Bonnō Embedding and the Scam-Detection Mechanism: Mapping Text into a Vector Space of Psychological Vulnerability

Note on terminology. “Mental afflictions” in this essay corresponds to the Buddhist concept of bonnō (煩悩) — in Sanskrit, kleshas — the mental states held to cause human suffering, classified into 108 categories. The translation of Buddhist terminology in this essay is preliminary; I plan to refine it in v0.2 in consultation with Buddhist scholars.

Introduction

What follows is my rewriting of §14 of the Concept Record — the founding document of the research program I am running at Mindseed Research, “Bonnō × Scam-Virus Mapping: An Integrated Database” — into a standalone piece. My intention is for the essay to stand on its own, without requiring readers to be familiar with the other chapters; the background I need is given in the body.

I have only one thesis. Scam texts carry, in vector space, signatures that target human psychological vulnerability — mental afflictions, in Buddhist terms. By constructing a function f that maps text into a 108-dimensional vector of affliction stimulation, and by judging composite-style across (a) the characteristic shape of that vector and (b) a law-grounded check on source attribution, it becomes possible to detect, at the abstract level and with stable accuracy, fraud that surface-level keyword matching cannot reach. I write this both as an application of Affective Computing research and as a technical framework that, I believe, can be immediately useful to consumer-protection and anti-scam practitioners.

About the author¹: I run Mindseed Research as an independent researcher based in Wakayama, Japan. Forty-plus years of infrastructure engineering (Kansai Electric Power Company, Samsung SDS) precede the current 20-year research program centered on Bonnō × AI. See the About page for my background.

1. The definition of the affliction-embedding space

I define affliction embedding as the following three-way relation:

f: (T, C, U) → V = (w₁, w₂, ..., w₁₀₈),  wᵢ ∈ [0, 1]

  T: an arbitrary text
  C: situation / context (time of day, medium, relationship, cumulative behavior, etc.)
  U: the individual's mental-affliction sensitivity profile
  V: a 108-dimensional affliction-stimulation vector

Here wᵢ represents “how strongly text T, in situation C, stimulates affliction i of individual U.” The value range is normalized to [0, 1].

This formulation is positioned in the lineage of embedding-representation learning as developed in Word2Vec, BERT, CLIP, and others. What I see as distinctive, though, is that the meaning of the embedding space corresponds directly to “human psychological vulnerability.” Conventional embeddings learn things like “semantic similarity of words” or “correspondence between images and text”; what I want to learn here is “the action on human emotion.”

Why 108 dimensions? Because, as I read the Buddhist canon, that is the empirical granularity of partition reached over 2,500 years of internal observation. For the methodological argument, see the companion essay Buddhism as a Coordinate System. As the practical implementation starting point, I begin from the “three poisons × three subcategories each = 9-dimensional minimum model” laid out in §0.4.5 / §4.3.0 of the Concept Record.

2. The operational definition of scam judgment

For an arbitrary text T, I decompose the scam probability P(scam | T) through the affliction vector V like this:

P(scam | T, C, U) = g(V(T, C, U), C(T), S(T))

  V(T, C, U): the affliction-stimulation vector
  C(T):       contextual information about the text (medium, situation, counterparty)
  S(T):       information about the source / sender
  g:          the integrating decision function

The core claim, in this decomposition, is that scam judgment is reduced to a pattern-recognition problem in vector space. What remains is the question of how to characterize V, C, and S concretely. The next section addresses V; §4 addresses S.

3. Three signatures of the scam vector

From empirical observation and theoretical reasoning, I see three common features in the affliction vectors of scam texts. Together, they give us the leverage we need to separate scam texts from legitimate ones.

Signature 1: concentrated stimulation of a few specific afflictions

A legitimate text stimulates several afflictions mildly. A scam text stimulates a few specific afflictions intensely:

A legitimate-text vector:
  affliction A: 0.3, affliction B: 0.2, affliction C: 0.4, ...  uniformly low–medium

A scam-text vector:
  affliction A (greed)   : 0.92  ★★★
  affliction B (loneliness): 0.78  ★★★
  affliction C (urgency) : 0.85  ★★★
  affliction X (the rest): 0.05–0.10

Compute the standard deviation of the vector and the scam case is plainly higher. Simple, but easy to deploy in practice as a feature.

Signature 2: suppression of cautionary afflictions

A scam text, simultaneously with stimulation, is designed to suppress the afflictions related to caution:

Typical examples of affliction suppression (negative stimulation):
  - Doubt:                "trust me," "keep this between us"
  - Caution:              "decide now or it's over"
  - Confirmation impulse: "you shouldn't consult your family"

This is a “suppression structure” not seen in legitimate advertising. I treat it as a decisive feature of fraud. The detection side has to learn the co-occurrence pattern of stimulation and suppression, not just stimulation intensity in isolation.

Signature 3: coexistence of mutually incompatible appeals

“Easy money” (stimulating greed) + “no risk” (suppressing fear) + “special access just for you” (stimulating sociality) — scams routinely place logically incompatible appeals side by side. Combinations that would not stand up in legitimate economic activity show up consistently in fraud.

