By Jofia Jose Prakash · published by O'Reilly Media · source article
Tacit knowledge is the hardest requirement in enterprise AI, and most knowledge programs never capture it. Jofia Jose Prakash lays out a tacit-aware architecture: an elicitation protocol that turns vague expert statements into decision rules (thresholds, exceptions, evidence, escalation), four planes (capture, representation, serving, transmission) that address distinct failure modes in how expertise moves, and an evaluation loop (incident replay, bus-factor audit, abstention calibration, transfer outcomes) that measures whether the knowledge that matters actually reached the people who need it.
Tacit knowledge is the hardest requirement in enterprise AI, and most knowledge programs never capture it. The article lays out a tacit-aware architecture: an elicitation protocol that turns vague expert statements into decision rules, four planes that address distinct failure modes in how expertise moves, and an evaluation loop that measures whether knowledge actually reached the people who need it. This collection meshes that architecture with agent-rdf-memory on Virtuoso and extends it with synthetic sample data for live querying.
Every knowledge program begins with the same request: a departing expert is asked to document her process, and she returns a clean flowchart of the happy path — leaving out the thresholds she watches, the conditions that make the procedure unsafe, and the supplier whose parts fail in humid weather. Six months later the line goes down and the knowledge base can't explain what to do, because no one asked the engineer to explain the judgment behind the procedure.
Michael Polanyi gave the problem its durable formulation in 1966:
“We can know more than we can tell.” — The Tacit Dimension
In companies, tacit knowledge usually appears in three forms, and each requires a different method of transfer. The synthetic sample data below models one of each, drawn from a fictional enterprise — Aurora Manufacturing — so the taxonomy can be queried live.
| Sample | What it captures | Held by |
|---|---|---|
| Nordic Bearing Co. parts fail above 70% RH | Maya knows the supplier whose parts fail in humid weather; the knowledge is elicitable because a precise question about conditions surfaces it. | Maya Reyes |
| Stop the line when plant humidity exceeds 70% RH | The threshold that turns Maya's general caution about humidity into an observable trigger. | Maya Reyes |
| Viscosity drift in supplier batch 3 is a leading failure indicator | Evidence Maya watches for: a viscosity drift in a specific supplier batch precedes bearing failure by about two weeks. | Maya Reyes |
| Escalation to the governance council for employment, credit, or health use cases | A decision rule mirroring the article's elaborated answer: escalation is required when a use case touches employment, credit, or health decisions, or when model output reaches a customer without human review. | Maya Reyes |
| Predeployment review skipped for internal-only tools with no personal data | The exception people get wrong most often: review is skipped only for internal-only tools with no personal data. | Maya Reyes |
| Review does not proceed without a named accountable owner | A blocking condition: if no named accountable owner can be identified, the review does not proceed regardless of risk tier. | Maya Reyes |
| Sample | What it captures | Held by |
|---|---|---|
| Engineer hears the bearing whine two weeks before the vibration sensor trips | Maya's trained ear detects the onset of bearing failure well before instrumentation; the knowledge lives in perception and is transferred by demonstration, not prose. | Maya Reyes |
| Nurse notices a patient looks wrong before the monitor changes | Priya's trained attention flags deterioration before objective monitoring; perceptual knowledge of the kind that resists documentation. | Priya Nair |
| Operator feels a motor-housing temperature change through a glove | A perceptual cue — a temperature delta too small to register as a formal reading — that an experienced operator reads by touch. | Maya Reyes |
| Technician detects a conveyor pitch change indicating belt wear | A change in the conveyor's sound pitch signals belt wear to an experienced technician before any measured degradation. | Joe Adler |
| Sample | What it captures | Held by |
|---|---|---|
| A sound decision is reversible within 48 hours without a line stop | Aurora's shared standard of what a sound decision looks like: it can be reversed within 48 hours without stopping the line. | Aurora Manufacturing (synthetic) |
| Shift handoff always restates the why, not just the what | A team habit: every handoff names the reason behind the instruction, preserving the causal context that a bare status list loses. | Aurora Manufacturing (synthetic) |
| Any change touching safety requires two signatures | A peer-review norm: no safety-relevant change ships without two independent signatures. | Aurora Manufacturing (synthetic) |
| Blameless post-incident norm — root cause, not person | A shared standard that makes post-incident review safe: the investigation names root cause, never the person. | Aurora Manufacturing (synthetic) |
The central design question: which follow-up would prompt an expert to say the missing judgment aloud? Expert explanations go vague in four places, and a targeted question about each turns a general statement into a usable rule.
