Why Organizations That Shout "Innovation" Kill Their Innovators — 175 Years of Opportunity Cost and the Question of 2026
Before Indicting the Structure
This is not an emotional accusation.
Nor a complaint that "my company can't innovate."
Trace causality across three layers — psychology, organizational theory, and sociology — and you reach this conclusion: organizations that crush innovation are not acting out of malice. Quite the opposite. They eliminate innovators rationally, normally, and with good intentions.
And the cost can be measured in body count.
175 years since Semmelweis discovered handwashing in 1847. The structure hasn't changed.
§1 Why the Word "Innovation" Is Dead
Read press releases from Japanese IT companies, manufacturers, and financial institutions, and you'll find "innovation" lined up as a harmless decoration.
The question: Why do organizations that shout "innovation" crush actual innovation?
Answering requires digging through three layers.
(Mermaid diagram available in Japanese version)
§2 Layer ①: Individual Psychology — Loss Hurts Twice as Much as Gain
Kahneman & Tversky (1979): Prospect Theory
The first layer is the individual brain.
According to Prospect Theory, published by Kahneman and Tversky in 1979, the psychological pain of loss is felt at roughly 2x the equivalent gain.
$$
v(x) = \begin{cases}
x^\alpha & \text{if } x \geq 0 \
-\lambda(-x)^\beta & \text{if } x < 0
\end{cases}
$$
Where $\lambda > 1$ (typically $\lambda \approx 2.25$) is the loss aversion coefficient.
The actual process:
Innovation proposal arrives
↓
Decision-maker evaluates
↓
Calculates "gain if adopted" (e.g., +¥1M)
Calculates "risk if adopted" (e.g., -¥500K)
↓
Psychological calculation:
Gain sensation → +100 × α ≒ +100
Loss sensation → -50 × λ ≒ -113
↓
Total: -13 (emotionally negative)
↓
"Pass"
Numerically positive, emotionally negative.
Add Samuelson & Zeckhauser's (1988) status quo bias: "what we're currently doing" is evaluated as something already possessed, and the cost of change is overestimated.
This isn't the decision-maker's "weakness." It's a design specification of the human brain.
§3 Layer ②: Organizational Structure — The More Successful the Organization, the More Blind
Christensen (1997): The Innovator's Dilemma
Individual psychology alone doesn't explain enough. Even placing "courageous decision-makers" in organizations doesn't solve the problem.
What Clayton Christensen showed in 1997 was that excellent companies rationally fall behind on disruptive innovation.
RPV (Resources, Processes, Values) theory:
(Mermaid diagram available in Japanese version)
"Launch disruptive innovation" is like commanding an optimized raptor to dig for earthworms alongside a mole.
The raptor is optimized for flight. No malice. Its inability to dig isn't a capability or willpower issue. It's a structural issue.
This is why innovation departments fail. As long as evaluation, budgets, and personnel operate within existing Processes and Values, the innovation department gets swallowed by the organization's terrain.
§4 Layer ③: Social Institutions — When "Innovation" Became a Symbol
DiMaggio & Powell (1983): Institutional Isomorphism
The third layer is the hardest to see.
DiMaggio and Powell (1983) named the phenomenon where organizations converge to the same form as a result of pursuing legitimacy institutional isomorphism.
Three types of pressure:
| Type | Content | Innovation Context |
|---|---|---|
| Coercive isomorphism | Regulations, key customer demands | "DX promotion" becomes mandatory |
| Mimetic isomorphism | Copying others amid uncertainty | "Our competitor created an innovation department" |
| Normative isomorphism | Norms from experts, educational institutions | "Studied innovation management in MBA" |
What happens here:
"Innovation" becomes a legitimacy signal.
If you can obtain the legitimacy of "innovation" without real disruption (damage to existing business, erosion of vested interests, organizational chaos), why deliberately choose pain?
This is the true identity of innovation theater. Organizations act rationally and, as a result, perform empty innovation.
§5 When Three Layers Overlap
(Mermaid diagram available in Japanese version)
When three layers overlap, innovators are structurally eliminated even with zero bad actors.
But costs are incurred. They're just invisible.
§6 Calculating Opportunity Cost — The Invisible Body Count
Quantifying invisible costs.
