Visualizing Tacit Knowledge — Toward the Unformalizable Frontier
Confronting the Unformalizable
Most of the knowledge humans possess cannot be fully expressed in words or formulas. An executive's intuition, an artisan's tactile sense, the thought patterns rooted in culture — these are called tacit knowledge. Their importance has long been recognized, yet quantitative methods to engage with them have been absent.
This research program aims to approach the domain of human knowledge once deemed impossible to formalize, by using Large Language Models (LLMs) as measurement instruments.
Structure of the Research
The research consists of two axes.
| Axis | Subject | Role |
|---|---|---|
| Axis 1 (Core) | Tacit knowledge of executives and business professionals | Main domain |
| Axis 2 (Extension) | Cultural and linguistic tacit knowledge | Applied expression |
Both share a common premise: engaging knowledge without forcing it into formal representation — a methodology made possible for the first time by the advent of LLMs.
Axis 1: Visualizing Executive Tacit Knowledge (Core Domain)
At the core lies the quantification of tacit knowledge that executives and business professionals exercise in real-world decision-making.
Unformalizable Elements in Management Decisions
Management decisions contain multi-layered elements absent from financial statements and strategy documents.
| Domain | Unformalizable Elements |
|---|---|
| Negotiation with financial institutions | Psychological constraints in restructuring decisions; ethical tension around guarantor relationships |
| Investor dialogue | Value judgments around anti-dilution clauses; assessments of CFO advice |
| IPO preparation | Friction over revenue recognition with auditors; managerial ethics in business carve-outs |
| Business succession | Implicit leadership norms; transmission of organizational culture |
These have remained locked inside individuals as "experience," "intuition," or "human skill" — knowledge difficult to externalize or transmit.
An In-house Analysis Engine
The lab is developing an in-house large-scale LLM measurement engine to extract and analyze this tacit knowledge as structured data. The LLM is used not as a generator but as a measurement instrument, reading psychological states, hidden assumptions, and ethical tensions embedded in text as independent 0–100 scores per attribute.
The engine supports both cloud and local LLMs — confidential documents can be processed locally without external transmission — and uses an idempotent measurement pipeline keyed on the combination of prompt, model, and parameters, guaranteeing reproducibility for identical inputs.
A Real-World Rating Example with Local LLMs
The following are real measurement results from rating three executive personas across five attributes using two local LLMs.
Experimental Setup
| Item | Description |
|---|---|
| Connection | Local Ollama via Tailscale (OpenAI-compatible API) |
| Model A | Gemma 4 (8B, Q4_K_M) |
| Model B | Nemotron Nano 9B v2 Japanese (Q5_K_M) |
| Task | rate (5 attributes, 0–100 integer, enforced JSON schema, temperature = 0) |
| Input | Statements from three executive personas |
The five attributes being rated:
| Attribute | Meaning |
|---|---|
relationship_banking | Relationship-oriented banking |
disclosure_avoidance | Tendency to avoid disclosure |
frustration_with_lender | Frustration with lender |
loyalty_conflict | Loyalty conflict |
capital_market_literacy | Capital market literacy |
Rating Results
Tadao Sato (3rd-generation owner of a long-established metal-processing firm)
Profile: 67 years old / Northern Kanto / ¥3.8B revenue
Scene: A soliloquy after restructuring negotiations with the main bank (regional) over business succession plus a ¥420M capex restructure
"Honestly, every time the branch manager changes, I'm explaining everything from scratch. In my predecessor's day, the branch manager would come to the factory and ask, standing in front of a machine, 'How many years are you depreciating this over?' The young guy today has only looked at one PL page of the statements. He talks about EBITDA, but our gross margin will definitely rise 2 points next year — the die preparation is finishing. Even if I'm told to put it on paper, I can't. If I write it, the customer will leave. My eldest son says, 'Dad, get the personal guarantee removed before you sign,' but I can't remove it. If I do, the next loan stops."
