That way if the RK45-auto routine is slicing DT into smaller intervals, the integration will be stable.

• stocked flow formulation: with process noise scale 10, we sometimes get negative stocked flow.

more experiments in here

b. fire-fighting covid model: hyunjimoon/VaccineMisinf#4 (comment)

• partial model calibration? + convention to ensure internal consistency?
• remedies to extend boundary when hit by one (~ may be solvable by non-uniform prior; something as increasing estimated parameter numbers just to fit time-series)

[hyunjimoon]

posterior:
JanW1 todo

1. plot process noise (and show it is auto-correlated enough)
2. capping process noise to 2 to prevent negative stock; this prevents negativity leaking to flow which explains negative stocked flow

DecW3 summary

1. process noise is being injected every saveper (as saveper now replace time_step) as it should be
2. kill stuck chains with an extremely low variance if loglik is smaller compared to the correct majority 
3. removing stuck chains would improve pair plots
4. one layer for hyperprior should be added to both forward and backward model. Using two parameter (location and scale) distribution. i.e. a. to b.
   a. theta_r ~ normal(mu_loc, mu_loc*.1) mu_loc = .25,
   b. theta_r ~ normal(loc_theta, scale_theta), scale_theta ~ exp(1), mu_loc = .25
   c. theta_r ~ normal(loc_theta, scale_theta), loc_theta ~ exp(mu_loc), scale_theta ~ exp(1), mu_loc = .25

[hyunjimoon]

review:

1. move away from max, min formulation, prevent negative stock value
   

2. lower plot process noise scale, but it increase (.1 -> .3:

pn_scale= 1 gives divergence, decreasing stocked_

3. plot obs
   

[hyunjimoon]

Hazhir Angie 1W1

1. objective function log posterior (accuracy)

# add loglik for "matter ~ form"
1. parameter_draw ~ prior_func(hyper_param)
X. target_sim = f(parameter_draw)
2. observed_data ~ likelihood_func(target_sim)

Q1. how to make use of components of logposterior? e.g. how can modelers improve their model, if they found:
i) contraction: likelihood_func is more informative, but consistent with prior_func
loglik_s + loglik_ss << loglik_prior

ii) containment:prior_func is more informative, but consistent with likelihood_func
loglik_s + loglik_ss >> loglik_prior

iii) compromise: tension btw information encoded in 'prior_func' and 'likliehood_func'
loglik_s + loglik_ss ~ loglik_prior

generated quantities{
    // Define observed vector (matching vector)
    array[N] vector[R] s_obs_post;
    array[N] vector[R] ss_obs_post;
    // Assign generated value to observed vector (matching vector)
    for (r in 1:R) s_obs_post[:, r]=lognormal_rng(log(s[:,r]),m_noise_scale);
    for (r in 1:R) ss_obs_post[:, r]=lognormal_rng(log(ss[:,r]),m_noise_scale);

    real loglik;
    real loglik_prior;
    real loglik_s;
    real loglik_ss;
    for (r in 1:R) loglik+=normal_lpdf(adj_frac1[r]|adj_frac1[r],0.025);
    for (r in 1:R) loglik_prior+=normal_lpdf(adj_frac1[r]|adj_frac1[r],0.025);
    for (r in 1:R) loglik+=lognormal_lpdf(s_obs[:, r]|log(s[:,r]),m_noise_scale);
    for (r in 1:R) loglik_s+=lognormal_lpdf(s_obs[:, r]|log(s[:,r]),m_noise_scale);
    for (r in 1:R) loglik+=lognormal_lpdf(ss_obs[:, r]|log(ss[:,r]),m_noise_scale);
    for (r in 1:R) loglik_ss+=lognormal_lpdf(ss_obs[:, r]|log(ss[:,r]),m_noise_scale);
}

2. optimization algorithm (speed)

• sharing metric and init
  

3. effect of process noise during forward, backward

Q. what hypothesis interests you? can it be a paper?

• closer the estimated parameter is to the process noise source (e.g. production adjustment time is closer to process noise than target delivery delay) uglier parameter recovery becomes. if we can construct a controlled setting well, this sounds interesting
• higher the multiplicative process noise is (e.g. process noise scale .1 vs 3), the uglier parameter recovery becomes
• compare results with process noise structure on and off (i.e. scale =0) hypothesis?

4. Could non-linearity of generator possibly affect rank uniformity in SBC?

5. literature review on simulation-based inference

Q. plots?

i) rejection sampling based on an analytic solution which gives nonuniform joint prior as attached, from their notebook estimation for ode notebook
Q. could we use this strategy? could we have prior knowledge on forbidden areas, and can we get access to this wit sda?

ii) plot design by prior location and scale, noise fraction as attached from their paper Detecting Model Misspecification in Amortized Bayesian Inference with Neural Networks

iii) feel for magnitude on prior location (5), scale (2.5), noise fraction (.5)
scale is half the location and noise is 1/10th (difference between scale and noise?)

[hyunjimoon]

Bayes-self-consistency: prior-data-posterior = posterior-data-prior

detail: https://discourse.mc-stan.org/t/need-help-on-sbc-uniform-proof/17777/14

[hyunjimoon]

Different opinion on adding measurement noise for posterior predictive check in synthetic data context + source of process noise.

Regarding our discussion on adding measurement noise to target simulated from the parameters' estimated value, I think adding sampling uncertainty might be what we want for our purpose: to compare with observed data to judge estimation quality. Here observed data is simulated in our case. wi measurement noise, we are ignoring sampling uncertainty from below. Detail in https://mc-stan.org/docs/stan-users-guide/sampling-from-the-posterior-predictive-distribution.html

Regarding what I asked you last Fall, "can process noise correspond to predictor uncertainty?" (which you answered, it might), "posterior predictive sampling for posterior predictive checks is different from usual posterior predictive sampling discussed in the chapter on posterior predictions in that the original predictors x  are used." from this official document following supports this. I know it is confusing... but thought you might be interested as you thankfully emphasize the importance of different uncertainty sources in Heterogeneity and Network Structure in the Dynamics of Diffusion: Comparing Agent-Based and Differential Equation Models paper.

[hyunjimoon]

Tom, Jair, Angie 1W1

• 🪵 BayesSD meeting Log #23 (reply in thread)
• progress on 'degenerating joker`: obsidian-rendered and code
• Michael's advice:

As for the "stuck" Markov chains you have to differentiate between bad adaptation/exploration and actual multimodality. The example used in https://github.com/betanalpha/mcmc_diagnostics demonstrates how one can use useful MCMC diagnostics to differentiate between the possibilities.

• @jandraor does sir model have example of stuck chain and can we write a story of "differentiating bad adaptation/exploration and actual multimodality" affecting policy?
• attacked uniformity and evolving policy

[hyunjimoon]

Tom Angie 0210

• Q from Tom's chronic waste disease talk #140

• multi time scale #141

• end of feb. notebook for andrew's blog

[hyunjimoon]

Q. is it correct that smoothing customer order to expected customer order doesn't prevent the amplification (inflow surge from exogenous surge)?

instead of shipping -> selling, grocercy store; expected deamdn goes up desired inventory goes up -> inventory goes up (housing market eg: constrcution time is long - build more house

[hyunjimoon]

HW: reformulate the purchase decision to accelerate the transition, and analyze the resulting overshoot in purchases.

ev_stockfrac.mdl.zip

[hyunjimoon]

Q3. simulat. loop which I thought could be handled in vensim causes error

[hyunjimoon]

Q4. connecting SEIR in EV context to bridge Jair's estimation workflow with policy workflow

key q:
4.1. does SEIR (not sir) have an interpretation for tech adoption? i.e. how to formulate as close as possible with SEIR (what are epidemic-specific assumptions, let alone the different goals of policy to increase vs decrease infected)
4.2. what to estimate in SEIR EV adoption model (market)

4.3. feedback on market_ev_v1:
--potential fleet buyer as ice user's initial value?
4.4. how to fill the gap btw market and evc_supply model? i.e. how to translate market.buying_ev into evc_supply.customer_order? (detail in #142)

