An analysis of the short talk from the same day on extracting tacit expertise into a graph while preserving ownership. The content is examined: the mechanics, the formats, the economics of contribution, and what was left unresolved.
This is the fourth talk of the same day, and it is short: a lightning format, five minutes and no demo. But by density it is more useful than many long ones, because it contains an engineering pipeline in full: ontology, interview agent, writing into the graph, circumstances of assertion, attribution, and an attempt to pay the author for their contribution. The speaker immediately sets aside the question of his own role: he is not a lawyer, he is an engineer, and he works through an engineering task inside the legal field that was discussed before him.
The talk comes in the block of short presentations after the main ones, and its subject is the same as the whole day, but from another side. The first three talks answered questions of who owes whom, where memory lives and what we dispose of. This talk poses a question that had remained to one side until then: where does the knowledge a system runs on come from at all, if the most valuable knowledge was never written down.
The framing of the topic sounds like "who owns what we know," and in the talk it is unfolded not as a philosophical question but as an extraction task: how to get knowledge out of a person's head without losing either its ownership or its circumstances, and so that its use can be settled up.
The speaker introduces himself briefly: twenty years of work with graphs and ontologies, co-founder of the Apache TinkerPop project, author of the open graph language Hydra. The line of work explains why the conversation is not about norms but about formats: a person who spent most of his career building tools for representing knowledge arrives at the question of sovereignty from the side where sovereignty has to be implemented in the data structure, not declared in a contract.
A second important detail: he says outright that he works on a new project and wants to bring it to the status of an Apache project. This places the whole talk in the context of an open tool. The claim that expertise extraction should be open and shared, not closed inside a single vendor, is not a declaration here but a consequence of the chosen licensing model.
A third element of the frame: he does not claim legal conclusions and speaks of an engineering task inside the legal field examined by the previous talk. This is a honesty rare for the subject: instead of trying to solve a legal question with code, he offers code that makes the legal requirements executable.
| As in the transcript | What it actually is | Status |
|---|---|---|
| Koguzi, the startup | Cognisee, the company building a cognitive infrastructure layer and describing itself as a laboratory of artificial collective intelligence | verified |
| My co-founded, Patrick Tinkerbop | Apache TinkerPop, the graph framework the speaker is a co-founder of. The "Patrick" fragment of the name is not confirmed in the transcript, and there is no such name in the project's list of co-founders | partial |
| a new open source project, called Hydra | Hydra, the open graph programming language, repository CategoricalData. The speaker calls it his main project and speaks of plans to bring it to an Apache project | verified |
| translingual programming | A way to write a program once and obtain it in several execution languages while preserving meaning. In Hydra this is realized through a correspondence between graphs and lambda calculus | verified |
| an Asian that you've talked to | An agent: a tool for interviews that conducts a conversation and writes the graph in parallel. The tool's name is garbled in the transcript and could not be confirmed | name unverified |
| the prominence strain of who said what | Provenance, source traceability: who exactly said what, when and under what circumstances | verified |
| your demanded discourse | Domain, the subject area: aviation operations, maintenance, healthcare | verified |
| modal logics / logic of proofs | Real formal systems. Modal logics describe the circumstances of assertion, the logic of proofs is applied to which connections are to be considered confirmed | verified |
| a quote that's making the rounds lately | The quote is not given in the transcript, so its source cannot be established. However, in Cognisee's documents this idea is fixed with a reference to Polanyi: we know more than we can tell | unverified |
The problem statement is built on three goals, and this is the most useful part of the talk for a conversation about data in general. The first goal: privacy. The second: ownership, that is a traceable chain of whose data was used. The third: sovereignty, and it is precisely this one, according to the speaker, that is less developed than the first two.
The difference is explained through content rather than terminology. Privacy answers the question of what is learned about you. Ownership answers the question of whose it is and where it came from. Sovereignty answers the question of how it is disposed of: whether a person is entitled to influence how their knowledge is used, and whether they receive anything when it brings benefit. The first two topics are already covered by norms and practices, the third remains poorly provided for even where the first two are respected.