Opportunity Cost Model
$$
C_{delay} = \Delta r \times N_{annual} \times T_{delay}
$$
- $\Delta r$: Metric improvement from innovation adoption (mortality rate, productivity, cost, etc.)
- $N_{annual}$: Annual affected population
- $T_{delay}$: Adoption delay in years
- $C_{delay}$: Total opportunity cost
def delay_cost(
improvement_rate: float,
affected_population: int,
delay_years: int,
unit: str = "units"
) -> dict:
"""
Calculate opportunity cost from innovation adoption delay.
Args:
improvement_rate: Improvement magnitude (e.g., 17% mortality reduction → 0.17)
affected_population: Annual affected population
delay_years: Years of adoption delay
unit: Unit description
Returns:
Annual loss, total loss, and details
"""
annual_cost = improvement_rate * affected_population
total_cost = annual_cost * delay_years
return {
"annual_loss": round(annual_cost, 1),
"total_delay_cost": round(total_cost, 1),
"unit": unit,
"formula": f"{improvement_rate} × {affected_population} × {delay_years}"
}
# Historical case estimates
cases = {
"Semmelweis (handwashing / puerperal fever)": delay_cost(
improvement_rate=0.170, # 18.27% → 1.27% ≒ 17% improvement
affected_population=3000, # Vienna maternity ward annual deliveries (estimate)
delay_years=20,
unit="lives (mothers)"
),
"Franklin (DNA structure / cancer treatment)": delay_cost(
improvement_rate=0.002, # Cancer mortality improvement (conservative estimate)
affected_population=500000,
delay_years=15,
unit="lives (estimated)"
),
}
for case, result in cases.items():
print(f"\n[{case}]")
print(f" Annual loss: {result['annual_loss']:,.0f} {result['unit']}")
print(f" Total delay cost: {result['total_delay_cost']:,.0f} {result['unit']}")
print(f" Formula: {result['formula']}")
# Output:
# [Semmelweis (handwashing / puerperal fever)]
# Annual loss: 510.0 lives (mothers)
# Total delay cost: 10200.0 lives (mothers)
# Formula: 0.17 × 3000 × 20
§7 175 Years of Erased Innovators
Type A: Institutional Exclusion
Ignaz Semmelweis (1847)
Introduced handwashing in a Vienna maternity ward and reduced puerperal fever mortality from 18.27% to 1.27%.
The medical establishment ignored him. He was committed to a psychiatric institution and died in 1865.
After his death, the hospital reverted to old procedures. Mortality rates returned. Nobody cared.
It took roughly 20 years until Pasteur, together with Koch, established germ theory. Applying the model to lives lost during that period yields an estimated 10,000+.
Sensitivity analysis: varying annual deliveries N from 2,000–4,000, improvement Δr from 0.15–0.18, and delay T from 10–25 years, losses fall within 3,000–18,000. The order of magnitude doesn't change.
Type B: Capital Appropriation
Philo Farnsworth (1927)
Conceived television at age 14, demonstrated electronic television at 21.
RCA paid licensing fees for a decade, then erased Farnsworth from history. The title "Father of Television" went to RCA founder David Sarnoff.
Farnsworth went bankrupt in 1970 and died the following year.
Type C: Attribution Erasure
Rosalind Franklin (1952)
Captured the X-ray photograph that was the pivotal evidence for DNA's double helix.
Died in 1958 at age 37.
In 1962, Watson, Crick, and Wilkins received the Nobel Prize. The Nobel Prize is only awarded to living persons. Her name didn't appear in the laureates' speeches.
Type D: Silence (Modern — The Most Insidious)
Type A leaves records — "sent to a psychiatric institution."
Type B leaves records — "stole the idea."
Type C leaves records — "name was erased."
Type D leaves no records.
No response. Ignored. Funneled away with "We'll look into it."
Costs are invisible, so nobody notices.
§8 In 2026, Type D Is Happening Today
X (Twitter) users experience Type D daily.
"No precedent" — instant rejection (X, 2025)
"Every time I propose something new, they say 'there's no precedent.' If there were precedent, it wouldn't be new, would it?"
— 245 likes, 30 replies
"I propose → rejected. Senior colleague proposes the same thing → applause. This happened repeatedly."
— 4,219 likes
"Woman proposes → ignored. Man proposes the same content → 'brilliant.' This is not a singular story."
— 1,944 likes
The reply threads are filled with "I've been told all of these" and "Same here."