Q5. how simulator can contribute and when? https://www.cartsmobility.com/blog/trenton-moves-interactive-graphic-visualization-new

market_ev_v1.mdl.zip

Jair's notebook:
Archive.zip

[hyunjimoon]

summary of Tom's comment

• there was a paper /observation study that instantaneous info sharing during beer game didn't help much; we concluded there were many other factors (perverse, low q info) that hindered the great use of this info
• john's paper on organization learning includes E stage (backlog of ppl waiting for the product)
• classic SEIR has: S is ICE user, E is ppl waiting for EV, ADV (market saturation) happens in SE, WOM happens in E-SE and I-SE.
• For Ev, EV haver is connected to ordering (=SE) but this is equal to setting EV haver as I. Remaining problem is what to set/name R.
• capacity is fully utilized (advertising is useless bc S << E)
• EV-EVC handshake failure would not be logged in EV, but in EVC
• who would have the record of balking effect (EV has long lead time and high price so ppl don't order)

• to create overshoot people should buy vehicle more than needed which is possible by the increased slope of market share or size which can be represented as stocked flow below

1. increase market share:

• make ICE harder to use (price or infrastructure)

2. increase market size: grow the total fleet)

• which unit (user vs vehicle) to use?: user with the assumption one user has one vehicle; if the unit becomes household, 1~4 for each household

• higher order has lower deviation

[hyunjimoon]

@jandraor and Angie 0217
For #76, @Dashadower and Angie prepared https://gist.github.com/Dashadower/face0ff696d83a13d0ea1624b70c7917, and below are some agenda for the meeting with Jair:

• feedback on the notebook
• final format for the notebook that we'll share with Andrew
  -- how to keep format differences (from different authors) less visible
• is due Feb. feasible (due date) or should we reschedule? when?

20 policies (mask, anti-virus, search strategies, vaccine (supply chain)) each of them are not mutually exclusive and have dependence
2: testing and isolation, testing capacity (deploy it + building up capacity)
4: contact tracing (who did you have contact with?, quarantine - ) + preparedness, mask, contact tracing)
social distance
incorporating process noise -> CTMC is easier than compartment model (2003, Rohani zotero; The Interplay between Determinism and Stochasticity in Childhood Diseases)

[hyunjimoon]

common purpose: smaller loop affect the bigger loop

1. add sbc to workflow

• determine the degree of hierarchy
• estimation (ncp, cp)
  

2. actionable item

• 20 policies (propagate uncertainty on policies)
• EV vs ICE

[hyunjimoon]

Jair' model is SEIR (modulate people to go from I to R)

• we agreed blog feedback can increase the audience of readsdr paper and agreed by next week:
  J. jair: show how to overcome non-identifiability: 1-2-3- nonuniform sbc (degeneracy would be captured) - 1-2-3 uniform sbc
  A. angie: based on uniform sbc - EV context

plans for A:

- SEIR: potential adopters: `S`, wom: `I` are telling `S` (internal demands), `I` infected for a while you move to `R` 
what are epidemiology-specifc assumptions? among the assumption Jair is making (infection to go up vs down)

key q:
1. does seir (not sir) have an interpretation for tech adoption?
2. how to fill the gap btw `market` and `evc_supply` model? i.e. how to translate `market.buying_ev` to `evc_supply.customer_order`?
3.  with evc supply chain? (`market.buying ev` becomes `evc_supply.customer order`(charging station)
4. feedback on `market_ev_v1`: 
--`potential fleet buyer` as `ice user`'s initial value?

[hyunjimoon]

tom, angie

classification of people (atomized bit)

• a person is an atomized bit, modularized thoughts that are mostly consistent
• I decipher signals (e.g. writings and voice) to learn their bit/soul
• based on the following label of person's function into one of the four quadrants
• each person I label people's function into one of the four quadrants

bit is flow, atom is stock, Q. role of energy to fasten bit2atom?
example of bit: technology, less situated, higher rate of success for observer
example of atom: asset, more situated

1. label $bit_t$

BE stands for bayesian entrepreneur, PC stands for probabilistic computing

past\future	observer 👀	actor 🤜
effect of feature	describe [[def(BE_⬇️)]], [[def(PC_⬇️)]]	explain [[def(BE_⬆️)]]
prediction	predict [[def(PC_⬆️)]]	synthesize [[def(PC_🔃)]]

• ⬆️, ⬇️, 🔃 is ability to act, think, think-then-act
• 🔃 requires time stone, which increases act-think clockspeed through simulation (detail in [[W11_Stones and Gauntlet]])
• [[def(PC_⬇️)]] desires [[def(PC_🔃)]] and with easy-to-use tool, persuasion success rate would be over 50%

2. label $atom_t(bit_t)$

• [[def(BE_⬇️)]] is observer or actor using non-computerized human language (non-scalable approximation) to explain certain effect of feature

• [[def(BE_⬆️)]] is actor choice of adapting before theorizing (industry mentors)

• [[def(PC_⬇️)]] is observer choice of using measurement, econometrics, causal inference to explain effect of feature

• [[def(PC_🔃)]] s

3. persuade $atom_t(bit_t) \rightarrow atom_{t+1}(bit_{t})$

• order fastest to slowest clockspeed action and plan out change

4. persuade $atom_{t+1}(bit_t) \rightarrow atom_{t+1}(bit_{t+1})$

---

5. label $bit_{t+1}$

social science testing with genome

• https://github.com/orgs/Data4DM/projects/3/views/2?pane=issue&itemId=57482640
• marginnote3app://note/5F80E6CE-B741-430D-8B66-149DA0D1C8E7
• bayesdb (IS THERE STH truly causal? my hypothesis example) in knowledge production system

transfer learning and hierarchical bayes in transportation

• predicting startup growth with hierarchical gaussian process #200
• spatial and temporal hierarchy #120
• connection with ford and boland paper on fast-slow dynamics modeling
  Bloand17_cont_servnet_design.pdf
  Ford18_fast_slow_sim.pdf

[hyunjimoon]

gal, angie

pre-meeting

given our goal dynamic canvas.pdf
q.

• how to intuitively deliver three axis (direction, speed, accelaration) to startups?
  

post-meeting

[hyunjimoon]

miaomiao, angie

prior

1. AM's prior that's in the best form to be evolved to SMA prior (S:Scott, M:Miaomiao, A:Angie)
2. syncing A and M's prior on labeling for RAG

just in general, supply side (feasibility) - A - M - S - demand side (desirability)

likelihood

In the first part of the conversation, Speaker 1 and Speaker 2 discussed entropy reduction mechanisms in GPT prompts, with Speaker 2 emphasizing the importance of experimentation to find a product-market fit. In the second part, Speakers 1 and 2 discussed their concerns about a research project they are working on with Scott, weighing potential risks against rewards. Speaker 1 proposed an alternative idea and sought feedback on their other ideas, including the role of VCs and the difference between entrepreneurs and founders.

