Designing for
the Unspoken

Tacit Knowledge

Michael Polanyi (1966) defined tacit knowledge as understanding that is "inherently difficult to articulate," summarized in his aphorism:

"We can know more than we can tell."

It encompasses know-how, mental models, and intuitions gained through action and experience.

Situated Learning

Brown, Collins & Duguid (1989) established that knowledge is inseparable from the activity, context, and culture in which it develops.

The "Master-Apprentice" model depends on legitimate peripheral participation: novices learn by working alongside experts in authentic contexts.

Information Overload

Miller (1956) established that human working memory can hold approximately 7±2 items simultaneously.

When data volume exceeds this capacity, processing degrades. Modern intelligence environments routinely exceed these limits, creating the "cognitive bottleneck."

Deep Smarts (Leonard, 2005)

Expert intuition built through years of experience: pattern recognition, judgment under uncertainty, and contextual awareness.

Unlike explicit knowledge in manuals, deep smarts are transferred through guided experience and Socratic dialogue.

Inheritance Imperative

Hang & Zhang (2024) argue that when experts retire, organizations face a "knowledge inheritance crisis."

Departing expert knowledge is organizational inheritance that must be actively claimed through structured protocols.

Recognition-Primed Decision

Klein (1993) showed that experts do not deliberate options; they recognize patterns.

TacitFlow preserves these pattern libraries by capturing the cues experts notice.

Workforce Crisis

• OECD (2025): 40% of civil servants report burnout; 13% intend to leave within 12 months.
• Europol (2025): Urgent need for "state-of-the-art analytical competence" to address capability gaps.

Hallucination
LLMs generate plausible-sounding but factually incorrect information because they predict tokens based on statistical patterns rather than verified facts. TacitFlow mitigates this via Retrieval-Augmented Generation (RAG). (Lewis et al., 2020)

Groundedness
The property ensuring AI outputs are anchored in verifiable source material. Unlike creative text generation, grounded reasoning requires every claim to trace back to documented evidence.

Provenance
W3C PROV-O (2013) provides the standard vocabulary for expressing provenance. TacitFlow uses PROV-O to maintain digital chain-of-custody, tracking who generated information, from what sources, through which processes.

Publication Strategy
The research follows a multi-pronged approach: establishing the theoretical gap (Scoping Review), validating the problem (Empirical Study), and proposing the solution (Technical Report).

Venues
Targeting high-impact journals in Knowledge Management and AI to bridge the gap between social science theory and technical implementation.

Research Design
This study employs Design-Based Research (DBR), iterating between theory and practice to develop a working artifact. DBR bridges the gap between laboratory findings and real-world practice. (Wang & Hannafin, 2005)

Emergent Inquiry
RQ4 is an "emergent" question. In DBR, the intervention itself (TacitFlow) acts as a hypothesis. The study tests if the existence of such a tool changes the nature of the problem.

Research Through Design
Artifacts embody theoretical propositions. The prototype becomes a vehicle for testing ideas rather than merely an end product. (Zimmerman, Forlizzi & Evenson, 2007)

Software as Hypothesis
Building systems specifically to validate theoretical constructs. TacitFlow is not a commercial product but a research instrument for testing the EASCI framework. (Leinonen et al., 2008)

Methodology

Arksey & O'Malley (2005) established the framework for scoping reviews, distinguishing them from systematic reviews by their broader exploratory purpose. Scoping reviews map key concepts, evidence types, and research gaps rather than synthesizing effect sizes.

PRISMA-ScR

Tricco et al. (2018) extended PRISMA reporting guidelines specifically for scoping reviews, adding items on rationale for review type selection and deviation from protocol. This enhances transparency and reproducibility in exploratory evidence synthesis.

The Externalization Gap

The SECI model assumes externalization is straightforward. The review found the opposite: converting tacit knowledge to explicit form is the primary bottleneck in knowledge transfer.

Key Insight

A gap exists between how people share knowledge in practice (socially) and how technology supports it (structurally).

Detailed Findings

Most workplaces rely on mentoring, storytelling, and job shadowing. AI or digital tools for experience sharing remain rare. While socialization methods are common, they rarely succeed in making tacit knowledge explicit.