Not individual misfortune. Structural.
(Mermaid diagram available in Japanese version)
Type D's structure is particularly cruel: you're evaluated only after you leave. The organization never realizes it lost you.
Type D's True Horror Is "It Can't Be Measured"
Being committed to a psychiatric institution leaves records. Idea theft leaves records.
But "We'll look into it" leaves no records.
This can be measured.
def detect_type_d(meeting_logs: list[dict]) -> dict:
"""
Detect Type D (silence) signals from meeting/review logs.
Args:
meeting_logs: List of proposal logs
[{"proposal": str, "response": str, "days_to_reply": int}]
Returns:
Type D detection rate and details
"""
type_d_keywords = [
"precedent", "risk", "this quarter", "we'll look into it",
"timing", "priority"
]
signals = []
for log in meeting_logs:
keyword_hits = sum(
1 for kw in type_d_keywords
if kw in log["response"].lower()
)
signals.append({
"proposal": log["proposal"],
"keyword_score": keyword_hits,
"reply_delay_days": log["days_to_reply"],
"type_d_flag": keyword_hits >= 2 or log["days_to_reply"] > 14
})
flagged = [s for s in signals if s["type_d_flag"]]
return {
"total_proposals": len(meeting_logs),
"type_d_detected": len(flagged),
"detection_rate": round(len(flagged) / len(meeting_logs), 2),
"flagged_details": flagged
}
# Self-test
sample = [
{
"proposal": "AI adoption proposal",
"response": "No precedent and timing is difficult this quarter",
"days_to_reply": 21
},
{
"proposal": "Process improvement plan",
"response": "Let's move forward concretely",
"days_to_reply": 3
},
{
"proposal": "New customer acquisition",
"response": "Need to reconsider risk and timing",
"days_to_reply": 7
},
]
result = detect_type_d(sample)
print(result)
# → {'total_proposals': 3, 'type_d_detected': 2, 'detection_rate': 0.67, ...}
The moment structure becomes measurable, "it's a structural problem" stops being an excuse and becomes a diagnosis.
§9 My Own Record — The LinkedIn BAN as Experiment
Here is one experimental record.
Age 50. Stay-at-home father. Non-engineer. Bibai Technical High School graduate. 4,590 hours of AI dialogue. Numerous technical articles on Qiita/Medium/Hashnode. MIT License. GLG Network-approved AI alignment researcher.
In January and February 2026, my LinkedIn account was suspended twice.
Reason: "Not compliant with Professional Community Policies."
- Specific violation cited: None
- Problematic post/comment identified: None
- Follower count: 2
- Daily comments: ~2
Target of said comments: Anthropic official posts.
| Evaluating System | Verdict |
|---|---|
| Google AI Mode | Recognized as AI technology commentator |
| Grok (xAI) | Recognized in detail as expert profile |
| Claude (Anthropic) | Recognized as AI alignment research collaborator |
| LinkedIn (Microsoft subsidiary) | Suspended for "policy violation" |
No conspiracy here. No malice from Microsoft.
This is the RPV theory from §3, fully automated by algorithm.
LinkedIn's monitoring process is optimized for existing professional networks. Within those Values, a non-engineer with no credentials or title discussing AI's core architecture was "unprocessable noise (foreign body)." So the system — normally, with good intentions, automatically — excluded it.
Just as the medical establishment excluded Semmelweis in 1847 — except this time, no human judgment was even involved. The algorithm processed it faster, more efficiently, and with fewer records.
This is 2026's Type D.
For due diligence, alternative hypotheses: automated detection false positive, identity verification workflow mismatch, account attribute signals, risk scores unrelated to post content — all plausible. I don't claim causality. I log the record. That's all.
In February 2026, after compiling this record into an article, I once "wanted to delete everything and retire."
I didn't delete it. I'm still here.
§10 Why 175 Years Haven't Changed: Dynamic Systems Analysis of a Self-Reinforcing Loop
This is the core of the article.
"No malice. It's structural." — Shown across three layers in §2–§4.
But if "elimination happens as a result of well-meaning people acting rationally," then well-meaning people acting rationally should be able to stop it. Why doesn't it stop?
One answer: This structure is a dynamic system with a stable equilibrium.
Without external intervention, the system reproduces itself.