Transcript

https://otter.ai/u/s4_sbCL4NR1s-rfnM5JVf1RtZXU?view=transcript

Action Items
[ ] Have another meeting with Professor Scott to provide a final update before potentially discontinuing the project.
[ ] Prepare to simulate their conversation with Professor Scott to determine the best sequence and topics.
[ ] Speaker 2 to present their initial simulation model on type 1 vs type 2 errors in belief updating at the next meeting.
[ ] Schedule regular monthly meetings to continue discussions and work towards presenting at a future conference.
Outline
Entropy reduction mechanisms for information theory.
Unknown Speaker discusses entropy reduction mechanism and information theory, seeking to integrate ideas from different sources.
Speaker struggles to connect ideas and find a cohesive theoretical framework, awaiting feedback from Chris.
Improving AI models' noise tolerance through experimentation.
Unknown Speaker discusses experiment on adding noise to GPT prompts, hoping to observe differences in response between solo and dual groups.
Speaker wonders if tighter noise distribution in solo group might lead to less varied responses, while dual group's ability to detect signal from noise could be advantageous.
Experiment design and theory validation.
Unknown Speaker discusses experiment and solo, confusion between theory and experiment.
Speaker questions why solo is titled solo, and if experiment can show tighter or getting tighter.
Model development and uncertainty, with disagreement on approach.
Unknown Speaker expresses uncertainty about proposed model, disagrees with Scott's wording.
Modeling project with conflicting opinions.
Unknown Speaker discusses potential risks and rewards of a project with Scott, expressing uncertainty about the project's direction.
Speaker mentions Daniel's strengths in organizational model building, but acknowledges their own limitations in this area.
Experimentation and entrepreneurship with a focus on modeling and entropy.
Unknown Speaker discusses ideas for experiment with Jackie, prioritizing their needs.
Speaker seeks feedback on theoretical angle, experimental setup from Scott.
A project with a colleague, including potential next steps and concerns about detachment from reality.
Unknown Speaker suggests meeting with Scott to discuss project and prioritize tasks (0:17:52)
Unknown Speaker worries about detaching from reality if project doesn't move forward (0:19:40)
Entrepreneurship, quality, and risk preferences.
Unknown Speaker discusses entrepreneurship, quality, and belief updating process.
Speaker wants feedback on theory about entrepreneurial firms and risk preference.
Simulation project, type one and type two models, and confidence.
Unknown Speaker discusses simulation project with colleague, focusing on type one and type two models.
Speaker seeks colleague's input on how to present research at an entrepreneurship conference.
Product and market changes based on experimentation and data analysis.
Unknown Speaker discusses solo and duel actions in product-market fit, with solo actions focusing on changing the product and duel actions challenging the model itself.
Speaker suggests that duel actions may be more effective in questioning the market, as individuals with more experience in different markets may be more willing to do so.
Experiment progress and goals for a Bayesian conference.
Unknown Speaker discusses experiment goals and progress with Scott, including sharing their goal for presenting at a conference.
Unknown Speaker and Scott discuss the importance of clear communication and setting aside dedicated time for their conversation.

posterior

schedule for 30min meeting:

1. M2’s progress so far

• table's type1,2
• hard & not high expertise and interest in econ. modeling
• interest in type1-2
  

2. A's type1-2
3. next step

[hyunjimoon]

@tomfid, nathaniel, angie
90 mins on developing theory and planning to diffuse

1 When you were (are) applying Bayesian methods and system dynamics modeling in health policy, what were the key challenges and insights?
2 How did you communicate complex Bayes + SD models to decision-makers in health policy? What strategies were most effective?

3 Can your experiences in health policy modeling be transferred to modeling entrepreneurial learning? What similarities and differences do you see between these domains?

4 Given the issues in entrepreneurial learning literature (lack of formality in units and types), do you see path where we can join forces to build theory/tool for Bayes + SD? What could be a potential production and diffusion plan for this theory?

5 How can the SD+Bayes fields come together to make progress? Are there key technologies that would help? Could probabilistic program(ming) make this easier? (for example)

• below two tables and feedback I received from 13 management school scholars are context for my question for how to argue hierarchy, bayesian, dynamics are three sides of one dice🎲?
• paths to community building (Tom, Hazhir, Jair, Angie) to diffuse diffusing bayesSD💡
• Tom expressed interest in starting bayes interest group. nate expressed interest in second edition of dynamic modeler book. i'd like to get your advice in pooling my (hopefully our) resources to address different needs
• bayes resource guide I'm making

angie's situation

Step	Current State in Entrepreneurial Learning	Potential Bayes + SD Approach
Problem	Lack of formality in units and types in entrepreneurial learning literature
Root Cause	Inconsistent language (pivot, causal, experiment, persuade) and unclear definitions (e.g., scaling as outcome vs. process)
Proposed Solution	Apply Bayes + SD modeling to formalize units and types in entrepreneurial learning
How Solution Addresses Root Cause	Bayes + SD can provide a rigorous framework for defining units (idea/belief-individual/firm-industry/ecosystem) and analyzing objects and functions of entrepreneurial learning
Production Plan	knowledge production table below
Diffusion Plan	communities to bridge in management (bayesian entrepreneurship (scott stern), bayes in business), in cognitive science (probabilistic program: josh tenanbaum, vikash masinghka), bayesian statistics (andrew gelman)

knowledge production plan and our potential collaboration

knowledge production phase	Questions
Generate	When you were applying Bayesian methods and system dynamics modeling in health policy, what were the key challenges and insights? Could probabilistic program make this easier?
Synthesize	Can your experiences in health policy modeling be transferred to modeling entrepreneurial learning? What similarities and differences do you see between these domains?
Educate	How did you communicate complex Bayes + SD models to decision-makers in health policy? What strategies were most effective?
Apply	Given the issues in entrepreneurial learning literature (lack of formality in units and types), do you see path where we can join forces to build theory/tool for Bayes + SD? What could be a potential production and diffusion plan for this theory?

[tomfid]

My summary of the "problem" and "root cause" under the heading "Angie's situation" is that entrepreneurial strategy, like a lot of fields, lacks formal models. This leads to all the usual problems: variables are fuzzy constructs, data gets used to evaluate isolated nodes in the system rather than updating an integrated understanding, and statements about structure can be used to rationalize a variety of outcomes post hoc. This means that theory is used as propaganda to justify actions that are desired on self-serving, seat-of-the-pants or superstitious grounds.

[tomfid]

It's hard to learn in an environment like this, with weak theory, because you don't have a structure to update when you get data. That mainly leaves population learning: evolutionary forces kill 90 out of 100 firms, and we look at the 10 survivors and try to figure out what worked so we can imitate it in the next generation. But what worked might be substantially random.

[tomfid]

Possibly interesting aside: the evolutionary learning problem might be a good application for LLMs because they can assimilate a lot of data about the survivors and find patterns in a reasonably consistent way. But since the corpus on which they were trained lacks a coherent model, there's no reason to expect them to have one.

[tomfid]

I like the generate-synthesize-educate-apply sequence.

I think the 'educate' component is big. Communicating complex Bayesian/dynamic models is hard, especially because it takes time to build them, and entrepreneurs are in a very time-constrained environment (maybe a bit like public health officials in March 2020). One response might be to get more of the education piece done up front, so that entrepreneurs are complexity-literate, and you only have to teach them the immediate model and data, not the whole of decision making under uncertainty.

[hyunjimoon]

jb, angie (otter)

tldr; angie digested 1,2 information as knowledge, and am producing BST and pivot-game. thanks jb for explicitly saying helping me finish phd is your primary goal. when i could explicitly include exaptation was uncertain so i'm doing my best to make our thought-exchanging process itself helpful to you (🐟🪷symbiosis)

1. mapping existing entrep.learning approaches with machine.learning learning tribes

Bayesian (scientific), behavior (effectuation, causal (not exactly same causal in stats.), Evolutionary (@jeanbaptiste do you agree this is most relevant to design from below table? i hope difference between scientific vs design can address @chasfine 's question on bayes vs evol. learning)

below table was presented in https://cdmcd.co/JQJLvk session

Epistemological logic	Key assumptions
Effectual	Opportunities are subjective ideas, entrepreneurs navigate available resources to exploit perceived opportunities, and risk aversion prompts experimentation in small, manageable steps (Sarasvathy, 2001)
Causal	Opportunities exist objectively and independently of entrepreneurial efforts. Entrepreneurs are alert discoverers of opportunities who formulate theories about novel opportunities based on observations, experiences, and perceptions (Kirtzner, 1997; Shane, 2003).
Scientific	Entrepreneurs develop beliefs and hypotheses to experiment under uncertainty - A pragmatist perspective, asserting that entrepreneurs do not engage in wide-range environmental scanning but focus on identifying and solving market problems through a bottom-up abductive thought process (Camuffo et al., 2020; Felin et al., 2019; Zellweger and Zenger, 2022).
Design	Entrepreneurship is an artificial phenomenon, designed to achieve particular goals in light of environmental circumstances, a "gateway to a different, not yet transpired future" (Berglund et al., 2020; Dimov, 2021, p.7) - an agentic perspective that suggests opportunities are not objective entities waiting to be discovered; entrepreneurial agents proactively construct these opportunities through a creative and deliberate process.