Next the speaker defines what he intends to extract, and the definition is worth unpacking, because the whole architecture depends on it. Tacit knowledge is what is specific to a particular person and exists only in their head. It may be conscious or not, it may have been said at some point or may never have been written down, and a person may not know how to write it down at all.
An important consequence the speaker spells out: this is not the same as a graph of facts about the world. The knowledge being extracted consists of assertions, not of truths. Every assertion has a modal part: a person does not say "this is so" but "I believe this is so," "I was told this is so," "this was true yesterday," "this is true in Australia but not in the US." Without this part knowledge loses its meaning, because it becomes unclear how far it can be relied upon.
In practice this means that expertise extraction never reduces to transcribing text. The circumstances of the utterance must be recorded, and the speaker calls this a separate requirement rather than an implementation detail.
An ordinary knowledge base answers the question "what is known." An expertise base must answer the question "who exactly believes this and how confident are they." The first is good for reference, the second is good for a decision, because a decision in real work always rests on someone's judgment with caveats.
The core of the engineering account and the speaker's main practical technique: the work begins not with an interview but with a schema. Before the conversation with the expert, the ontology is defined, that is a set of concepts and types for a particular subject area. If the conversation is to be with someone from aviation operations, an aviation ontology is given; if with a maintenance technician, a maintenance ontology; if with a doctor, a medical one.
The ontology does two jobs at once. First: it sets which questions are appropriate, that is it guides the agent and makes its questions precise. Second: it serves as validation on the way in, because everything the agent said and everything that came back from the person is translated into the form of a triple, namely subject, predicate, object, and checked against the ontology. If the check fails, the record does not enter the graph in an incorrect form.
The speaker says outright what is wrong with the option without an ontology: in a narrow technical area the model asks less precise questions and cannot record the data in the required format. In other words, the ontology is not an ornament but what keeps both the questions and the records within the subject.
Next the extraction tool is named, and its design is worth describing precisely. It is an agent with which a person converses as with an ordinary interlocutor, but which in parallel builds a knowledge graph in the background and writes it into the graph database, checking each record as it appears. The speaker formulates the goal as an attempt to reproduce a person's expertise in the format of a graph on a shared vocabulary.
An essential detail that distinguishes this approach from common practices: validation happens not afterwards but during. This changes the role of the person. The expert does not fill out a form but answers questions, and the knowledge immediately lands in the structure in a verifiable form. The speaker did not show a demonstration because of the format and suggested looking at it at the demo table.
Now the technical core. In order not to lose the caveats mentioned above, modal representations are used. This is a way to record not only the content of an assertion but everything that surrounds it: when it was said, which space it belonged to, whether one country or another was meant, who exactly said it, whether the person was confident, whether they asserted it themselves or passed on what they had heard.
The second element is the logic of proofs, applied to which connections actually make it into the answer to a query. The difference between the two levels matters: the modal layer stores everything, including the doubtful and the disputed, while the logic of proofs selects from it what can be presented as grounds for an answer.
| Circumstance | Why it is recorded | What breaks without it |
|---|---|---|
| Who said it | Allows recognition and payment to be returned to the author | The ability to settle up for a contribution is lost |
| When it was said | Separates the outdated from the current | The system relies on information that is no longer true |
| Where it holds | Distinguishes regional conditions and norms | A rule from one country is applied in another |
| How confident the author is | Separates assertion, guess and hearsay | A retelling of someone else's opinion is presented as established fact |
| From whom it was heard | Builds the chain of transmission of information | The ability to assess the reliability of the source disappears |
The most substantive part of the talk and perhaps the most underrated. Next a scheme is built in which knowledge belongs to its author, and on use, recognition can be returned to the author and micropayments can even be organized: part of the revenue from answering queries goes to the one whose knowledge formed the basis of the answer.
And here a problem immediately emerges, which the speaker names honestly, without trying to disguise it. If two hundred people confirm the same fact, then the graph will have two hundred equal valid grounds. When a query comes, the system indexes it and picks some one, and as a rule picks the same one. As a result there forms what the talk calls a hotspot: one person gets paid every time, while the remaining one hundred and ninety-nine never do, even though their contribution is no smaller.