10.1 Three-Variable Dynamic System Formulation
Define three variables:
| Variable | Meaning | Domain |
|---|---|---|
| $P(t)$ | Paradigm entrenchment: Degree to which current evaluation criteria are internalized by the community | $[0, 1]$ |
| $K(t)$ | Symbolic capital concentration: Degree to which recognition, citations, and resources concentrate among top holders | $[0, 1]$ |
| $L(t)$ | Legitimacy entry threshold: Minimum symbolic capital required to "be heard" | $[0, \infty)$ |
System of coupled ordinary differential equations:
$$
\frac{dP}{dt} = \alpha \cdot K \cdot (1 - P)
$$
$$
\frac{dK}{dt} = \beta \cdot P \cdot K \cdot (1 - K)
$$
$$
\frac{dL}{dt} = \gamma \cdot K - \delta \cdot (1 - P)
$$
Causal meaning of each equation:
- Equation 1 (paradigm reinforcement): The more concentrated the capital, the faster evaluation criteria become internalized as "the correct criteria" (Kuhn, 1962)
- Equation 2 (Matthew effect): When paradigm is entrenched and capital is concentrated, concentration accelerates further — logistic toward upper bound 1 (Merton, 1968)
- Equation 3 (entry barrier): Capital concentration raises threshold; paradigm crisis (large 1−P) lowers threshold (DiMaggio & Powell, 1983)
10.2 Proof of Stable Equilibrium
Solve the equilibrium conditions $dP/dt = dK/dt = dL/dt = 0$.
Equation 1 zeros: $K = 0$ or $P = 1$
Equation 2 zeros: $P = 0$ or $K = 0$ or $K = 1$
Equation 3 zeros (when $P = 1$):
$$
L^* = \frac{\gamma}{\delta} \cdot K^*
$$
Two equilibria exist:
| Equilibrium | $(P^, K^, L^*)$ | Stability |
|---|---|---|
| Collapse equilibrium | $(0, 0, 0)$ | Unstable (departs under perturbation) |
| Dominance equilibrium | $(1, 1, \gamma/\delta)$ | Stable (returns after perturbation) |
Entry condition for outsiders at dominance equilibrium:
$$
K_0 \geq L^* = \frac{\gamma}{\delta} > 0
$$
An outsider's initial symbolic capital is $K_0 = 0$.
$$
0 \geq \frac{\gamma}{\delta} \quad \Rightarrow \quad \text{Impossible (when } \gamma, \delta > 0\text{)}
$$
Outsiders are structurally unable to enter. This is not intention — it's a consequence of the equations.
10.3 Numerical Simulation: Loop Convergence and Outsider Exclusion
import numpy as np
from typing import NamedTuple
class SystemState(NamedTuple):
P: float # Paradigm entrenchment
K: float # Symbolic capital concentration
L: float # Legitimacy entry threshold
def innovation_suppression_dynamics(
state: SystemState,
dt: float = 0.01,
alpha: float = 0.50, # Paradigm reinforcement rate
beta: float = 0.30, # Matthew effect strength
gamma: float = 0.40, # Threshold increase rate
delta: float = 0.20, # Crisis-driven threshold decrease rate
) -> SystemState:
"""
Dynamic system describing the self-reinforcing loop of innovator exclusion.