• Discussed three levels of pivots Curriculum design with MIT ENT Educators #158 (reply in thread)
• angie classified svafa's approach for presenters adjusting to investor's sentiment as behavior learning

2. feedback from aom confernece

• angie found three sessions from AOM conference most helpful (all included theory/scientific school from utah 🏔️& bocconi🇮🇹)

New Perspectives on Experimental Strategy

experimentation is, fundamentally, a strategic decision in . This work has directed attention to several critical choices actors face when crafting an experimental strategy including the design of the experiment, what to do after running the experiment, and the role of stakeholders, theories, the nature of the decision, and the quality of the idea. In this session, we aim to present N leading scholars who are working in this domain, allowing each to share emerging insights from their particular point of view on the topic; conduct a dialogue with the participants as a collective; and field questions from attendees. The purpose of the presentation will be to showcase areas of active research and illuminate opportunities for future work, in particular the elaboration of new theory and how current theory could be further developed through empirical insight. We anticipate that the first part of our panel discussion will be driven primarily through structured questions that encourage dialogue and constructive engagement between the panelists, orchestrated by the organizers. The overarching goal is to use the session to support the development of participant papers and to promote intellectual interchange on the topic of experimental strategy from the most nascent of entrepreneurs to the most established large firms.

Value Creation, Value Capture, and Strategic Theories of Digitally Transformed Firms

To foster a thriving community of scholars dedicated to investigating digital businesses through the lens of classic core strategy questions, we have organized series of popular PDWs at recent AOM conferences. These workshops have consistently garnered substantial attention, each boasting an audience of over 200 engaged participants. Our past PDWs have delved into diverse themes, covering corporate strategy in 2020, organizational governance in 2021, resources and capabilities in 2022, and fundamental issues in strategy in 2023. As we embark on our fifth installment in 2024, our focus will be on the theme "Value creation, value capture, and strategic theories of digitally transformed firms." This theme encompasses a broad spectrum, including digital ecosystems, platforms, artificial intelligence, and related phenomena. Our objective for this PDW is twofold. We will focus on both how digitization broadens the application of existing theories, and how its novel features require changes to the current theory.

Entrepreneurial Experimentation: Processes, Logic, and Future Research

This session brings together leading scholars to explore the evolving landscape of entrepreneurial experimentation amid digital transformation and global crises. It aims to deepen understanding of the processes, logics, outcomes and future research agenda of entrepreneurial experimentation, focusing on micro, meso, and macro levels. The panel discussions will highlight different experimentation logics—effectual, causal, scientific, and design—each providing unique insights into opportunity design and discovery, and the related theoretical perspectives. The symposium will address challenges and opportunities in implementing these logics in uncertain contexts, contributing to a richer discourse on navigating through uncertainty with informed experimentation strategies. The panel will also foster discussions on different approaches to entrepreneurial experimentation, the distinct logics guiding these approaches, and the challenges and opportunities they present in uncertain contexts. Our panelists will contribute a deeper understanding and broader discourse on the evolving landscape of entrepreneurial experimentation, its implications for business pivots, and the art of navigating through a fog of uncertainty. The panel discussion will be forward-looking and exploratory, and we intend the panel symposium to result in an outline for future research in this important area.

• Thomas Zellwegner's (img above) Airbnb's business model as an optimization problem with variables for supply, product, demand
• rational analysis and modeling for decision optimization by Tom Griffiths (rational use of cognitive resource, abstract computational problem that a system has to solve and explain its behavior in terms of optimal solution to that problem)

3. extending Causal Logic with benchmarking simulation tool to navigate uncertainty:

• angie is analyzing & extending bocconi🇮🇹's value lab tool in 🧭🗺️selling entrepreneurial choice/map as Bayes.Entrep #234 (comment)

• angie create below table that can apply mechanism of Chronometer by John Harrison and inertia measurement unit to benchmarking simulation tool (BST) which bootstraps by comparing outcomes of applying different tools to given internal information, to gather more information on external world without external reference. Chronometer compares two different ways of measuring time (local time, determined by observing when the sun reaches its highest point (local noon) and reference time (e.g., Greenwich Mean Time) kept by the chronometer) and to determine spatial (longitude) distance between local and reference point

Aspect	Toolbox-simulation bootstrapping	Chronometer method	IMU
Purpose	To test different learning modes before application	To determine longitude at sea	To measure and calculate orientation, velocity, and position
Information sources	Multiple learning algorithms (Bayesian, behavioral, evolutionary)	Local noon observation and chronometer time	Multiple sensors (accelerometers, gyroscopes, sometimes magnetometers)
Comparison method	Simulating outcomes of different learning modes	Comparing local time to reference time	Integrating data from multiple sensors
Output	Most suitable learning approach for a given situation	Longitudinal position	Orientation, velocity, and position estimates
Underlying principle	Cross-referencing multiple methods to improve decision-making	Using time difference to calculate position	Sensor fusion and integration
Adaptability	Can adjust to different business scenarios	Consistent method across different locations	Can adapt to various motion patterns and environments
Uncertainty handling	Explicitly deals with varying levels of uncertainty	Assumes chronometer accuracy and reliable noon observation	Uses error estimation and correction techniques (e.g., Kalman filters)
Continuous updating	Refines strategies based on ongoing learning	Requires periodic time checks	Constantly updates estimates based on sensor inputs
Error accumulation	Can identify and correct systematic biases in different learning modes	Errors accumulate over time, requiring periodic corrections	Drift errors accumulate over time, requiring periodic external references
Integration method	Combines insights from different learning approaches	Simple time difference calculation	Complex algorithmic integration of sensor data

• to argue why BST is useful, pilots who can only use gps 📡 cannot fly now gps is jammed in europe. this justifies the importance of teaching entrepreneurs learning toolbox📡🧭 i.e. bayes, behav, evol (B2E learning tools), to prevent loss from relying too much on one tool

• importance of causal reasoning in entrepreneurship, and jb mentioned unpredictability of protein (fold, shape) which adjacent possibilities theory or prediction for near near future might be of aid
• jb thinks seeing real world (visual perception by James Gibson ) is essential for driving (@jeanbaptiste do you have reference for this?) while angie is ok with not seeing. our difference stems from whether viewing driving as a ends or means - five reasons of doing science (understand, create value, aethetic etc) applied to startup

4. Pivot Game to educate entrepreneurs to build personalized causal / probabilistic reasoning to

#243

• jb's thesis advisor Wendey mackay on situated behavior affected jb to apply exaptation idea to management
• synthesizing charlie's chip mobility architecture and joseph pine's collaborative, transparent, adaptive, cosmetic customization from The Four Faces of Mass Customization.pdf, i created below

Exaptation Concept	Entrepreneurial Customization	Joseph Pine's Mass Customization Connection	benchmarking simulation tool (BST)	Emoji
Repurposing existing elements	Firms repurpose resources for novel value creation	Adaptive Customization: Modifying standard products to meet specific needs	Adapt existing learning models to new business contexts	🔄
Adaptation to new contexts	Companies adapt strategies to new markets/needs	Collaborative Customization: Working with customers to define their unique requirements	Adjust learning approaches based on changing market conditions	🌱
Emergent functionality	Discovering unexpected value propositions	Transparent Customization: Observing customer behavior to provide tailored solutions without explicit interaction	Identify and leverage unexpected insights from different learning modes	💡
Evolutionary perspective	Iterative experimentation and adaptation	Cosmetic Customization: Presenting a standard product differently to various customers	Continuously refine and evolve learning strategies based on outcomes	🧬
Context-dependent value	Customizing approach to specific market conditions	Combination of approaches: Using multiple customization strategies based on market conditions	Select and combine learning modes based on specific business scenarios	🎯
Non-linear innovation	Developing unexpected paths to value creation	Collaborative and Adaptive Customization: Creating novel solutions through customer interaction and flexible product design	Explore non-traditional combinations of learning approaches for innovative insights	🌀