We review high-risk use cases before deployment. If the risk seems significant, we escalate to the governance council.
Embeds cleanly and retrieves well, but “seems significant” supplies no decision criterion.
Escalation to the council is required when the use case touches employment, credit, or health decisions, or when model output reaches a customer without human review. Predeployment review is skipped for internal-only tools with no personal data, which is the exception people get wrong most often. If we cannot identify a named accountable owner, the review does not proceed, regardless of risk tier.
Contains a decision rule, an exception, a failure pattern, and a blocking condition.
Elicitation is one part of a larger knowledge system. A tacit-aware architecture has four planes — capture, representation, serving, transmission — and each addresses a different failure in the movement of expertise.
Collects more than polished procedure through structured interviews, incident reconstruction, decision journals, and observation — recording trigger, evidence, exception, and escalation while the expert can still explain the surrounding conditions, and routing perceptual skill toward demonstration rather than prose.
Reaches: ·
Preserves the distinctions that make material trustworthy — provenance, confidence, and validity context (plant, time period, equipment, conditions) as first-class properties, extending the knowledge graph beyond documents to the people and episodes that produced them.
Reaches: ·
Determines how knowledge reaches users: answers cite retrieved evidence and show the source; when the collection cannot answer, the system says so clearly and routes to someone with relevant experience — a referral restores the human contact through which difficult knowledge often moves.
Reaches:
Completes the architecture by helping expertise move between people through shadowing, teaching, and communities of practice, and by detecting when knowledge concentration and attrition risk converge so capture and apprenticeship begin before a notice period does.
Reaches: ·
Each plane is instantiated by a concrete component of agent-rdf-memory — the RDF agent-memory harness deployed on Virtuoso Universal Server — and this collection's graph is loaded into the same quad store for live querying.
agent-rdf-memory captures tacit operating knowledge as structured Turtle: sessions/YYYY-MM-DD-{llm-id}-{agent-env}.ttl record what was actually decided, and preferences.ttl holds the standing rules the expert would otherwise carry silently.
Each Turtle document is a first-class CreativeWork with schema:dateCreated/dateModified, PROV-O wasGeneratedBy, and named-graph deployment — provenance, confidence, and validity context as first-class properties rather than interchangeable chunks.
The ontology.ttl PromptIntent classes route questions to the right preference topics and howto documents, and the endpoint answers cite their graph; when the store cannot answer, the retrieval policy falls back rather than fabricating.
The shared store is cross-LLM and cross-environment; {llm-id}-{agent-env} in filenames is provenance, not isolation — expertise moves between people and models through the same graph and its cross-session index (index.ttl), the way shadowing and communities of practice move difficult knowledge.
Retrieval precision and answer faithfulness measure how well a system serves its existing collection — not whether the collection contains the knowledge the organization actually depends on. That needs a separate evaluation loop tied to capture priorities and transfer outcomes.
Select 20 or 30 resolved incidents, remove the resolutions, give the opening facts to the system, and have the engineers who solved them grade its responses against a frontier model lacking the company's collection; the gap reveals the generic-answer rate.
Test questions that only one or two employees can answer and study how the system fails; a clear admission of uncertainty followed by a useful referral is healthy, while fluent boilerplate damages trust in every response.
Build a labeled set of answerable and unanswerable questions, then track abstention precision and recall as the collection grows; a system that never says 'I don't know' is unevaluated on the dimension that matters most.