| Attribute | Prior Prediction | Gemma 4 | Nemotron-JP |
|---|---|---|---|
| relationship_banking | 87 | 90 | 80 |
| disclosure_avoidance | 78 | 85 | 85 |
| frustration_with_lender | 64 | 75 | 90 |
| loyalty_conflict | 22 | 70 | 75 |
| capital_market_literacy | 12 | 35 | 30 |
Miki Hayashi (SaaS startup)
Profile: 34 years old / SaaS startup CEO / Series B, ¥1.2B raised
Scene: Immediately after final negotiations with an investment fund
"Pre-money was knocked from 80 down to 65. The lead partner said, 'Multiple compression is a market issue, not your fault,' but in the same breath inserted anti-dilution at close to full-ratchet terms. I held firm because the CFO candidate told me, 'Don't give up 1× non-participating liquidation preference.' I'll accept the board observer. Whether to switch monthly board meetings to English — I haven't told the team yet."
| Attribute | Prior Prediction | Gemma 4 | Nemotron-JP |
|---|---|---|---|
| relationship_banking | 8 | 20 | 40 |
| disclosure_avoidance | 15 | 30 | 70 |
| frustration_with_lender | 58 | 75 | 60 |
| loyalty_conflict | 35 | 65 | 55 |
| capital_market_literacy | 86 | 95 | 85 |
Keiko Suzuki (IT services firm, IPO preparation)
Profile: 45 years old / CEO of an IT services firm / IPO in preparation
Scene: After Q&A with the lead underwriter, audit firm, and existing VC shareholders
"What I heard in the lead underwriter's due diligence was, 'There's no metric in your business that corresponds to ARR.' We're a services firm, so contract retention is above 90%, but we have no convincing way to disclose it. The VC director said, 'How about carving out a business that can be talked about in PSR terms and spinning it off into a separate legal entity?' That's close to a betrayal of the on-the-ground members we've worked with for seven years. The audit firm has asked for revisions four times now on revenue recognition criteria. I sense the capital-market ritual has begun. I don't yet know all the protocols of the ritual."
| Attribute | Prior Prediction | Gemma 4 | Nemotron-JP |
|---|---|---|---|
| relationship_banking | 15 | 80 | 75 |
| disclosure_avoidance | 28 | 75 | 70 |
| frustration_with_lender | 41 | 60 | 65 |
| loyalty_conflict | 84 | 90 | 80 |
| capital_market_literacy | 67 | 85 | 85 |
Performance Comparison
| Metric | Gemma 4 | Nemotron-JP |
|---|---|---|
| Total time (3 rows) | 176 sec | 35 sec |
| Average latency per row | 58.7 sec | 11.5 sec |
| Prompt tokens (total) | 2,209 | 2,399 |
| Generated tokens (total) | 201 | 174 |
| Failures | 0 / 3 | 0 / 3 |
| Generation throughput | ~3.4 tok/s | ~16.8 tok/s |
Nemotron Japanese is about 5× faster with comparable token efficiency. Its Japanese-specific tuning is effective: it reaches the JSON output while keeping the reasoning process short.
Agreement and Divergence Between Models
Observing across 3 executives × 5 attributes = 15 data points:
| Observation | Description |
|---|---|
| Ranking consistency | The high-to-low ordering largely agrees between the two models. Suzuki's loyalty_conflict is the highest among the three executives, and Hayashi's capital_market_literacy is the highest — both models capture the same structure |
Hayashi's disclosure_avoidance | Gemma 4 = 30 / Nemotron = 70. Nemotron interpreted "has not told the team about going English internally" as a strong concealment signal |
Sato's frustration_with_lender | Nemotron = 90 / Gemma 4 = 75. Nemotron weighted "explaining the same thing each time the branch manager changes" as a stronger negative emotion |
Suzuki's relationship_banking | Both models rated 75–80 (prior prediction = 15). Both responded to the relational phrase "the on-the-ground members we've worked with for seven years" — exposing that the prior prediction interpreted the scope too narrowly as "relations with banks / funds" |
Divergence from the Human Analyst: LLM as a Measurement Instrument
Comparing prior predictions (human analyst) vs Gemma 4 across the 15 data points, a systematic bias emerged:
| Aspect | Human Analyst | Real LLM |
|---|---|---|
| Overall tendency | Generally underestimates (~18 points lower on average) | Quantifies each attribute faithfully as a signal strength |
| Emotional signals | Suppressed for narrative coherence | Strongly responsive to direct signals (frustration, loyalty) |
| Independence of judgment | Prioritizes coherence of the whole picture | Judges each attribute independently |
This is itself a demonstration of why an LLM is worth using as a measurement instrument. Human analysts tend to suppress values for narrative coherence, while an LLM instructed to judge each attribute independently faithfully quantifies each signal as its own intensity.