It follows that paying for knowledge is not an accounting task but a distribution task. Provenance tracking is ready, while fair selection of the source is not, and the speaker says outright that more work is needed here. I would note that the same problem exists even without money: any mechanism for ranking grounds creates invisible contributors through its selection.
The scheme with payment looks like a detail, but it has a strategic meaning, and it is not developed in the talk. If a person's knowledge brings benefit and the person receives nothing, then they have no reason to give it further. That is, payments are not charity but a condition for the flow of expertise to continue.
The reverse side is not named, but it is worth keeping in mind when applying this: direct payment for a contribution changes behavior. If people are paid for confirming a fact, an incentive appears to confirm what is profitable to confirm, and the quality of expertise can fall precisely where it is needed most. This risk is not examined in the talk, and in this analysis it is noted as an open question.
Next comes a short insert about the open project, and it matters as a link, not as promotion. Hydra is a programming language for graphs based on a correspondence between graphs and lambda calculus. From this correspondence grows a property the speaker calls translingual programming: a program is written once and then unfolds into other languages while preserving meaning.
The publicly stated state of the project is this: the language has reached mutual self-hosting, that is each of the implementations can generate all the cores, including its own. There are now implementations for Python, Java and Haskell, with Scala, Rust and Go announced next. The language supports ontologies directly, and it is precisely this that makes it relevant to the extraction task: the schema and the data are described in one tool rather than assembled from different ones.
The practical meaning for the theme of sovereignty: an open language with open ontologies allows different organizations to describe knowledge compatibly, without moving to someone else's closed platform. This is exactly the property the first talk of the day lacked: an obligation without a shared data structure remains a declaration.
Separately it is emphasized why ontologies must be open and shared. If knowledge is gathered from many people, the vocabulary must be one, otherwise incomparable descriptions cannot be put together. The speaker calls this a direct consequence of the task: openness here is not ideology but a technical requirement of aggregation.
The order of work follows from this on its own: first the ontology is defined, then it is used both for posing questions and for checking what has been recorded. Such an order reverses the usual practice, where data is collected first and the schema is derived afterwards. In the case of expertise the reverse order is impossible, because without a schema it is unclear what exactly to ask.
The speaker himself marks the boundary of his work, and this is an honesty rare for a short format. Tracking who said what, they have: the data is stored in a graph format, and this format allows storing attribution, following it, and settling payment for contributions. The remaining part, in his words, is query answering, and he promises to cover it separately, next time.
What this means in substance. Collecting knowledge together with its circumstances is possible and already works. Answering a question so that the right basis is chosen from among many equivalent ones does not work, because it is exactly at this step that the skew arises that was examined above. That is, what remains open is not a technical trifle, but precisely the area where sovereignty turns into money and recognition.
The talk is short, but a ready-made order of work can be drawn from it, and it applies to any task where knowledge must be extracted from a person and preserved.
A finished product cannot be taken from the presentation, because there was no demonstration and query answering is not closed. But the order of work can be taken from it in full: it answers the question of how to collect expertise so that it can later be paid for without losing the source.
All four talks of the day describe one system from different ends, and the joints are worth showing here, because taken separately each looks like a conversation about its own thing.
| Talk | The question it answers | What it adds to the overall picture |
|---|---|---|
| Agent obligations | Who owes what to whom and how to build this into the construction | A rule of agent behaviour and a mechanism to enforce it |
| Memory and verifiable identity | Where memory lives and how to verify a person | The context and credentials on which the agent operates |
| Control and infrastructure | What exactly we have authority over | Authority at four levels, from the agent to the megawatt |
| Expertise extraction | Where knowledge comes from and who gets paid for it | A format for knowledge with its circumstances and a mechanism for recognising contribution |
The joint between expertise extraction and memory is the tightest. Both talks are about the same thing: knowledge about a person and his work should belong to the person, not to the service. The difference is in direction: in one case memory is collected about the person himself, in the other he is himself the source of knowledge for others. The mechanics turn out similar: limited outputs, circumstances of the record, the right to decide what to pass on.