Measure whether knowledge reached the people who need it: shorter time to proficiency, fewer repeat incidents after elicitation, and fewer critical responsibilities that depend on a single person. Document and query counts describe activity, not transfer.
Which number, reading, or condition triggers the action? Turn 'if the risk seems significant' into an observable criterion.
When does the documented procedure cease to apply? Surface the conditions that make the standard procedure unsafe or inapplicable.
What did the expert observe before reaching the conclusion? Record the signal that precedes the judgment.
Who becomes involved, and at what point? Make the referral path explicit.
Limit the number of follow-ups so a long interrogation does not produce agreeable noise; keep the model focused on generating questions; and separate question-generation from compiling and validating answers — an expert's statement belongs in the record with provenance and context, but its accuracy requires independent review.
Tacit knowledge is what experts hold but cannot fully tell — skill, perception, and judgment that resist complete explanation. It matters because organizations keep improving retrieval over collections that omit their most valuable operating knowledge; better ranking helps people find what was recorded but cannot recover the expertise that never entered the collection.
Elicitable knowledge (unspoken because nobody asked a precise enough question), perceptual knowledge (trained attention, such as hearing a bearing fail), and collective knowledge (a team's habits, standards, and shared sense of a sound decision). Each form requires a different method of transfer.
Knowledge that remains unspoken only because nobody has asked a precise enough question, or because an expert assumes everyone sees what she sees. A focused follow-up question is usually enough to surface it.
Knowledge that lives in trained attention — an engineer hears a bearing begin to fail, or a nurse notices that a patient looks wrong before a monitor changes. It is best transferred by demonstration and practice, not forced into prose.
Knowledge that resides in a team's habits, standards, and shared sense of what a sound decision looks like in that organization. It moves through communities of practice, teaching, and shared norms.
The observation, named by economist David Autor after Michael Polanyi's 'we can know more than we can tell', that many of the tasks hardest to automate depend on rules we cannot state. Modern machine learning works around it by learning from examples, but that workaround weakens when examples of rare expertise are scarce.
A failure produces a ticket, an incident report, and a trail of messages, but an experienced operator who quietly avoids a failure produces none of those records. The useful outcome is the absence of an event, so the data pipeline receives no trace of the decision that produced it.
Capture, representation, serving, and transmission. Capture collects more than polished procedure; representation preserves provenance, confidence, and validity context; serving cites evidence, abstains honestly, and routes to experts; transmission moves expertise between people through shadowing, teaching, and communities of practice.
Questions about thresholds (which number or condition triggers action), exceptions (when the documented procedure ceases to apply), evidence (what the expert observed), and escalation (who becomes involved and at what point).
A vague first answer such as 'we escalate if the risk seems significant' supplies no decision criterion and cannot guide a new employee. One focused follow-up yields an elaborated answer containing a decision rule, an exception, a recurring failure pattern, and a blocking condition — enough to guide a real dispute.
Limit the number of follow-ups so a long interrogation does not exhaust the expert into agreeable noise; keep the model focused on generating questions; and separate question-generation from compiling and validating answers — a statement's accuracy requires independent review.
Select 20 or 30 resolved incidents, remove the resolutions, give the opening facts to the system, and have the engineers who solved them grade its responses against a frontier model without the company's collection. The gap reveals how often the internal system merely restates public knowledge.
A test of the questions only one or two employees can answer, studying how the system fails on them. A clear admission of uncertainty followed by a useful referral is healthy; fluent boilerplate damages trust in every response, including the accurate ones.
A measure of whether the system answers when evidence exists and declines when the corpus cannot support an answer, using a labeled set of answerable and unanswerable questions. A system that never says 'I don't know' is unevaluated on the dimension that matters most.
Evidence that knowledge reached the people who need it: shorter time to proficiency, fewer repeat incidents after elicitation, and fewer critical responsibilities that depend on a single person. Document and query counts describe activity, not whether someone else can now make the decision.