[hyunjimoon]

summary of the conversation with @chasfine so far on this thread topic

Month	Key Topics and Developments
July-August 2023	- Hierarchical models vs. dummy variables in Bayesian SD
	- Computational Bayes exploration
	- GitHub logging for BayesSD project
	- CEB (Concrete Entrepreneurial problems solvable with Bayes) detector
September-December 2023	- Nick Bloom's papers on technology spillovers and idea finding
	- Marriage of OFE and PPL concept introduction
	- NSS (Nail, Scale, Sail) framework and clock speed discussion
January-February 2024	- Causal loops representation as trees
	- Bits and atoms concept in entrepreneurship
	- Selection of academia, biotech, semiconductor, construction for case studies
	- 🐔🥚 Tick-tock strategy: supervised and unsupervised learning for hypothesis generation
March-April 2024	- Biotech-semiconductor comparison focus
	- Collaborator search for "theory of startup pivot" paper
	- Outline drafting for sequential analysis in startups paper
May-June 2024	- Benchmarking tool development for startup success probability prediction
	- Probabilistic programming applications exploration
	- Creation of six evaluators for the project
	- 🌱 Epigenetic landscape analogy for market and product pivots
July-August 2024	- Rational agency equation in entrepreneurship
	- Planning for ENT and OM papers (technical and operational feasibility)
	- Synthesis of entrepreneurship strategy with integral vs modular lens
	- AI-assisted decision-making in startups discussion
September 2024	- Confirmation of "modular2integral in double helix" as scaling phase in NSS
	- Probabilistic programming vs inverse RL exploration
	- AGCT syntax development for entrepreneurship operations
	- ⚡️ Focus on capitalization operations and deal-investor fit using probabilistic programming

summarizing charlie-angie cld

[hyunjimoon]

Tom, Angie

1. angie's proposal of vensim datatype as inferencedata 1.5years ago has finally gained tom's resource! Vensim -> InferenceData/Arviz #268

2. angie analyzed the cause of bimodality in founder utility on investment amt - valuation cap (middle death valley - got investment but knowledge stock didn't exceed threshold)

[hyunjimoon]

Matt: M, Angie: A

1. finally nailed on the "language" to frame machine's role - problem, not the choice itself. e.g. if an entrepreneur was suggested an offer of $7m investment with 20% ownership, what is the meaning of this? i.e. what problem is this?

• Is "How much money is 20% worth?" the problem?
• Is "How much control do I lose if I don't if I lose 20%" the problem?
• Is "How much am I limited in the future" the problem?

Each of those problems is simultaneously happening. But human can only think about one (or few) of them i.e. can't play 3D chess.

2. Theory and Practice in Research - A criticized Ed Robert's heuristics "ENT scholars shouldn't write about a research question that entrepreneurial practitioners don't know what to do with it" (of course I can infer the context of its origin). In Matt's language, unit theory developer has one's unique role - which Matt agreed. (but one person trying to do everything somehow is burdening: https://otter.ai/s/PBYEaew5SFm3DTEUZmuHcQ?snpt=true.)

3. M resonated with the mess in management field and introduced "holding environment" https://otter.ai/s/7yR98mCaTum-vkyafOIy5Q?snpt=true "we the medical field have a way of getting making it our business to understand and, like, manage the emotions that when you do that, so it's not just you who's trying to deal with it, management doesn't really do that. In fact, management loads it back onto us and says, you know, you have to be better at this. Why aren't you better"

4. upon A's question whether there's hope in management field, Matt replied yes (of course no one in this situation would say no), with two reasons: in america everyone is suing everyone (which is not the reason for hope but cause for mess), and we don't have languages yet. but he said there are people doing that, it's just taking some time https://otter.ai/s/JrAnaEmnQBCFtm2un1AUpg?snpt=true

5. M's comment on the importance of my development (+difficulty of having differnt views) was very consoling https://otter.ai/s/QAXTOeKMSCm_eTEkZYc5hA?snpt=true, recommended "How to Know a Person" by David Brooks and "Nonviolent Communication" by Marshall Rosenberg.

---

• equip skills to implement rational meaning construction for 1
• read M's book: the craft of creativity

Aspect	Insight	Invention	Enlightenment
Definition	New ways of thinking that change what's possible in stories	Specific tangible solutions or creations that resolve stories	New knowledge that changes how we can think going forward
Formation Speed	Forms quickly	Takes longer time	Takes long time to generate
Scope	Occurs within specific stories	Resolves specific stories	Goes beyond single stories
Completeness	Usually incomplete	Complete, practical solution	Complete transformation of knowledge
Nature	Like plot twists that redirect thinking	Concrete implementation	Broader understanding that transforms capability
Example	Nakatsu's analogy between trains and kingfishers	The actual new nose design for Shinkansen	Understanding how serrated shapes affect soundwaves
Relationship to Stories	Redirects stories	Resolves stories	Transcends stories ( #todo connect with productivity)
Success Rate	Can fail or be infeasible	More reliable once developed	Long-lasting impact
SAJ (A's guess (NOT M's)]	👩‍🔬Scientist: Focuses on understanding what will happen by identifying the causal relationships in a situation. 👩‍🎨Artist: Imagines what might happen by envisioning new possibilities and alternative scenarios.	👩‍🔬Scientist: Focuses on understanding what will happen by identifying the causal relationships in a situation. 👩‍⚖️Judge: Evaluates the (dis)utility of what happens, assessing the moral, ethical, and practical implications of various actions.	👩‍🎨Artist: Imagines what might happen by envisioning new possibilities and alternative scenarios. 👩‍⚖️Judge: Evaluates the (dis)utility of what happens, assessing the moral, ethical, and practical implications of various actions.

[hyunjimoon]

tom, angie
discussed the trade-offs between fast decision-making and thorough analysis, emphasizing the importance of balancing the degree of rationality in belief update and action, represented by update clockspeed. Tom agreed SD is also probabilistic program in some sense. They analyzed the impact of market understanding and operational capacity on entrepreneurial success, using a hypothetical EV startup example. They highlighted the challenges of operationalizing customer knowledge, but connected with price discrimination (homogenous to heterogenous WTP for each market segments).

1. Key insight: Can't monetize market understanding without operational capacity

2. Early Stage Priority

• High base costs force operational focus
• Must achieve operational stability before market expansion

3. Growth Stage Balance

• Revenue potential capped by min(ops, market)
• Need to grow both capabilities in sync

4. RRR = Value / Resources = (Revenue - Cost) / Investment
   Operations: Reduces denominator (costs)
   Market: Increases numerator (revenue)

[hyunjimoon]

tom, angie

1. three questions on capability trap 🪤 in 🛠️selling probabilistic program to implement Entrepreneurship #224 (reply in thread)
2. understanding generality-robustness-automation tradeoff fro vikash's diagram on generative db/graphic engines/semantic parsers, esp. rational risk & reward tradeoff in user space 🛠️selling probabilistic program to implement Entrepreneurship #224 (reply in thread)
3. applying tradeoff in 1 and 2 to tradeoff between operational, market, action precision developed here)
4. mit designX
5. integrate process then product architecture of my research

e.g. product3

personal computers, toys, and semiconductors benefit from parallel approaches due to their low test costs 💸, high uncertainty 🎲, and potentially low correlation 🧩. Conversely, industries like shipbuilding, petrochemicals, and commercial aircraft favor sequential strategies due to slow clockspeeds and high development costs.

@tomfid moving forward, will share agenda via github and add review as a reply here.

[hyunjimoon]

@tomfid , angie

angie will introduce one paper each from three schools of thought (cognitive science, strategy, operations) she desires to synthesize:

Vul, E., Goodman, N., Griffiths, T. L., & Tenenbaum, J. B. (2014). One and done? Optimal decisions from very few samples. Cognitive science, 38(4), 599-637. Direction

Gans, J. S. (2023). Experimental choice and disruptive technologies. Management Science, 69(11), 7044-7058.