Thematic Analysis

Braun & Clarke (2006) established thematic analysis as a qualitative method. Their six-phase process (familiarization, coding, theme generation, review, definition, write-up) enables systematic pattern identification while remaining accessible. Unlike grounded theory, it permits both inductive and deductive coding.

From Interviews to EASCI

This study revealed that tacit-to-explicit conversion does not happen in one leap (as SECI suggests). Interview data showed knowledge transfer requires multiple deliberate stages: initial recognition of experience, tentative articulation, structured formalization, social validation, and adaptive refinement.

Participating Organizations

• EASS: Estonian Academy of Security Sciences
• KRA: Defence Resources Agency
• MoD: Ministry of Defence
• PBGB: Police & Border Guard Board
• RIA: Information System Authority
• SMIT: IT & Development Centre

"Sticky" Tacit Knowledge

Von Hippel (1994) coined "sticky information" to describe knowledge costly to transfer. Tacit knowledge is deeply embedded in context, making it resistant to codification.

Current Mechanisms

Transfer relies on ad-hoc mentoring ("watch me do this"). This breaks down during rapid turnover or when key personnel leave unexpectedly.

Practitioner Interest

Participants expressed strong preference for tools making sharing easier. Voice-based AI assistants resonated well for quick, natural capture.

SECI Model Context

Nonaka & Takeuchi (1995) describe four knowledge conversion modes. While organizations excel at Socialization (mentoring), they struggle with Externalization (articulating concepts).

Convergent Validity

Literature and field data independently identify the same failure point, strengthening confidence in this finding.

Documented Loss

Cho et al. (2020) documented this in Security Operations Centers: tacit expertise remains "trapped in heads" until personnel depart, then vanishes.

Why Voice?

• Natural: Experts articulate experience through storytelling.
• Focus: Speaking allows capture without breaking concentration.
• Context: Voice preserves narrative structure and reasoning.

Software as Hypothesis

TacitFlow is a research instrument (Leinonen et al., 2008). It operationalizes concepts from knowledge management theory to allow empirical validation. The study tests: Can AI mediate the tacit-to-explicit conversion?

Workflow Integration

TacitFlow requires no separate login or forms. It integrates into existing work activities, capturing knowledge through passive listening or brief voice prompts.

Why SECI?

Nonaka and Takeuchi's The Knowledge-Creating Company (1995) is among the most cited works in management science. SECI has shaped knowledge management practice for three decades.

The Spiral Concept

SECI operates as a continuous spiral rather than a linear sequence. Knowledge cycles through all four modes repeatedly.

Academic Critiques

• Gourlay (2006): Externalization lacks operational definition. It's a metaphor without a mechanism.
• Tsoukas (2003): Tacit knowledge cannot be fully "converted". It remains context-dependent.
• Snowden (2002): SECI oversimplifies context, assuming universal applicability.

"We know more than we can tell."

— Michael Polanyi (1966)

Collins' Taxonomy (2010)

• Relational: Concealed by social dynamics.
• Somatic: Physically embodied skill.
• Collective: Distributed across a team.

Theoretical Limits

• Gourlay (2006): "Externalization" lacks operational definition.
• Tsoukas (2003): Context is inseparable from tacit knowledge.
• Von Hippel (1994): "Sticky information" resists transfer.

Abductive Reasoning

Peirce (1903): The process of forming explanatory hypotheses. Unlike deduction (proving) or induction (generalizing), abduction infers the best explanation for a specific situation.

Intuition Pumps

Dennett (2013): Thought experiments designed to elicit intuitions. They bypass the Polanyi Paradox by asking "what would you do if...?" rather than "what do you know?".

Key References

• Polanyi (1966): The Tacit Dimension.
• Gourlay (2006): "Conceptualizing Knowledge Creation."
• Busch (2008): "Tacit Knowledge in Organizational Learning."

The Socratic Method

Named after Socrates (470-399 BCE), the method uses guided questioning rather than direct instruction. The teacher asks probing questions that expose contradictions, forcing the learner to refine their understanding. TacitFlow's AI probes follow this tradition: asking questions that help experts articulate what they know.