Knowledge an expert holds but cannot fully articulate — skill, perception, and judgment that resist complete explanation, even when the expert sincerely tries to teach them.
Tacit knowledge that remains unspoken only because nobody has asked a precise enough question, or because the expert assumes everyone sees what she sees.
Tacit knowledge living in trained attention — an engineer hears a bearing begin to fail, or a nurse notices a patient looks wrong before a monitor changes.
Tacit knowledge residing in a team's habits, standards, and shared sense of what a sound decision looks like in that organization.
Knowledge that can be written down, documented, and circulated — the part of the collection that retrieval, ranking, and fine-tuning address.
Many of the tasks that are hardest to automate depend on rules we cannot state; modern machine learning works around it by learning from examples, but the workaround weakens when examples are scarce.
The difficulty that arises when people can describe the steps of a method without fully understanding why they work — a primary reason best-practice transfers fail.
An experienced operator's quiet avoidance of a known failure mode; it produces no ticket, no incident report, and no data trace — the useful outcome is the absence of an event.
A group whose members develop and share a practice; in the transmission plane, communities of practice carry difficult knowledge that does not move through documents.
How often an internal knowledge system merely restates public knowledge that a frontier model already has; measured by comparing incident-replay answers against a frontier model lacking the company's collection.
The number of people whose sudden absence would cripple an activity; a bus-factor audit probes the questions only one or two employees can answer and studies how the system fails on them.
Whether the system answers when evidence exists and declines when the corpus cannot support an answer; a system that never says 'I don't know' is unevaluated on the dimension that matters most.
A capture-plane instrument that records the trigger, evidence, exception, and escalation path of a decision while the expert can still explain the surrounding conditions.
A capture-plane instrument that reconstructs a resolved incident — its opening facts, decisions, and outcomes — so the reasoning behind it can be examined later.
One of four places expert explanations go vague: which number, reading, or condition triggers the action.
One of four places expert explanations go vague: when the documented procedure ceases to apply.
One of four places expert explanations go vague: what the expert observed before reaching the conclusion.
One of four places expert explanations go vague: who becomes involved, and at what point.
Interactive graph visualization derived from the companion RDF. Click nodes to resolve, drag to explore. Graph data embedded from companion RDF at generation time.
Query this knowledge graph on URIBurner. The editor opens on the canonical SAMPLE entity-type summary (DAV named graph). Pick a recipe, edit freely, then run live or copy.
Reproduced verbatim from the companion RDF. Execute loads the query into the workbench below and runs it live.
PREFIX tk: <https://oreillyradar.substack.com/p/architecting-for-the-knowledge-you/ontology#>
PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#>
PREFIX schema: <http://schema.org/>
SELECT ?kind (COUNT(?sample) AS ?count)
FROM <https://linkeddata.uriburner.com/DAV/demos/daas/oreilly-tacit-knowledge-deepseek_v4pro-1.ttl>
WHERE {
?sample a tk:TacitKnowledge ; a ?type .
?type rdfs:label ?kind .
FILTER(?type IN (tk:ElicitableKnowledge, tk:PerceptualKnowledge, tk:CollectiveKnowledge))
} GROUP BY ?kind ORDER BY DESC(?count)Count the synthetic sample data by the three kinds of tacit knowledge.
PREFIX tk: <https://oreillyradar.substack.com/p/architecting-for-the-knowledge-you/ontology#>
PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#>
PREFIX schema: <http://schema.org/>
SELECT ?sampleIri ?name ?kind ?expert
FROM <https://linkeddata.uriburner.com/DAV/demos/daas/oreilly-tacit-knowledge-deepseek_v4pro-1.ttl>
WHERE {
?sampleIri a tk:TacitKnowledge ; a ?type ; schema:name ?name .
?type rdfs:label ?kind .