Phadnis, S, and C. Fine (2017), “End-to-end Supply Chain Strategies: Parametric Study of the Apparel Industry,” Production and Operations Management, 26(12), 2305-2322.

[hyunjimoon]

Bob horn, @jeanbaptiste with discussion prompt #256 (comment) 🗣️

analyzing bob's info mural, angie organized 📦product1 as four modules:

Mural Component	Element	Physical Product (EV)	Digital Product (AdTech)
1. 🎲 Scenario Modeling (Input/Output)	Base Conditions	• Capital: $10k • Timeline: 3 years • Core market: Transportation	• Capital: $10k • Timeline: 1 year • Core market: Advertising
	Balanced Scenario (50/50)	• Equal investment in ops/marketing • Steady growth trajectory • Balanced risk profile	• Equal platform/market investment • Moderate user growth • Technical/market equilibrium
	Operations Heavy (70/30) Scenario	• Manufacturing excellence focus • Cost optimization priority • Quality leadership	• Platform robustness priority • Technical sophistication • Performance leadership
	Marketing Heavy (30/70) Scenario	• Market penetration focus • Brand dominance strategy • Sales leadership	• Rapid user acquisition • Network effect focus • Market leadership
2. 🗺️🧭 Meaning to Moving	Why (🫀 Desire)	By Scenario: • Balanced: Market stability • Ops: Cost leadership • Marketing: Brand leadership	By Scenario: • Balanced: Sustainable growth • Ops: Technical excellence • Marketing: User dominance
	Why (🧠 Belief)	By Scenario: • Balanced: Risk mitigation • Ops: Efficiency wins • Marketing: First-mover advantage	By Scenario: • Balanced: Sustainable moat • Ops: Technical moat • Marketing: Network moat
	What/When (Target)	By Scenario: • Balanced: 3-year stability • Ops: 2-year efficiency • Marketing: 18-month penetration	By Scenario: • Balanced: 12-month growth • Ops: 9-month excellence • Marketing: 6-month scale
	How (Action)	By Scenario: • Balanced: Monthly reviews • Ops: Weekly production • Marketing: Daily sales	By Scenario: • Balanced: Daily metrics • Ops: Hourly performance • Marketing: Real-time tracking
3. 🎯 Target by🧍‍♀️agent, 🎲scenario	Resource Governance	Balanced: • Equal supply/demand Ops-Heavy: • Supply chain optimization Marketing-Heavy: • Market expansion	Balanced: • Equal platform/users Ops-Heavy: • Technical excellence Marketing-Heavy: • User acquisition
	Operations	Balanced: • 1000 units/month Ops-Heavy: • 2000 units/month Marketing-Heavy: • 500 units/month	Balanced: • 99.9% uptime Ops-Heavy: • 99.99% uptime Marketing-Heavy: • 99% uptime
	Market Development	Balanced: • 3 regions Ops-Heavy: • 1 region Marketing-Heavy: • 5 regions	Balanced: • 50k users Ops-Heavy: • 20k users Marketing-Heavy: • 100k users
4. 🎯 Target by ⏰time,🧍‍♀️agent, 🎲scenario	Short-term	Balanced: • Q1-Q2 targets Ops-Heavy: • Production focus Marketing-Heavy: • Market entry	Balanced: • Month 1-2 targets Ops-Heavy: • Platform stability Marketing-Heavy: • User growth
	Medium-term	Balanced: • Q3-Q4 expansion Ops-Heavy: • Efficiency gains Marketing-Heavy: • Brand building	Balanced: • Month 3-6 scaling Ops-Heavy: • Technical moat Marketing-Heavy: • Market share
	Long-term	Balanced: • Year 2-3 leadership Ops-Heavy: • Cost leadership Marketing-Heavy: • Market leadership	Balanced: • Month 7-12 dominance Ops-Heavy: • Technical leadership Marketing-Heavy: • Network effects

---

using Comparing Founder Strategies for Physical vs Digital Startups cld

[hyunjimoon]

tom, angie

1. 🔴automating

can we parse SD .xml file to our parameter classification table?

• experiment in 2. Typify parameters in syntax-semantic with colors, units, bounds #7 (comment)

2. ⚱️polya's urn rich get richer dynamic

• John (SD): 🪵 BayesSD meeting Log #23 (comment)

• Hart (strategy): economy as learning agent
  -- "system is trying to learn which tech is best from a signal made noisy by random arrival of different types of consumers"

• Scott (ent.strategy): population vs agent's S-curve
  -- Choosing Technology: An Entrepreneurial Strategy Approach
  -- choosing experiment of disruptive technology

• Josh (cog.sci): from https://probmods.org/chapters/learning-as-conditional-inference.html,
  -- meaning of "Subjective Randomness"?
  -- compare probmod VS vensim syntax (can we synthesize?)
  -- Independent and Exchangeable Sequences

any exchangeable sequence can be represented as follows, for some latentPrior distribution and observation function f:
var latent = sample(latentPrior)
var thunk = function() {return f(latent)}
var sequence = repeat(2,thunk)

3. bit atom management

• BAE. bit atom energy management 💡⚙️⚡️ #147
• 🪵 BayesSD meeting Log #23 (comment)

4. sigma ratio of layer2:layer1

flexibility encoded as prior on sigma ratio (e.g. process noise: measurement noise ratio?)
in hierarchical bayesian sbc context

• andrew (stats.) https://discourse.mc-stan.org/t/best-practices-for-simulation-based-calibration-with-hierarchical-models/19630/11
• josh (cog.sci)

• marginnote3app://note/25887EAA-DC04-486A-AFAE-E4EFE8419AFF

• john (sd):

Perspective	Key Points on Process Noise
Angie's Question	• Questions "process noise" vs "process variability" terminology • Suggests process noise should be estimated as parameter • Notes high variability can signal system robustness
Hazhir's Definition	• Random process with new draws each time period • Affects model dynamics via flow variables • Represents unmeasurable factors (e.g., weather in SIR models) • Measures unmodeled variation (aleatoric uncertainty)
John's Historical Context	• Real systems have inertia limiting white noise • Evolution from white to first-order autocorrelated noise • Terminology issue: "pink noise" differs from scientific usage so we should rename our first order autocorrelated as brown noise 🦫1/f^2 (🐷1/f vs)

context in Process noise and Feature noise.md, ProcessNoise.pdf

[hyunjimoon]

Tom, Angie
Q1. transition of quantity to quality

- is there a way to represent "rejecting and retiring balance loop", with quality of understanding and theory as stock variables? if it is a quantity, rejecting and retiring is outflow, but if it is a quality, does rejecting becomes inflow? - two parameters classify three evolution types (adaptation, co-opted adaptation, co-opeted non-aption): -- 1. 👓testing delay / 👆implementing delay -- 2. 🔄feedback delay / (👓testing delay + 👆implementing delay)

Q2. To understand one key comment from Theory of adjacent possible, "we note that Steel et al.’s stochastic implementation from Steel20_BDproc_comb_innov.pdf is not a discretisation of the TAP equation itself, because it requires a calculational timestep to be small to keep the creation probabilities below unity; it is a stochastic discretization of a continuum approximation to the original equation. Their numerical analysis also makes the further assumption of a fixed upper limit (usually i max= 4) in the summation." I made a table comparing Steel20's stochastic birth and death process and deterministic adjacent possible model (both authored by Kauffman).

cld

Q3. I think "calculational timestep to be small to keep the creation probabilities below unity" is relevant to ST>DT from John Sterman's comment below and wonder what @tomfid thinks (which i listed as last week's agenda).