| Claim | What verification shows | Status |
|---|---|---|
| Knowledge is preserved together with ownership, and its use can be paid for | The scheme is described as working at the collection stage; the mechanism of payment and of fair choice of basis is not shown | claimed |
| Ontologies must be common and open | Consistent with the task of aggregating knowledge from many people | confirmed |
| The graph format allows storing attribution and following it | Technically correct: graph models support provenance and relations at the data level | confirmed |
| Query answering is solved separately and later | Stated directly by the speaker as the open part of the work | open question |
| The Hydra language has reached mutual self-hosting and is moving toward version 1.0 | A publicly claimed technical milestone, implementations for Python, Java and Haskell, with Scala, Rust and Go claimed next | confirmed |
| The name of the interview tool | Distorted in the transcript, could not be confirmed | unverified |
| The quote cited at the start of the talk | Not given in the transcript, the source cannot be established | unverified |
| A live demonstration of the work | Not shown because of the format, offered at the demo table | not shown |
First: the expertise graph becomes a new point of concentration. If the knowledge of many people is piled into one common vocabulary and one base, then the owner of that base receives exactly the position against which the first talks of the day were written. An open ontology reduces the risk but does not eliminate it: the vocabulary may be common while the storage is still one.
Second: attribution turns into a profile. An exact record of who said what and when is useful for settlements, but it is also exactly a detailed profile of a person: his judgements, his reservations, his knowledge. In the talk attribution is treated only as a good thing, and the flip side is not discussed.
Third: the quality of the ontology sets the ceiling of the whole system. If the domain schema is incomplete, the agent will not ask the needed questions, and knowledge that could have been extracted simply will not be extracted. An error in the schema at the input is not corrected later by the quality of the model, and this makes the ontology the most accountable node of the entire pipeline.
It honestly describes how to collect knowledge together with its circumstances and how to mark authorship, but it does not answer the question of what happens if the collector and keeper of knowledge turns out to be the same player against whom the whole defence is built. In the terms of the third talk of the day: control over collection is a separate authority, and in this scheme it is not verified.
| Term | Meaning in substance |
|---|---|
| Tacit knowledge | Knowledge specific to one person, residing in his head, often unwritten and hard to formulate |
| Ontology | A set of the concepts and types of a domain, which determines which questions are appropriate and what counts as a correct record |
| Assertion instead of fact | A record that stores not the truth about the world, but someone's judgement together with its circumstances |
| Modal logic | A formal apparatus for fixing circumstances: when, where, who, how confident, from whom it was heard |
| Logic of proofs | An apparatus that determines which relations count as confirmed and enter the answer to a query |
| Provenance | Traceability of the source of information: the chain of who said what and from whom it came |
| Triple | The basic unit of the graph: subject, predicate, object |
| Hotspot | A skew in distribution: out of many equivalent bases one is always chosen, the rest receive nothing |
| Micropayments for knowledge | Settling with an author for the fact that his piece of information became the basis of an answer |
| Translingual programming | Writing a program once, with subsequent unfolding into several execution languages while preserving meaning |
| Mutual self-hosting | A state of a project in which each implementation of the language is capable of generating all the cores, including its own |
The talk gives a working order for extracting expertise from a person into a graph together with the circumstances of its utterance, but it breaks off exactly where the question of money and fairness begins: collecting and signing knowledge is possible, but choosing from two hundred equivalent bases the one worth paying for is not yet possible.
This analysis was made from the transcript of a short presentation. Names and titles distorted by automatic transcription have been restored from primary sources and marked with a verification status. The name of the interview tool and the quote from the start of the talk have not been confirmed. The claims about the readiness of the payment mechanism and about the fair choice of source rest on the speaker's words: no demonstration of the work was carried out, and these parts he himself calls open.
ARLI membership: the notes from the workshop
One site of yours, measured, with the price for the work in the same document
ARLI, Inc., a Delaware corporation. Questions: hq@arliwork.com.