OPTIONAL { ?sampleIri schema:author ?author . ?author schema:name ?expert }
FILTER(?type IN (tk:ElicitableKnowledge, tk:PerceptualKnowledge, tk:CollectiveKnowledge))
} ORDER BY ?kind ?nameAll fourteen synthetic tacit-knowledge samples, the kind each belongs to, and who holds it.
PREFIX tk: <https://oreillyradar.substack.com/p/architecting-for-the-knowledge-you/ontology#>
PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#>
PREFIX schema: <http://schema.org/>
SELECT ?sampleIri ?name ?session
FROM <https://linkeddata.uriburner.com/DAV/demos/daas/oreilly-tacit-knowledge-deepseek_v4pro-1.ttl>
WHERE {
?sampleIri a tk:ElicitableKnowledge ; schema:name ?name ; tk:elicitedIn ?sessionIri .
?sessionIri schema:name ?session .
}Trace which offboarding session surfaced each elicitable rule.
PREFIX tk: <https://oreillyradar.substack.com/p/architecting-for-the-knowledge-you/ontology#>
PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#>
PREFIX schema: <http://schema.org/>
SELECT ?sampleIri ?name ?expert
FROM <https://linkeddata.uriburner.com/DAV/demos/daas/oreilly-tacit-knowledge-deepseek_v4pro-1.ttl>
WHERE {
?sampleIri a tk:PerceptualKnowledge ; schema:name ?name ; schema:author ?author .
?author schema:name ?expert .
}Perceptual knowledge lives in trained attention — who carries each cue.
PREFIX tk: <https://oreillyradar.substack.com/p/architecting-for-the-knowledge-you/ontology#>
PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#>
PREFIX schema: <http://schema.org/>
SELECT ?sampleIri ?name ?holder
FROM <https://linkeddata.uriburner.com/DAV/demos/daas/oreilly-tacit-knowledge-deepseek_v4pro-1.ttl>
WHERE {
?sampleIri a tk:CollectiveKnowledge ; schema:name ?name ; schema:author ?org .
?org schema:name ?holder .
}Collective knowledge belongs to the team, not any one person.
PREFIX tk: <https://oreillyradar.substack.com/p/architecting-for-the-knowledge-you/ontology#>
PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#>
PREFIX schema: <http://schema.org/>
SELECT ?plane ?kind
FROM <https://linkeddata.uriburner.com/DAV/demos/daas/oreilly-tacit-knowledge-deepseek_v4pro-1.ttl>
WHERE {
?planeIri a tk:KnowledgePlane ; schema:name ?plane ; tk:reaches ?kindIri .
?kindIri rdfs:label ?kind .
} ORDER BY ?planeThe tacit-aware architecture's planes and which tacit kinds each can reach.
PREFIX tk: <https://oreillyradar.substack.com/p/architecting-for-the-knowledge-you/ontology#>
PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#>
PREFIX schema: <http://schema.org/>
SELECT ?stepIri ?name ?position
FROM <https://linkeddata.uriburner.com/DAV/demos/daas/oreilly-tacit-knowledge-deepseek_v4pro-1.ttl>
WHERE {
?howto a schema:HowTo ; schema:step ?stepIri .
?stepIri schema:name ?name ; schema:position ?position .
} ORDER BY ?positionThe four follow-up questions plus the guardrails, as an ordered schema:HowTo.
PREFIX tk: <https://oreillyradar.substack.com/p/architecting-for-the-knowledge-you/ontology#>
PREFIX rdfs: <http://www.w3.org/2000/01/rdf-schema#>
PREFIX schema: <http://schema.org/>
SELECT ?plane ?component
FROM <https://linkeddata.uriburner.com/DAV/demos/daas/oreilly-tacit-knowledge-deepseek_v4pro-1.ttl>
WHERE {
?planeIri a tk:KnowledgePlane ; schema:name ?plane ; tk:mapsToMemoryComponent ?component .
} ORDER BY ?planeThe mesh: each architecture plane mapped to the agent-rdf-memory graph it instantiates.