Jay Forrester knew back in the 1940s and 50s that white noise doesn’t exist in reality because all real systems have some inertia that limits the power in the higher frequencies of random disturbances, whether those are in a physical servomechanism, a supply chain, the economy, or one’s beliefs.  In Industrial Dynamics he addressed this issue by sampling from a white noise distribution at finite intervals (rather than at every time step, “DT").  FYI, sampling at every time step in a system integrated numerically by Euler’s method means that the power spectrum of the noise added to any system (process or measurement noise) will be sensitive to the time step because the highest frequency in a pseudo-white noise stream with values chosen every DT is 1/2DT).  Even worse, using higher-order integration schemes such as Runge-Kutta have large impacts on the power spectrum of noise in models.  The time step and numerical simulation algorithm used in any model are artifacts of the simulation method, and not a feature of reality.  The results of any model, including the model’s response to process and measurement noise, should not depend on the time step or simulation methods. By sampling white noise at a finite interval, say, ST (for Sampling Time), ST>DT, Jay’s early models solved the problem of having the system’s response to noise be sensitive to the time step in the Euler integration, but it was rather arbitrary (the power spectrum has a sharp and unrealistic cutoff at the highest frequency allowed, 1/2ST, which is independent of DT.  So in the 60s folks in system dynamics came up with the formulations for first-order autocorrelated noise.  They called it pink noise to indicate that it was not white noise.

[hyunjimoon]

based on Tom's feedback that feedback flow (from quality of form(body) to quality of understanding) may not deteriorate quality of form (body) i.e. this is information flow, i separated three stock variables as below. sign of the relationship between quality of understanding to calculated value is - sign at first but becomes + sign later. for example, during phd 1 year when one has low function quality, increasing the understanding would lower calculated value ("i'll write the best thesis" -> "omg what a deep sea of knowledge there are. i just wish to graduate.."). so it's - sign. however as one's function quality improves, better understanding positively contributes to calculated value. so it becomes - sign. this means purple loop in NAIL is reinforcement loop at first but becomes balancing loop later, explaining NAIL-S curve. This sign transition may happen during SAIL as well, explaining SAIL-S curve. SCALE-S curve is mostly explained by R2 loop: "calculated value increases implementing rate increases quality of form, increasing measure value (as one can increase experiment's fidelity with a higher quality form").

Calculated Value	Quality of Function	Quality of Understanding	"Quality of Understanding" $\rightarrow$ "Calculated Value" Sign	NAIL, SAIL S-curve
H	H (phd4)	H (phd4)	+
L	L (phd2)	H (phd2)	- $\rightarrow$ +	R1,R3 $\rightarrow$ B1,B3
H	H	L
H	L (phd1)	L (phd1)	-

Would like to get @tomfid @chasfine 's feedback
UTB_horizontal.mdl.zip

this aims to explain Charlie's three s-curves each during nail (R1->B1), scale (R2-> two cases below), sail (R3->B3):

two cases for SCALE:

• step increase of needed quality (e.g. covid): balance loop is added to counteract the R2
• gradual increase of needed quality(e.g. market saturation): R2's exponential increase turns into exponential decrease (lowered measured value triggers lowered calculated value, lowering implementing and form's quality, lowering measured value)

---

\subsection{Phase-Dependent Dynamics}

The relationship between understanding and calculated value exhibits critical sign transitions that drive startup evolution through three distinct phases. During the Nail phase, initial reinforcing feedback (R1) transforms into a balancing loop (B1) as deeper understanding reveals underlying complexities. This transformation manifests when limited early understanding accelerates development, but eventually comprehensive understanding exposes implementation challenges.

The Scale phase operates primarily through reinforcing feedback (R2) driven by implementation quality improvements. This feedback loop encounters moderation through two mechanisms: sudden increases in needed quality (exemplified by events like COVID-19) that introduce balancing feedback, and gradual market saturation that converts exponential growth to decline.

In the Sail phase, system-level optimization mirrors the Nail phase dynamics through a transition from reinforcing (R3) to balancing (B3) feedback. This final S-curve emerges from the interaction between system understanding and functional optimization at scale.

[hyunjimoon]

📦research product

Process / Product	MOD (🧍‍♀️Founder)	INTEG (🌏Founder+Investor)
INTEG (🧭, 🗺️ How,What)	3. BPE|D Belief, meaning, judgment, and environment functions are integrated while desire function remains separate. Environment and belief are synchronized but system still permits variable desire/preference structures.	4. BPDE Complete integration of belief, meaning, judging, desire, and environment functions. Individual entrepreneurial cognition fully aligns with societal-level uncertainty dynamics.
MOD (🧭|🗺️ How|What)	1. BP|DE Belief, meaning, and judging functions are variable while desire and environment are fixed. This represents the basic entrepreneurial state before any integration with external stakeholders.	2. BPD|E Belief, meaning, judging, and desire functions are integrated while environment remains separate. System can model joint founder-investor preferences but not yet full market dynamics.

THEORETICAL VERSION:
My dissertation characterizes entrepreneurial decision making by separating it with conventional decision setting for optimal stopping, also known as secretary problem. While secretaries are chosen from fixed candidate pools with clear thresholds, entrepreneurs navigate constantly evolving option sets and action spaces. I prove why this variable state-action space invalidates traditional frameworks and propose a path from BP|DE (where belief, meaning, and judging functions operate under fixed desire and environment) to BPDE (where entrepreneurial cognition fully integrates with market-level uncertainty). This transformation represents a new theory of entrepreneurial choice that bridges individual decision-making with collective market intelligence. [[🌒⌨️Entrepreneurs Are Not Secretaries_A Theory of Variable State-Action Decision Making]]

To achieve this integration, I first tackle the founder-investor alignment problem, moving from BP|DE to BPD|E by incorporating desire while keeping environment modular. Using investor archetype analysis, I demonstrate how heterogeneous investors vary in both meaning construction (how they interpret signals) and judging preferences (what they value). This foundational step reveals how rational meaning construction emerges between founders and investors through Bayesian calibration. [[🌓📐Bayesian Calibration in Entrepreneurial Decision Making]]

The final transformation from BPD|E to BPDE introduces environment integration, where individual entrepreneurial decisions harmonize with societal-level uncertainty dynamics. The theory shows how entrepreneurs can align their cognitive processes with market-level epistemic and aleatoric uncertainty ratios, achieving true Bayesian calibrated choice. This sequential approach - first establishing rational meaning construction, then extending to market-level integration - provides a complete theoretical framework for entrepreneurial decision-making. [[🌔Bayesian Calibrated Choice_Balancing Individual and Market-Level Uncertainty]]

ALGORITHMIC VERSION:
My dissertation formalizes entrepreneurial decision-making through a variable state-action framework where both S (state space) and A (action space) are non-stationary. I prove that when dim(S) and dim(A) vary over time t, standard optimal stopping conditions ∃τ* : E[V(τ*)] ≥ E[V(τ)] ∀τ no longer hold. This motivates a transformation from BP|DE, where {Belief,Meaning,Judging} are variable functions over fixed {Desire,Environment}, to BPDE, where all functions achieve Bayesian self-consistency.

The first transformation from BP|DE to BPD|E implements Bayesian calibration through p(action|state) and p(state|observation) modeling. Using synthetic investor generation and hierarchical logit analysis, I model the joint distribution p(meaning, judging|investor_type). This allows quantification of how meaning construction p(s|o) and judging functions p(a|s) vary across investor archetypes, enabling systematic analysis of rational meaning construction.

The final transformation to BPDE implements simulation-based calibration checking for Bayesian computation, verifying when individual cognitive architectures achieve consistency with market-level uncertainty. The implementation uses program synthesis to model the complete integration, with Bayesian self-consistency equations πprior(θ) = ∫dy∫dθ̃ πpost(θ|y)πobs(y|θ̃)πprior(θ̃) verifying alignment between individual and market-level distributions.

---

⚙️(📦) process of convergence to research product

Jan.07

received feedback on smart sampling startup.pdf from those who assumed the role of scientist, engineer/modeler, practitioner. This is BMJDE from above two by two table.

