When Observation Loses Authority

WORKING PAPER

When Observation Loses Authority
Three operational problems, one structural discipline, and why organisations that
produce more may know less




Author                         Alessio Virgili — Virgili Studio

Status                         Working Paper — First Release

Version                        1.0 — July 2026

Audience                       Advisory firms · Executive search · Private equity · Venture builders ·
                               Knowledge-intensive organisations

Contact                        virgilistudio.com




Abstract



   Generative AI has dramatically increased the speed at which organisations produce text, analysis,
decisions, and strategic documents. This expansion does not automatically translate into greater
knowledge. When no structure distinguishes what has been observed from what has been interpreted,
what is demonstrated from what is merely plausible, and what has been decided from what has only been
generated, production accumulates faster than understanding does. Observations become hypotheses
without being labelled as such. Hypotheses are repeated until they acquire the appearance of facts. The
most recent formulation silently replaces the source it was derived from. Each new project reconstructs
the existing patrimony rather than building on it.

   This paper documents how three distinct operational problems — the representation of complex
professional authority, the identification of real opportunities from well-formed ideas, and the
governance of knowledge through a growing body of generative output — revealed a shared structural
weakness. Each required not a better tool, but a discipline capable of keeping observation, knowledge,

decision, and generation rigorously separated. On this basis, Virgili Studio developed a practical
infrastructure: the CV Engine, the Opportunity Engine, and the Virgili Core Engine — a canonical
knowledge kernel that governs what can be treated as authoritative, in which conditions, and why.

   The contribution is not the invention of provenance, knowledge graphs, or executive assessment. It is
their operational configuration around a principle that existing frameworks address only partially: an
organisation does not accumulate knowledge simply by storing information or producing documents. It
accumulates knowledge when what it learns can survive the project that generated it, retain its epistemic
status, and be interrogated again without losing its connection to the reality from which it came.

   The paper concludes that the primary limit of generative systems is not their capacity to formulate
responses. It is the absence, in most organisations, of an infrastructure capable of establishing what
possesses authority, how that authority was built, and under what conditions it can be transferred to a
new decision.




1 The Problem That Precedes the Solution


   The organisations that work with knowledge — advisory firms, executive search practices, private
equity teams, venture builders, family offices — share a structural tension that generative AI has made
more visible without resolving it.

   The tension is this: producing more does not mean knowing more.

   When producing a document cost substantial time, each document received proportional attention.
When production approaches zero friction, the cost moves from creation to verification, from writing to
classification, from formulation to the governance of what status a formulation is allowed to acquire.
Most organisations have built tools for the first problem. Very few have built the discipline required for
the second.

   The consequences are observable and recurring. A preliminary observation is repeated across several
documents and progressively acquires the appearance of a verified conclusion. An estimate loses its
label as estimate. An example becomes a pattern. A plausible relationship between two elements is read,
in later versions, as a demonstrated one. The most recent formulation of a claim becomes more
authoritative than the source it summarises — not because anyone decided this, but because the system
rewards linguistic fluency over provenance.

   This paper does not treat this tension as a technological problem. The generative systems currently
available did not create it. They amplified it, because they accelerated the accumulation of output
without altering the underlying absence of structure. The discipline required precedes the tools. It
concerns the process by which an observed reality becomes a representation, a representation becomes a
decision, and a decision becomes an operational output.

The question is not how to produce more. It is how to ensure that what is produced does
     not erode the authority of what was known before it existed.

   Three problems made this tension concrete for Virgili Studio. They appeared, initially, to be
unrelated. What they revealed — progressively, through their failure modes — was that they shared the
same structural cause.




2 Three Problems, One Structure


2.1 The CV Engine: representing authority that exceeds chronology

   The first problem emerged from the work of representing a senior executive whose professional value
did not reside in a single function or a linear career path. The value derived from the combination of
capabilities exercised across different contexts: product architecture, manufacturing governance,
commercial scaling, brand construction, category management, operational transformation — applied
across more than two decades in roles that were simultaneously entrepreneurial, industrial, technical,
and commercial.

   Traditional CV systems organise professional authority chronologically: company, title, period,
responsibilities, results. This structure is effective when the value of a person corresponds largely to the
continuity of their functional path. It becomes insufficient when the real contribution consists of having
built systems, solved problems across multiple domains, or exercised the same capability in formally
different contexts.

   The market for senior mandates presents a structural characteristic that compounds this insufficiency.
At the highest levels of seniority — Chief Product Officer, Operating Partner, General Manager, Board
Advisor — the encounter between supply and demand rarely occurs through publicly observable
channels. A significant proportion of C-level, board, interim, and advisory opportunities is intermediated
by a limited number of international executive search firms, funds, investors, family offices, and
professional networks that identify and assess candidates before any formal mandate is published. The
window through which the market can read a profile is narrow, and it opens only when the profile has
been made legible in advance.

   The system built to address this problem identified four distinct failure modes. A Translation Failure
occurs when a capability exists but the market reads it incorrectly — the experience is present, but the
framing makes it unrecognisable to the buyer. A Depth Failure occurs when the capability exists but not
at the depth the mandate requires. A Domain Failure occurs when the required domain is genuinely
absent. A Visibility Failure — the most frequent — occurs when the capability exists, has been exercised
at the correct depth, and sits in the correct domain, but the market does not see it, because the profile has
not been made visible to the right interlocutors through the right channels.