None of these three signatures, alone, cleanly separates fraud from legitimate advertising. Only by combining them composite-style do we reach into the territory that surface-level keyword matching never touches. That, at least, is my reading.

4. The source-attribution principle — a law-grounded composite axis

Pure affliction vectors alone cannot cleanly separate scams from legitimate advertising — because legitimate advertising also stimulates afflictions. Cosmetics legitimately stimulate vanity, diet products legitimately stimulate anxiety about the body, investment products legitimately stimulate desire for gain.

The second axis I want to combine here is source attribution, grounded in law.

In Japan:

• Act against Unjustifiable Premiums and Misleading Representations: obligation to identify the advertiser
• Act on Specified Commercial Transactions: obligation to display business information
• 2023 (Reiwa 5) revision on stealth-marketing regulation: obligation to disclose that content is an advertisement

In other words, legitimate advertising is legally obligated to disclose its source. Stated as a decision rule:

[source clearly disclosed] + [strong affliction stimulation]  →  legitimate advertising
[source unclear]           + [strong affliction stimulation]  →  very likely a scam

My core view is that this is a far more powerful discrimination axis than affliction vectors alone. The concrete source-attribution checks look like this:

Item	Scam pattern
Company name	none, or fabricated
Corporate registration number	none
Address	none, or overseas (US / UK / Hong Kong feigned)
Contact	LINE / Telegram only, no phone number
FSA registration number	none, or forged
Website	hastily set up, no operator info
Privacy policy	none, or machine-translated
Specified Commercial Transactions disclosure	none

These absences are themselves violations of law, so the detection logic has legal grounding. When a detected fraud later needs to be brought to enforcement or administrative guidance, the evidence of missing source attribution is directly usable. My program performs affliction-stimulation analysis and source-attribution analysis in tandem — a design that minimizes false positives while keeping the detection rationale legally explicable.

5. The strategic advantage at the abstract level — robustness to changing techniques

The strategic advantage of the affliction-vectorization approach, in my view, is that it fights one level of abstraction higher:

[concrete level]   the perpetrators change techniques every month
  ↓
[middle level]     paraphrasing, new names, new platforms
  ↓
[abstract level ← this approach] the afflictions targeted cannot be changed

However the concrete techniques evolve, the structure of exploiting human vulnerability remains invariant. Capturing that invariance, I believe, makes it possible to build a detection system that holds up against unknown techniques.

This is a textbook instance of what I think of as “combining mature logics”:

• Buddhist affliction theory (2,500 years)
• Source-attribution obligation in law (decades)
• Pattern-recognition engineering (70 years)

Integrating three mature logics, to take on a new problem (SNS-mediated investment fraud). The Samsung SDS philosophy I internalized — “mature technology is safety first” — aligns directly with this positioning.

As I argue in the companion essay Economies of Precision, the giants stay at surface-level keyword matching because of Economies-of-Scale incentives. My place is the opposite: deep, abstract-level analysis. This essay is the technical implementation of what I call Economies of Precision.

6. False-positive risk and countermeasures

If affliction vectorization becomes too strong, legitimate economic activity and ordinary communication can be misclassified as fraud. This is, realistically, one of the heaviest responsibilities a consumer-protection system bears. I want to address it with layered defense:

Countermeasure	Description
Source-attribution check (§4)	exclude legally registered sources
Context filter	distinguish public advertising from one-to-one DMs
Cumulative-behavior analysis	judge on a sequence of actions, not single ones
Transparent false-positive reporting	let the user push back with “this was a false positive”
Threshold tiers	three-level output: strong warning / caution / information
Thorough explanation	always present the reasoning behind the warning

The design principle I want to hold firm to is this: the output is a “warning,” not a “block.” Final judgment rests with the user. For details, see the companion essay The External Prefrontal Cortex (ExPFC). Once ExPFC is positioned as a device that substitutes for the “examination” module, the final decision-maker has to remain the user themselves.

7. Ethical questions — six of them

By this point, I expect it is clear that affliction-vectorization technology is double-edged. What can be used for detection can be used for attack. The ethical questions I see as continuing items for inquiry are six:

1. Misuse potential. A technology that understands the affliction-stimulation structure of text can also become a tool for crafting more sophisticated scams. This is the defender’s dilemma itself.
2. Manipulation detection vs freedom of expression. Where does “fraud” end and “strong opinion” begin? The boundary is context-dependent, and machine judgment alone cannot draw it.
3. Cultural bias. Does the Buddhist taxonomy of afflictions function appropriately in non-Buddhist cultures? The cultural validity of the coordinate system bears on the entire program.
4. Personal privacy. The ethics of subjecting individual speech to affliction analysis. Whose speech, for whose sake, to what depth?
5. Value judgment by AI. Is it right for an AI to tell you “you are being stimulated to greed”? The philosophical implications of delegating value judgment to AI.
6. Risk of dependence. Do users abdicate their own judgment? This is isomorphic to the question raised in the companion essay §13 ExPFC.