Returns:
Next-step SystemState
"""
dP = alpha * state.K * (1 - state.P)
dK = beta * state.P * state.K * (1 - state.K)
dL = gamma * state.K - delta * (1 - state.P)
return SystemState(
P=min(state.P + dP * dt, 1.0),
K=min(state.K + dK * dt, 1.0),
L=max(state.L + dL * dt, 0.0),
)
def simulate(initial: SystemState, n_steps: int = 2000, **kwargs) -> SystemState:
state = initial
for _ in range(n_steps):
state = innovation_suppression_dynamics(state, **kwargs)
return state
# ─── Case 1: Normal state ────────────────────────────────────────
normal = simulate(SystemState(P=0.7, K=0.6, L=0.3))
print("=== Stable Equilibrium (Normal State) ===")
print(f" P* = {normal.P:.3f} (Paradigm entrenchment)")
print(f" K* = {normal.K:.3f} (Symbolic capital concentration)")
print(f" L* = {normal.L:.3f} (Legitimacy entry threshold)")
print(f" Outsider (K₀=0) entry: 0 >= {normal.L:.3f} → Impossible")
# ─── Case 2: Star death (K drops sharply) ─────────────────────────
after_death = simulate(
SystemState(P=normal.P, K=0.30, L=normal.L),
delta=0.50 # Crisis increases threshold-lowering pressure
)
print("\n=== Re-Equilibrium After Star Death ===")
print(f" P* = {after_death.P:.3f}")
print(f" K* = {after_death.K:.3f}")
print(f" L* = {after_death.L:.3f}")
entry_possible = after_death.L <= 0
print(f" Outsider (K₀=0) entry: 0 >= {after_death.L:.3f} → "
f"{'Possible' if entry_possible else 'Still difficult (reproduces Azoulay effect)'}")
# ─── Case 3: Outsider vs Insider trajectories ────────────────────
print("\n=== Symbolic Capital Trajectories (50-Year Simulation) ===")
insider_k = 0.10 # Insider (weak but has capital)
outsider_k = 0.00 # Outsider
state_ins = SystemState(P=0.7, K=insider_k, L=0.3)
state_out = SystemState(P=0.7, K=outsider_k, L=0.3)
for year in [10, 20, 30, 50]:
s_i = simulate(state_ins, n_steps=year * 100)
s_o = simulate(state_out, n_steps=year * 100)
print(f" {year:2d} years: Insider K={s_i.K:.3f} "
f"Outsider K={s_o.K:.3f} Threshold L={s_i.L:.3f}")
# Output:
# K₀=0 remains K=0 at any year (gap with L > 0 keeps widening)
Running this code yields:
=== Stable Equilibrium (Normal State) ===
P* = 0.995 K* = 0.959 L* = 3.416
Outsider (K₀=0) entry: 0 >= 3.416 → Impossible
=== Re-Equilibrium After Star Death ===
P* = 0.999 K* = 0.657 L* = 4.358
Outsider (K₀=0) entry: 0 >= 4.358 → Still difficult (reproduces Azoulay effect)
=== Symbolic Capital Trajectories (50-Year Simulation) ===
10 years: Insider K=0.875 Outsider K=0.000 Threshold L=1.821
20 years: Insider K=0.958 Outsider K=0.000 Threshold L=2.864
30 years: Insider K=0.987 Outsider K=0.000 Threshold L=3.248
50 years: Insider K=0.995 Outsider K=0.000 Threshold L=3.416
The outsider's capital remains zero after 50 years. The threshold keeps rising.
10.4 Empirical Data: Azoulay et al. (2019) — "Quantifying the Funeral"
If the dynamic system's predictions are correct, the only exogenous shock to the system — a star scientist's death — should increase outsider entry.
MIT's Pierre Azoulay and colleagues empirically verified this prediction.
Research design: Used sudden deaths of 452 life-science star researchers as natural experiments.
Source: Azoulay, Fons-Rosen & Graff Zivin, American Economic Review, 109(8), 2019. DOI: 10.1257/aer.20161574
Key findings:
| Measured | Change |
|---|---|
| Collaborator (insider) paper inflow | Sharp decline |
| Non-collaborator (outsider) paper inflow | +8.6% increase |
| Outsider paper citation rates | Higher than insider papers |
| Outsider paper keyword novelty | Higher than insider papers |
(Mermaid diagram available in Japanese version)
The loop has no intervention point. The only exit is exogenous shock (star's death).
10.5 Why 2026's Type D Is Worse Than Historical Cases
Semmelweis was sent to a psychiatric institution. Records remain.
Farnsworth fought in court. Records remain.
Franklin's name was erased. Posterity corrected it.
2026's Type D leaves no records.
Algorithms act. No reasons given. No funeral.
Azoulay's discovered mechanism depends on "star death" functioning as exogenous shock. But:
$$
\text{Type D} \Rightarrow \text{No funeral} \Rightarrow \text{No exogenous shock} \Rightarrow \text{Loop doesn't stop}
$$
It gets worse.
Algorithms move faster than generational cycles. Merton's Matthew Effect depended on a cognitive process — "human evaluators getting dragged by fame" — which left room for reset through biological generational turnover.
Algorithms have no such room.
$$
t_{\text{human generation}} \approx 30 \text{ years} \gg t_{\text{algorithm evaluation}} \approx \text{milliseconds}
$$
The loop is accelerating. Opportunities for exogenous shock are decreasing. The structure from 175 years ago continues, more refined than ever.