Iterative Refinement

DBR's iterative cycles distinguish it from traditional experiments. Each TacitFlow iteration generates usage data that informs the next design phase, embodying DBR's commitment to theory-practice integration.

Cognitive Load Theory

Sweller (1988): Working memory has strict limits. TacitFlow's voice interface minimizes extraneous load (typing/navigating) to preserve germane load (reflection/articulation).

Recognition-Primed Decision

Klein (1993): Experienced practitioners recognize patterns rather than deliberating options. TacitFlow preserves these pattern libraries by capturing the cues experts notice.

Complex Thinking

Morin (2008): Complex phenomena require multiple perspectives. TacitFlow synthesizes pragmatism, process philosophy, organizational learning, and cognitive science.

Deep Smarts

Leonard (2005): Expert intuition built through years of experience. Unlike explicit knowledge in manuals, deep smarts are transferred through guided experience and Socratic dialogue.

Psychological Safety

Edmondson (1999): Essential for sharing. Experts must feel safe revealing their "tricks" without fear of judgment. TacitFlow addresses this through anonymous draft submission and mentor-mediated validation.

Mental Models

Internal representations (e.g., "suspicious behavior"). TacitFlow captures mental models through structured articulation and makes them visible for peer validation.

Avoiding Competency Traps

March (1991): Organizations that over-exploit existing knowledge fall into "competency traps." EASCI balances exploitation (Consolidation) with exploration (Innovation).

Process Philosophy

Whitehead: Actual Entities are "drops of experience." Concrescence is the integration of data into a new unity. In TacitFlow, each Knowledge Object is a concrescence of prior data, perspective, and AI structuring.

Collins' Taxonomy (2010)

Relational: Concealed by social dynamics.
Somatic: Embodied.
Collective: Social.
TacitFlow targets relational and somatic knowledge through conversational elicitation.

Experiential Learning

Kolb (1984): Concrete Experience → Reflective Observation → Abstract Conceptualization → Active Experimentation. EASCI maps directly onto this cycle.

Miller's Law (7±2)

Working memory holds ~7 chunks. EASCI's structured stages respect this limit by processing knowledge incrementally.

Curse of Expertise

Hinds (1999): Experts cannot "unlearn" expertise to imagine a beginner's perspective. TacitFlow's AI probes address this by asking the questions novices would ask.

Theoretical Backbone

This framework is the theoretical backbone that TacitFlow operationalizes. Each stage maps to specific system components and AI agents in the architecture.

Sensemaking (Weick, 1995)

Sensemaking is the retrospective process through which people construct "plausible accounts" of ambiguous situations. Unlike rational decision-making models, sensemaking acknowledges that meaning emerges through social interaction and narrative construction. TacitFlow's Consolidation stage operationalizes this by requiring peer dialogue before knowledge enters the canon.

Field Example

David Kolb's Experiential Learning

David Kolb (1939-present) developed the Experiential Learning Theory (ELT) synthesizing work by Dewey, Lewin, and Piaget. His four-stage cycle shows learning as a spiral: (1) Concrete Experience (doing), (2) Reflective Observation (watching and reflecting), (3) Abstract Conceptualization (thinking and concluding), (4) Active Experimentation (planning and trying). Each stage requires different abilities. EASCI's five stages map directly onto Kolb's cycle with Consolidation added for social validation.

Dave Snowden's Cynefin Framework

Dave Snowden developed the Cynefin (Welsh: "habitat") framework distinguishing five domains: Simple (clear cause-effect; best practice), Complicated (expertise needed; good practice), Complex (emergent; probe-sense-respond), Chaotic (no patterns; act-sense-respond), and Disorder (unknown domain). Snowden critiqued SECI for treating all knowledge as if it existed in the "Complicated" domain. TacitFlow is designed for the Complex domain where knowledge emerges through interaction rather than transfer.

Dynamic vs. Linear

SECI presents a linear spiral: Socialization → Externalization → Combination → Internalization. EASCI models knowledge as a complex adaptive system with multiple feedback loops. Innovation can trigger new Experience; Consolidation can require return to Articulation.

Process Philosophy View

Whitehead's (1929) "creative advance into novelty" reframes knowledge as flux, not static entity. What appear as stable organizations are merely "relatively stable" patterns emerging from underlying processes. EASCI embodies this dynamic view.