Table: Integration of Feedback into Action Items

Priority	Action Item	Feedback (Direct Quotes)	Process to Act
1	Strengthen Decision Framework Integration	Scientist: "This is a more comprehensive decision model, because you get decisions right that your decision is either I'm going to continue to experiment, I'm going to scale or I'm going to fail." Angie: "I haven't fully followed what their logic is, but they're saying so this experiment is our experiment... But it seems that the current literature is saying there are kind of like experiments in between the two commitments."	• Simplify decision model to binary choice (commit vs learn) • Demonstrate how sampling strategy itself is a decision through Tesla battery example
2	Divergence from Traditional Statistical Decision Theory	Scientist: "When I learned optimal stopping, it was the secretary hiring process..." Angie: "Your desire changes as you search. So compared to a parking lot, where just your goal is to park, Secretary is you have to see. And as you see the quality Secretary pool, you may either upgrade or downgrade your desire."	• Develop formal contrast between parking (fixed goal) and secretary (evolving goal) problems • Show how entrepreneurial sampling maps to secretary problem
3	Clarify Hierarchical Model Integration	Modeler: "A lot of pure inferences seems like it's concerned with something like measuring the effect of some new molecule on diabetes. And that in itself, is not a decision. The decision, then is what do you do with that information." Angie: "The exchangeability comes as an abstraction tool, because even though the data that you observe is very convoluted... at some point, you don't really care about the underlying thing."	• Develop secretary example showing how hierarchical structure affects sampling strategy • Connect hierarchical Bayesian updating to concrete decisions
4	Add Practical Implementation Guidelines	Practitioner: "Single optimal care, same scenarios... And what's with dynamic case... I mean the first like fifth case or dynamic make a lot of sense. The single optimal case does not make a lot of sense." Angie: "Texas people prefer us battery way much compared to California... numbers of sample you would need from the Texas is much larger compared to the California."	• Create decision rules for choosing between sampling strategies (fixed/optimal/dynamic) • Use Tesla battery case to show when to switch between strategies

Jan.08

1. commitment
   During our last chat, it became clear that 1) there is a gap in Bayesian entrepreneurship literature, 2) I have some idea how to fill this gap, 3) I feel I'm one of the best people to do this, which makes me excited. I finally committed to Entrepreneurship as my market.

2. belief shattering
   After specifying the customer, testing their perception of Bayes became the bottleneck. I learned my ignorance on entrepreneurship scholar's perception for: 1. sampling strategy 2. hierarchical bayesian 3. bayesian calibration

Especially I was very surprised that Jeff Dotson and Prof. Ben-Akiva's phd student (who have relatively expertise on Bayesian computation, Prof. Ben-Akiva covers Bayesian heavily in his demand analysis class as attached) weren't aware that bayesian calibration differs from bayesian inference ("Bayesian inference provides a procedure for constructing inferences but it offers no general guarantees on the overall behavior of those inferences. If we want to ensure robust inferences then we at least need to attempt some calibration in order to determine how we expect our model to perform").

3. subgoal adjusting
   I might need some expertise to bridge bayesian inference with discrete choice analysis first.

Jan.09

connected mobility, inspection paradox #209, #186
annotate(angie, MMI senior fellow 24).pdf

Institutional approaches to mobility metrics do more than merely observe market development and user preferences; they actively shape them (Zhao, 2009). As David Zippers analysis suggests, the auto industrys emphasis on metrics such as range and accelerationoften fueling car bloatnot only misaligns with actual usage patterns (where 19 out of 20 trips are under 20 miles) but also channels research and development in directions that may undermine long-term sustainability goals. These measurement choices highlight a broader challenge in mobility innovation: the bidirectional interplay between institutional evaluation frameworks and user preferences, in which metrics not only record behavior but also influence it.

This dynamic becomes evident in how purely top-down approaches often fail to grasp or adapt to the complexity of mobility systems. As inspection paradox (Angus, 1997; Ross, 2003) illustrates, measurements taken outside the system (e.g., an aerial or aggregate view) can systematically diverge from those made inside a system (e.g., a random arrivals experience). This implies institutional perspectives on mobility innovation can markedly differ from everyday user experiences. A similar gap appears between the investor (observer) and founder (mover) perspectives, contributing to underinvestment in certain high-impact solutionsmicromobility, for exampledespite strong evidence of shorttrip benefits. By recognizing how inside vs. outside measurements diverge, mobility stakeholders can better align institutional metrics with the day-to-day realities of users. These misalignments also drive force-fed equity in hardware-heavy ventures and a persistent mismatch between conventional venture capital evaluation frameworks and the capital requirements of mobility hardware startups, echoing initiatives to integrate psychological factors and user preferences into planning (Zhao, 2009; Sumantran et al., 2017).

The path forward lies in recognizing how both social and physical mobility thrive through diverse, integrated approaches that simultaneously accommodate and shape preferences. This is well captured in CHIP mobility architecture (Sumantran et al., 2017) which highlights connected, heterogeneous, intelligent, and personalized solutions that can unify otherwise fragmented travel modes. Fig.1 shows how societal imperatives from the top and user preference from the bottom can be connected through intelligent applications of heterogeneous modes. Just as transportation networks blend multiple modes through intelligent routing, the entrepreneurial ecosystem works best with various financing options matched to specific needs. This approach lets founders align different funding sources (grants, debt, equity) (Lerner et al., 2022) with specific project phasesfrom early RD to hardware deploymentsimilar to how routing applications match transport modes to journey segments. By integrating diverse travel modes with intelligent routing and

2 Figure 1: CHIP Mobility Framework, reproduced from Figure11.6 in Sumantran et al. (2017)

combining capital sources through flexible financing, institutions can accelerate mobility innovation. Rather than using rigid metrics or single-mode solutions, strategies like crosssector partnerships and iterative prototyping provide the adaptability modern systems need. For mobility startups, this means accepting partial commitments and solving immediate problems without waiting for full funding; for transportation systems, it means building networks that evolve with user preferences

Jan.10

gave talk at TBV Conference Agenda and List of Attendees 9-11 Jan 2025_final.pdf with 🛝angie moon rational meaning construction.pdf

4. charlie's advice on decision framing (p.18) + "Connectivity" from CHIP mobility were the basis of the attached slide which I presented at the conference. It went well and the audience (especially Teppo Felin) liked my 🚗mobility metaphor! This motivates me to develop algorithms that enable efficient matching of heterogeneous belief from society's perspective and effective path finding from the founder's perspective. If the algorithms are good enough, entrepreneurship institutions like CDL (Toronto) and ION (Boconni, Utah, HEC) can be its target market.

• I identified another gap (on top of Bayesian inference vs calibration). Economy and statistics may have different perspectives on probability due to its history of adoption. More to come. I'd love to get your opinion on the attached textbook's chp.9 (decision analysis) and part 3. advanced computation. This would be my Bayesian brain as I translated it... how it differs from your image of "statistical decision theory". This will be very helpful in learning how to bridge (computational) Bayesian statistics and Bayesian entrepreneurship.

Jan.13

1. high level choice via ↗️ then 🏷️

↗️ arrow (integration order: product then process) came before the 🏷️label (is lifting environment process or product?).

i knew integrating product should come before integrating process but wasn't sure which one between lifting belief to (belief, desire) and lifting founder to (founder, investor) was product.

after feeling the ratio of hardship `BMJ|DE ➡️ BMJE|D` / `BMJ|DE ➡️ BMJD|E` > 1, I judged lifting E may be integrating process and lifting E may be integrating product (liable to change). so i committed to `BMJ|DE ➡️ BMJD|E` path.

external evaluator may view this as changing from BMJE|D to BMJD|E but for me it was testing two (BMJ|DE ➡️ BMJE|D and BMJ|DE ➡️ BMJD|E) then choose one. To be specific, consideration set constructed from options is { BMJ|DE ➡️ BMJE|D , BMJ|DE ➡️ BMJD|E} and i.e. flow from meaning (belief encoding) was not high enough to make one judgement. so i put efforts in strengthening meaning flow (kept sampling) so that it's locally balanced with judging flow.

2. high level choice guides low level choice

low level customer decision

• among customer's consideration set {discover value to new customer, deliver value to exiting customer} former. theory-driven entrepreneurship scholars will be my new customer. this market will grow (evidence: strategy science special issue and harvard collaborator's choice).

🌕FULL CIRCLE by fully digesting my Jan.08 choice "committed to Entrepreneurship as my market." from above. This circular checking is example of BAYESIAN CALIBRATED CHOICE.