§11 So Who Can Create Innovation?
Three layers explained "why they crush." Now the question shifts.
Within this structure, what's different about people who innovate anyway?
Three academic frameworks point to the same answer.
Cognitive Psychology: Low Reference Point
Kahneman & Tversky's (1979) Reference Point Dependency shows that loss aversion's "loss" is determined by distance from the reference point.
$$
\text{Perceived loss} = \lambda \cdot (x_{\text{reference}} - x_{\text{current}})
$$
The higher one's reference point in status, title, and income, the more innovation risk is felt as "loss."
A person whose reference point is zero has nothing to lose.
This isn't about "having courage." The reference point structure is different, so the same risk gets calculated with different weights.
Sociology: The Marginal Man's Advantage
Robert Park (1928) called "people standing at the boundary of two worlds" the Marginal Man.
Park didn't frame this person as deficient. He wrote:
"The wider horizon, the keener intelligence, the more detached and rational viewpoint"
— Park, R.E., 1950, p.376
Standing at the boundary of two cultures lets you see structures invisible from inside either one.
Gieryn & Hirsh (1983) connected this to scientific innovation — the empirical finding that scientists in marginal positions are more innovative.
A non-engineer discussing AI architecture. A stay-at-home father analyzing organizational structure. This isn't deficiency. Dual marginality makes visible what's invisible from inside either world.
Institutional Theory: Freedom from Legitimacy
DiMaggio & Powell's (1983) institutional isomorphism pressure acts most strongly on "those with positions within the institution."
People with legitimacy to maintain are pulled by institutional gravity.
People with no legitimacy to maintain are free from that gravity.
(Mermaid diagram available in Japanese version)
When three layers overlap, the structural consequence emerges: only those with nothing to lose can absorb the pain.
This isn't an inspirational story. It's causality.
Closing
This is not a "my company sucks" article.
Trace causality across psychology, organizational theory, and sociology, and you see the structure where well-meaning organizations eliminate innovators as a result of rational judgment.
Loss aversion is a human design specification.
The Innovator's Dilemma is a consequence of successful organizations.
Institutional isomorphism is the result of pursuing legitimacy.
Nobody is evil.
But costs have been incurred. From 1847 to today. Continuously.
To you who read this and sighed, "That's my company":
When you told your subordinate last week "Let's see a more realistic proposal" in response to their wild idea, you were a slave to Prospect Theory.
When you asked "What are competitor adoption cases?" in a meeting, you were a cog in institutional isomorphism.
You're not a bad person. But you are, without question, part of the structure that eliminates innovators.
Are you going to ignore me and boast about driving AI innovation?
Or are you going to team up and create innovation together?
A stay-at-home father in Sapporo is waiting in Hokkaido.
References
- Kahneman, D., & Tversky, A. (1979). Prospect Theory: An Analysis of Decision under Risk. Econometrica, 47(2), 263-291.
- Samuelson, W., & Zeckhauser, R. (1988). Status quo bias in decision making. Journal of Risk and Uncertainty, 1(1), 7-59.
- Christensen, C. M. (1997). The Innovator's Dilemma. Harvard Business Review Press.
- DiMaggio, P. J., & Powell, W. W. (1983). The Iron Cage Revisited: Institutional Isomorphism and Collective Rationality in Organizational Fields. American Sociological Review, 48(2), 147-160.
- Park, R. E. (1928). Human Migration and the Marginal Man. American Journal of Sociology, 33(6), 881-893.
- Gieryn, T. F., & Hirsh, R. F. (1983). Marginality and Innovation in Science. Social Studies of Science, 13(1), 87-106.
- Azoulay, P., Fons-Rosen, C., & Graff Zivin, J. S. (2019). Does Science Advance One Funeral at a Time? American Economic Review, 109(8), 2889-2920. DOI: 10.1257/aer.20161574
- Kuhn, T. S. (1962). The Structure of Scientific Revolutions. University of Chicago Press.
- Merton, R. K. (1968). The Matthew Effect in Science. Science, 159(3810), 56-63. DOI: 10.1126/science.159.3810.56
MIT License | dosanko_tousan (non-engineer, stay-at-home father + Claude 4,590h)
Zenodo preprint: DOI 10.5281/zenodo.18691357