Process Philosophy Explained

Process philosophy treats reality as fundamentally composed of events and processes rather than static substances. Whitehead, its primary architect, argued that "actual occasions" of experience are the ultimate constituents of reality. For knowledge management, this means: knowledge is not a thing to be stored but an ongoing process of becoming. Each moment of knowing integrates past experiences into present understanding and propels toward future inquiry.

Pragmatism Explained

Pragmatism (Peirce, James, Dewey) judges ideas by their practical consequences rather than abstract truth. Knowledge is validated through action: "truth is what works." For EASCI, this means: a KO is valid if it enables effective action. The framework doesn't seek abstract truth but practical utility: knowledge that helps practitioners navigate real situations.

Why Five Stages?

Each stage transforms knowledge through specific cognitive and social mechanisms. Unlike SECI's four modes, EASCI separates Articulation from Structuring because they require different AI interventions: dialogue-based elicitation vs. abductive formalization.

Situated Learning

Brown & Duguid (1991) argue that canonical descriptions of work often diverge from actual practice. Real learning happens through legitimate peripheral participation within communities of practice.

Embodied Cognition

Merleau-Ponty (1945) argued knowledge is grounded in bodily experience. TacitFlow's Experience stage captures pre-reflective patterns through behavioral observation.

Maurice Merleau-Ponty (1908-1961)

French phenomenologist who established the body as the primary site of knowing. Against Cartesian dualism, he argued that perception is active embodied engagement. TacitFlow captures this through contextual metadata.

Phenomenology

Study of conscious experience from a first-person perspective (Husserl). It treats subjective experience as valid data, not bias to be eliminated.

Cartesian Dualism

Descartes' separation of mind and body. Merleau-Ponty's critique grounds tacit knowledge theory: knowing involves the whole embodied person.

Abductive Mapping

Peirce distinguished abduction (guessing best explanation) from deduction and induction. Structuring uses abduction to generate hypotheses from sparse evidence.

Exploration vs. Exploitation

March (1991): Organizations must balance efficiency (exploitation) with innovation (exploration). EASCI structures this tension.

Capture Phase in Practice

Communities of Practice

Lave & Wenger (1991): Knowledge transfer occurs through "legitimate peripheral participation". TacitFlow uses mentor-novice pairing.

Grounded Cognition

Barsalou (2008): Conceptual knowledge involves modal simulations. Expert judgment relies on "feeling" if a situation matches prior patterns.

Cognitive Apprenticeship

Collins et al. (1989): Modeling, coaching, scaffolding, and fading. AI probes function as scaffolding that fades as reasoning improves.

Reflective Practice Feedback Loop

Schön (1983): "Reflection-in-action". Innovation reveals gaps, prompting refinement of Articulation methods.

Pattern Recognition Feedback Loop

Simon (1996): Experts recognize patterns. Innovation helps identify recurring structures, improving Structuring.

Legitimate Peripheral Participation

Newcomers learn by participating at the edges of practice. TacitFlow operationalizes this through guided observation.

Bounded Rationality

Simon (1957): Humans use heuristics due to limited processing power. KOs provide pre-organized knowledge to reduce search costs.

Smart Forgetting Protocol

Innovation requires unlearning. This protocol ensures outdated intelligence (e.g., disproven hypotheses) is actively pruned to prevent "zombie facts."

Social Verification as Quality Gate

Von Krogh et al. (2000): Knowledge creation requires "enabling conditions" like trust and empathy. TacitFlow's mentor circles operationalize this, surfacing tacit disagreements to prevent hallucination loops.

Double-Loop Learning

Argyris & Schön (1978): Unlike single-loop learning (correcting errors), double-loop learning questions governing assumptions. It supports revising protocols when standard methods fail.

Social Learning Theory

Bandura (1977): Learning occurs through observation and modeling. TacitFlow's mentor circles create observational learning opportunities.

Observational Learning

Learning by watching others. Components: attention, retention, reproduction, motivation. TacitFlow makes validated expert KOs visible to learners.

Validate Phase in Practice

Embodied Knowledge Feedback Loop

Merleau-Ponty (1945): Structuring creates cognitive schemas that become the pre-reflective lens for future Experience.