These are continuing considerations. I plan to develop ethical guidelines in parallel with development. This essay only raises the questions; the answers will be sharpened through dialogue with collaborators. The points here are related to §10.3 (general / cross-cutting issues) and §13.11 (ExPFC-specific ethics) in the Concept Record, and will eventually be consolidated into a single ethical guideline.

8. Conclusion

Let me restate my thesis one more time, in my own words.

Scam texts carry, in vector space, signatures that target human psychological vulnerability — mental afflictions. By constructing a function f that maps text into a 108-dimensional affliction-stimulation vector, and by judging composite-style across (a) the characteristic shape of that vector — concentrated stimulation + suppression of cautionary afflictions + coexistence of incompatible appeals — and (b) a law-grounded check on source attribution, we can reach, at the abstract level and stably, into the territory that surface-level keyword matching cannot cover. That is the core of my detection mechanism.

The position of fighting at the abstract level is of a piece with the strategic argument of the companion essay Economies of Precision. And this detection mechanism is positioned as the concrete implementation device of what is discussed in the companion essay The External Prefrontal Cortex (ExPFC):

“Map text into an affliction vector; from its shape, make the action on psychological vulnerability visible; and substitute, externally, for the user’s ‘examination’ module (the prefrontal-cortex function).”

This is the core technology of my research program. To Affective Computing researchers, I offer this as the proposal of a new embedding space — one that learns the action on human emotion. To anti-scam practitioners, I offer it as a design guide for a detection mechanism that holds up, stably, at the abstract level. The two readings, in my mind, collapse into a single question: how far can the psychological vulnerability of a human being be described, machine-readably?

Big AI blocks scams with keywords. I capture the structure of scams with affliction vectors. Don’t look at the surface. Look at psychological vulnerability itself. That is the core of the scam detection I am aiming at.

---

Related materials

• Concept Record (full text) — including §14, the parent research program from which this essay is derived
• Concept Record §17: ExPFC Core Specification v0.1 — the formalization of the function f and three-signature detection of this essay as a core specification (API contracts, JSON Schemas, Python reference implementation)
• Companion essay: §0 Buddhism as a Coordinate System — the methodological tier that backs the 108-dimensional coordinate system (Level 1)
• Companion essay: §13 The External Prefrontal Cortex (ExPFC) — the device that this detection mechanism drives (neuroethics + AI Safety)
• Companion essay: §15 Economies of Precision vs Economies of Scale — the strategic tier supporting the fight at the abstract level (Level 2)
• 日本語版
• PYOL Mind Mirror — a fraud-awareness tool I run in production, serving as the implementation foundation for this detection mechanism

Contact

If anything in this essay resonates with you — as a researcher, a critic, or a possible successor — I would welcome hearing from you. Constructive critique or collaboration proposals from Affective Computing, AI Safety, anti-scam practice, consumer protection, law, or Buddhist studies are all welcome via the contact page. English and Japanese are both fine.

References

1. Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT ‘21), 610–623.
2. Carroll, M., Chan, A., Ashton, H., & Krueger, D. (2023). Characterizing manipulation from AI systems. Proceedings of the 3rd ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization (EAAMO ‘23).
3. Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019). BERT: Pre-training of deep bidirectional transformers for language understanding. NAACL-HLT 2019, 4171–4186.
4. Marcus, G., & Davis, E. (2019). Rebooting AI: Building Artificial Intelligence We Can Trust. Pantheon.
5. Mikolov, T., Sutskever, I., Chen, K., Corrado, G., & Dean, J. (2013). Distributed representations of words and phrases and their compositionality. Advances in Neural Information Processing Systems, 26.
6. National Police Agency of Japan. (2025). Statistics on SNS-mediated investment and romance fraud, 2023–2025.
7. Park, P. S., Goldstein, S., O’Gara, A., Chen, M., & Hendrycks, D. (2024). AI deception: A survey of examples, risks, and potential solutions. Patterns, 5(5), 100988.
8. Picard, R. W. (1997). Affective Computing. MIT Press.
9. Radford, A., et al. (2021). Learning transferable visual models from natural language supervision. Proceedings of the 38th International Conference on Machine Learning, 8748–8763.
10. Whitty, M. T. (2013). The scammers persuasive techniques model: Development of a stage model to explain the online dating romance scam. British Journal of Criminology, 53(4), 665–684.

Footnotes

1. See the About page linked above. ↩

@misc{matsuura2026bonnoEmbedding,
  author       = {Toshinobu Matsuura},
  title        = {Bonnō Embedding and the Scam-Detection Mechanism: Mapping Text into a Vector Space of Psychological Vulnerability},
  howpublished = {Mindseed Research},
  year         = {2026},
  month        = {May},
  url          = {https://research.pyol.net/en/essays/bonno-embedding/}
}

Matsuura, T. (2026, May 18). Bonnō Embedding and the Scam-Detection Mechanism: Mapping Text into a Vector Space of Psychological Vulnerability. Mindseed Research. https://research.pyol.net/en/essays/bonno-embedding/

Matsuura, Toshinobu. "Bonnō Embedding and the Scam-Detection Mechanism: Mapping Text into a Vector Space of Psychological Vulnerability." Mindseed Research, May 18, 2026. https://research.pyol.net/en/essays/bonno-embedding/.