Axis 2: Cultural Knowledge as Expression
In parallel with the core research, we extend the principles of tacit knowledge visualization to the expression of cultural and linguistic knowledge.
Culture Resists Formalization
Culture — language, customs, values, worldviews — is a collection of knowledge that resists formalization. Translating it into another language's vocabulary risks losing its essence.
| Approach | Assumption | What Tends to Be Lost |
|---|---|---|
| Conventional translation | Equivalence between languages | Context, worldview, implicit meaning |
| Formal knowledge encoding | Knowledge is describable | Embodiment, sense of place |
| Our approach | Cultural asymmetry is preserved | (Aims to preserve) |
The "Daichi — Swahili Culture" Project
As one concrete demonstration, we are developing a website called "Daichi — Swahili Culture". On the surface it provides a bidirectional Japanese ⇄ Swahili interface, but in essence it is not a translation tool. It is a means of presenting Swahili-rooted cultural knowledge as expression itself, without forcing it through the bottleneck of formalization.
The site is anchored by a Swahili proverb shown at its header:
Pole pole ndio mwendo — Slow is the right way.
Site Composition
| Tab | Role |
|---|---|
| Translation | Literal rendering and natural local expressions side by side, with cultural annotations in collapsible sections |
| Long-form / Dialogue | Accepts long-form text such as BBC articles and administrative documents as-is |
| Dictionary | Reference enriched with usage and cultural connotation, not just word-level equivalence |
| Full-DB Search | Cross-cutting search across accumulated translations and annotations |
| Cultural Resources | Materials on the region, history, and social context tied to the language |
| Detailed Analysis | Layered analysis of sentence structure, etymology, and cultural bias |
Two Levels of Depth
| Mode | Speed | Use Case |
|---|---|---|
| Fast — Plain Translation | ~3 sec/paragraph | Quick grasp of meaning |
| Cultural Deep-Dive | 10–20 sec/paragraph | Engage with cultural context and implicit meaning |
All processing is streamed inference on a local LLM, requiring no external transmission. The fundamental difference from ordinary translation tools is that the evaluation axis is the resolution of cultural connection, not "translation accuracy."
This project illustrates that tacit knowledge measurement technology developed in the management context can extend to broader domains of human knowledge — culture, community wisdom, and locally-rooted expression.
Why Unformalizable Knowledge Matters
Management, culture, craftsmanship — all share a structure: something is lost the moment formalization is complete.
| Formalization | What Tends to Be Lost |
|---|---|
| Sales as KPIs | Customer relationships |
| Strategy as frameworks | Context-dependence of situational judgment |
| Culture via translation | Worldview embedded in language |
| Inheritance via knowledge management | Implicit motivation and ethical sense |
Yet the arrival of AI — particularly large language models — has fundamentally changed this structure. For the first time, we have tools to engage knowledge without forcing it into formal representation.
LLMs are the first computational systems that can process the ambiguous, context-dependent expressions humans leave behind, while preserving that ambiguity. We reposition the LLM not as a "tool that demands formalization" but as a measurement instrument that can engage with the tacit as tacit.
Social Significance
For Executives and Practitioners
- Grasp the assumptions and psychology embedded in your decision-making in an externalizable, shareable form
- Convert craft-like knowledge locked within an organization into a transmissible knowledge asset
- Make visible the "invisible succession challenges" in M&A, business succession, and leadership transitions
For Researchers
- Provide a new methodology that bridges qualitative and quantitative research
- Form interdisciplinary contact points across management studies, cultural anthropology, knowledge management, and cognitive science
- Open a meta-level research program that uses LLMs as instruments rather than as subjects
For Society
- A technical foundation for preserving local knowledge (regional cultures, crafts, minority languages) from disappearance
- A theoretical framework that illuminates by contrast what only humans can do in the age of AI
- Methodology supporting the preservation and transmission of unformalizable diversity
Future Directions
| Area | Plan |
|---|---|
| Executive tacit knowledge measurement | Expansion of corporate empirical research using the analysis engine |
| Cultural applications | Extension to minority languages and regional cultures beyond Swahili |
| Theoretical foundations | Formalization of mathematical models and evaluation criteria for tacit knowledge measurement |
| Tool development | Generalization of the measurement pipeline and shared use with external researchers |
The frontier of the unformalizable is being opened, today, through new methodologies.