Organizational Memory Feedback Loop

Walsh & Ungson (1991): Consolidated KOs become the institutional vocabulary that shapes how practitioners articulate new experiences.

Double-Loop Learning Feedback Loop

Innovation back to Experience reframes the entire knowledge capture process, creating fundamental shifts in practice.

Stable Tacit Practices

Practices are not static but actively performed. Consolidation creates "relative stability" where knowledge is routinized yet subject to re-enactment.

Destabilization Events

External shocks or anomalies force a shift from unreflective habit to active sensemaking (Dewey's "problematic situation").

Metacognition

Flavell (1979): "Thinking about thinking." TacitFlow's reflective probes develop awareness of one's own cognitive processes.

Lures for Feeling

Whitehead: Validated KOs become "lures" that shape how practitioners perceive future situations.

Resilience Engineering

Hollnagel (2011): Innovation enables adaptive capacity by updating mental models based on accumulated experience.

Self-Determination Theory

Deci & Ryan (2000): Intrinsic motivation depends on autonomy, competence, and relatedness. TacitFlow's voice-first design supports personal agency (choosing what to share) rather than forcing structured input.

Self-Efficacy

Bandura (1977): Belief in one's ability to succeed. TacitFlow builds efficacy through structured success experiences and visible attribution.

Paralinguistic Information

Voice captures hesitation, emphasis, and tone: metadata for tacit judgment. Polanyi (1966) noted experts often "feel" rightness before they can explain it. Audio preserves these signals; text discards them.

Contextual Metadata

Data about data capturing creation circumstances. TacitFlow captures:

• Temporal: Timestamp, duration.
• Spatial: GPS, zone.
• Social: Presence, supervisor.
• Operational: Case #, incident type.
• Physiological: Heart rate (stress).

Situated Interaction

Brown, Collins & Duguid (1989): Knowledge is inseparable from the activity and context in which it develops.

Subtle Cues

Audio captures confidence or uncertainty. Experts often "feel" rightness before they can explain it.

Knowledge Object Structure

Each KO contains:

1. Claim: the asserted knowledge statement
2. Evidence: supporting sources and observations
3. Confidence: weighted score based on evidence quality
4. Provenance: full derivation chain per W3C PROV-O
5. Metadata: timestamps, classification labels, contributor IDs
6. Relations: typed links to other KOs in the graph

PROV-O Vocabulary

PROV-O (W3C, 2013) is the OWL representation of PROV-DM. TacitFlow serializes all KO provenance as PROV-O-compliant RDF, enabling interoperability with any PROV-aware system.

From "Sticky" to Traceable

Von Hippel (1994) identified tacit knowledge as "sticky": costly to transfer because context is lost. PROV-DM embeds context (situational metadata, derivation history) directly into the knowledge structure.

Addressing the Trust Barrier

Reluctance to share often stems from lack of attribution. PROV-DM's wasAttributedTo ensures contributors receive explicit credit, transforming sharing into a documented contribution.

Ontology (Information Science)

A formal specification of a conceptualization. Unlike a taxonomy, ontologies capture complex relationships (e.g., "an officer can supervise multiple incidents").

Taxonomy vs. Ontology

Taxonomy: Tree structure (is-a). Dog is an animal.
Ontology: Graph structure (arbitrary relations). Dog is owned-by Person.

The Provenance Chain

Creation Trail (who/when) → Derivation Trail (reasoning logic) → Access Trail (retrieval). All trails are W3C PROV-O compliant.

Audit Trail Architecture

TacitFlow maintains three audit layers ensuring full accountability and traceability for every piece of knowledge in the system.

The Tacit Knowledge Gap

40% of public servants 50+ exit in 7 years. The solution applies the EASCI Framework to systematically capture this tacit knowledge.

The "Flywheel" Effect

Every round makes the system smarter. Verified KOs help with future cases, so knowledge grows faster over time.

Why "Compounding"?

Every solved case becomes a reference for the next one. The system doesn't just "process" data; it learns from it, but only through a strictly governed pipeline.

Why Gates Matter

Without strict gates, AI systems accumulate "zombie facts" - outdated or unverified information that degrades decision quality. Every promotion requires passing the gate.

Nonaka's SECI Model

Our EASCI framework is an evolution of Nonaka & Takeuchi's (1995) SECI model, adapted for AI-human hybrid teams. (Irbe, 2025b)

Smart Forgetting

Innovation requires not just learning, but unlearning. Kluge & Gronau (2018) showed that organizational knowledge must be actively pruned. "Smart Forgetting" invalidates outdated KOs to prevent "zombie facts."

Smart Forgetting & Creative Advance

Whitehead's (1929) "creative advance into novelty" reframes this: the settled past (validated KOs) becomes the platform for future innovation. The Innovation stage retires obsolete knowledge while generating new experiential context.

Complexity at the Edge

Stacey's Complex Responsive Processes (2000) and Kauffman's At Home in the Universe (1995) explain why the feedback loops produce non-linear, emergent outcomes. The compounding effect emerges from self-organizing interactions across EASCI stages.

Edge of Chaos

A metaphor from complexity science describing the boundary zone between order and randomness where complex systems exhibit maximum adaptability and creativity. Effective organizations operate at this productive boundary.

Transactive Memory System

Wegner (1995): Organizations remember collectively through a distributed cognitive network where members know "who knows what." TacitFlow operationalizes TMS by explicitly mapping expertise networks.

GDPR Compliance

GDPR Art. 17: The Innovation phase includes data retirement protocols. When KOs are superseded, the system maintains audit trails while pruning operational data per Art. 17 (Right to Erasure) requirements.

Micro vs Macro Loop

Micro-Loop

Tactical, real-time reasoning (OODA: Observe, Orient, Decide, Act).

Macro-Loop

Strategic knowledge governance (EASCI).

GraphRAG (Edge et al., 2024)

Connecting LLMs to knowledge graphs produces more accurate, traceable responses than document-based RAG alone. The graph structure enables reasoning over relationships, not just retrieving text chunks.

Graph-of-Thought

Structuring reasoning as a graph (not a chain) enables exploring multiple hypotheses simultaneously and selecting the best explanation through mutual exclusion (Besta et al., 2024).

Confidence Propagation

Each hypothesis in the GoT graph carries a confidence score derived from evidence quality. Pruning events update confidence in connected nodes.

Mutual Exclusion & Modus Tollens

GoT uses modus tollens ($((P \to Q) \land \neg Q) \to \neg P$) for hypothesis pruning: if a hypothesis implies a consequence that contradicts evidence, the hypothesis is eliminated. This prevents "hallucination cascades."

Complexity & Emergence

Kauffman (1995) showed complex systems self-organize at the "edge of chaos." The Micro Loop's cyclic process mirrors this: emergent intelligence arises from constrained iteration, not linear prediction.

Heuristics

Heuristics (availability, representativeness) enable fast decisions but introduce bias. TacitFlow's structured KOs provide reliable anchors to correct these biases (Kahneman, 2011).

Memory Architecture

LLMs have ephemeral context windows (128k–200k tokens). TacitFlow's Knowledge Graph persists indefinitely, transforming the assistant from a "stateless oracle" to a "learning partner."

Token Prediction vs. Reasoning

Standard LLMs predict the next token based on statistics. TacitFlow forces a different process: retrieve evidence, reason over it, then validate.

Multi-Hop Reasoning

Answers questions requiring multiple connected facts (e.g., "Who trained the officer who handled Case X?"). Vector search cannot traverse these relational chains; GraphRAG can.

Semantic vs. Keyword Search

Keyword: exact matches. Semantic: conceptual similarity ("routine patrol" $\approx$ "standard rounds"). TacitFlow combines both.

Dense Retrieval

Uses neural networks to encode queries and documents into dense vectors. Similarity is computed via cosine distance, capturing semantic relationships unlike sparse methods (BM25).

Theoretical Foundations

• Dewey: Knowledge is dynamic transformation ("situated inquiry").
• Dennett: "Intuition pumps" force externalization of tacit assumptions.
• Peirce: Abduction generates hypotheses from sparse evidence.
• Weick: Knowledge is "enacted" through social narrative.
• Whitehead: "Creative advance" - past KOs become data for new models.

Workplace Learning

Eraut (2004): Most learning is informal. The "Experience" stage captures authentic practice rather than formal documentation.

Social Knowledge

Tacit transfer depends on trust (Leonard & Sensiper). Mentor circles and verification address sharing barriers (Cabrera & Cabrera).

Integration of Theory

Integrates three theoretical components: (1) EASCI stages as macro-level knowledge lifecycle, (2) micro-loop inference cycle (Boyd, 1987), and (3) W3C PROV-O for provenance tracking.

Demo Instructions

Click "Run Simulation" to see a knowledge capture scenario. Use step controls (⏮ ⏭) for pedagogical walk-through. The visualization shows stage transitions, KO validation, and feedback loops.

Simulation Scenario

A security officer notices an anomaly during patrol. The system captures their observation, elicits reasoning through intuition pumps (indirect "what if?" questions), structures it into a KO, seeks peer validation, and updates the threat model.

Production Telemetry

TacitFlow tracks token consumption, retrieval latency (ms), confidence scores (%), and provenance node count. Three trace panels show GraphRAG data flow, GoT reasoning graphs, and PROV-O chains.

Retrieval Latency

Time from query to retrieval. Must be <200ms for voice. Optimized via ANN indexing and caching. High latency indicates graph growth or query complexity.

Confidence Scores

Based on embedding similarity, provenance completeness, peer validation, and recency. <70% triggers retrieval; <50% triggers expert review.

Real-Time Confidence

Confidence scores update live. Low confidence (<70%) triggers additional evidence retrieval. Very low confidence (<50%) flags the KO for manual expert review. This prevents premature consolidation of uncertain knowledge.

Real-Time Telemetry

The telemetry panel shows: PROV nodes (provenance chain length), KO count (Knowledge Objects referenced), Confidence % (weighted evidence score), Token count (LLM resource usage).

Production Metrics

Tracks Capture latency (time from observation to KO creation), Articulation completeness (AI probe iteration count), Validation rates (peer acceptance %), and Decay ratios (KOs archived vs. promoted).

Why Visualization Matters

Tufte (2001) emphasizes that complex processes become comprehensible when represented visually with appropriate detail. This simulation shows data at multiple levels simultaneously: individual reasoning steps (micro), stage transitions (macro), and system-wide metrics (telemetry).

Small Multiples

The three trace panels (GraphRAG, GoT, PROV-O) follow Tufte's "small multiples" principle (Tufte, 2001): identical visual structures showing different data streams allow direct comparison and pattern recognition across the inference stages.

Software as Hypothesis

Leinonen et al. (2008): The prototype is a research instrument for testing theoretical assumptions, not a commercial product. Features are hypotheses about what enables tacit knowledge transfer.

Cognitive Apprenticeship

Collins et al. (1989): Modeling, coaching, scaffolding, fading. TacitFlow's AI probes function as scaffolding that "fades" as analysts internalize reasoning patterns.

The Prototype

Unlike chatbots that predict tokens, TacitFlow structures claims with evidence. Every output follows W3C PROV-O provenance standards, making reasoning auditable.

KO Example (Simplified JSON-LD)

{
  "@context": "https://w3id.org/ko/v1",
  "@type": "KnowledgeObject",
  "id": "ko:uuid-...",
  "claim": "Suspect A linked to Van B",
  "evidence": ["ev:log-001", "ev:cam-02"],
  "confidence": 0.85,
  "author": "agent:lepik"
}

KOs encode claims with evidence following W3C PROV-O for legal chain-of-custody.

EU AI Act Compliance

TacitFlow addresses requirements for "High-Risk AI Systems" (Annex III, Law Enforcement) by implementing rigorous data governance, record-keeping (logging), and human oversight measures.

Human Agency

The system supports, not replaces, human judgment. The "Gate" mechanism ensures that AI proposes structure, but humans dispose of truth. The analyst remains the "human-in-command."

Bias Mitigation

Bias often hides in training data. By using RAG over a curated Knowledge Graph, the system bypasses the biases inherent in the LLM's pre-training, grounding answers only in verified institutional knowledge.

Data Sovereignty

Air-Gapped Deployment: To ensure absolute data sovereignty and prevent leakage, the architecture is designed for on-premise operation, disconnected from public cloud providers.