The problem was not to write a better document. It was to build a system from which
     better documents of different kinds could be derived without reconstructing the
     underlying person each time.

   The response was not a better CV. It was a Capability Pattern — a structured map separating entities
(companies, brands, platforms) from the capabilities that manifested within them. The same verified
corpus could then produce a Chief Product Officer CV, an Operating Partner profile, an executive
snapshot for private equity and family office contacts, a B2B presentation for search firms, or a response
to a specific mandate. The content does not change each time; it is selected, organised, and translated
relative to the buyer and the problem, while maintaining its connection to verified evidence.

   The architecture distinguished at least three capability clusters that the market tends to acquire
separately: brand building, product architecture, and platform building. It also distinguished, within each
cluster, between what had actually occurred, the available evidence, the capability that could be
reasonably inferred, and the way that capability was relevant in a new context. Without this separation,
representation oscillates between two failures: remaining so close to the chronology that transferability
becomes invisible, or becoming so interpretive that the connection to evidence is lost.




2.2 The Opportunity Engine: distinguishing a real opportunity from a well-
formulated idea

   The second problem emerged in a different domain: the identification and evaluation of
entrepreneurial and investment opportunities.

   Generative AI has made it relatively straightforward to produce concepts, market scenarios, category
extensions, and possible business models. This abundance does not solve the more important problem —
how to distinguish an interesting idea from an opportunity with sufficient commercial, industrial, and
strategic potential to justify investment, time, and organisational construction.

   Existing strategic and financial research addresses parts of this problem. Consulting firms use AI to
extend research, accelerate analysis, identify patterns, and compare scenarios. In investment and M&A
processes, analytical systems filter targets, evaluate markets, and reduce the field of alternatives. These
systems operate, however, on a universe that already exists: companies, assets, startups, technologies, or
investment programmes that can be subjected to scoring.

   The problem Virgili Studio encountered was located before this phase. The object to be evaluated
was not yet a company or a business case. It could be a daily friction, a change in human behaviour, an
industrial capability not fully used, a material with applications not yet built, a brand asset not yet
translated into a platform, or an intermediate space between existing categories. A normal scoring model
was insufficient, because scoring requires the object to be already defined. The problem was to construct
the object without confusing the quality of the narrative with the quality of the opportunity.

The response was a methodological chain in which the idea is not the starting point. The starting point
is the observation. An observation records a friction, a behaviour, a change of context, or an unused
capacity — without immediately transforming it into a thesis. A requirement clarifies what condition
should be satisfied and for whom. An hypothesis constructs a possible response while keeping visible its
provisional character. Evidence establishes what genuinely supports the hypothesis and what remains to
be verified. Evaluation compares attractiveness, accessibility, capability, risk, and development
conditions. A decision does not produce only a yes or a no; it preserves the reasons why an opportunity is
developed, suspended, or excluded.

     Even an opportunity that is not pursued can generate transferable knowledge. A search
     can produce a protocol, an industrial map, an access path, an evidence, a capability, or
     a new observation domain that survives the outcome of the original project.

   This principle — that the unit of learning is not the successful project but every verified element that
remains available for a future decision — changed how results were evaluated. The question is not only:
what did the project produce? It becomes: what did it leave to the organisation that will remain usable
after its conclusion?




2.3 The Knowledge Kernel: preventing generation from rewriting its own
sources

   The third problem was not initially visible as an autonomous challenge. It emerged as the first two
systems grew.

   With the accumulation of analyses, documents, stress tests, and versions, each new output tended to
reinterpret the previous ones. A definition could be rendered more incisive and, precisely for that reason,
lose some of the qualifications that had made it accurate. A hypothesis could be taken up in a subsequent
document and progressively become a fact. A number introduced to summarise a corpus could be
repeated without it being any longer clear whether it derived from a verified source, from a count, or
from a formulation created to reinforce the narrative.

   The problem was not the obvious hallucination. It was the gradual modification of the epistemic
status of information. The difference is precise. Stating that a mechanism was observed in an
anonymised case is different from stating that it recurred across four independent mandates. The second
sentence may be more authoritative and persuasive, but it introduces a quantitative claim that must be
documentable. When a system rewards narrative fluency over provenance, this kind of drift can occur
without anyone explicitly deciding to invent a datum.

   A radical editorial rule was introduced: no information could enter the corpus unless it came from a
verified source; plausible fillers, undeclared inferences, reasonable assumptions, and automatic
completions were excluded. This rule was correct, but the growth of the system demonstrated that an
editorial rule was insufficient. What was required was an architecture capable of enforcing the

distinction between source, canonical object, interpretation, and output — not through human vigilance
alone, but through structural design.

   The research context around this problem is well-developed, though its focus differs from what was
required here. The NIST AI Risk Management Framework insists on the need to document provenance,
govern risk, maintain human oversight, and evaluate how content and feedback propagate through
generative systems. Microsoft Research's VeriTrail addresses a fundamental challenge in multi-step
generative workflows: verifying only the final output is insufficient; it is necessary to identify where
unsupported content was introduced and reconstruct the path connecting reliable elements to original
sources. PROV-AGENT extends the W3C PROV model to include prompts, responses, and agent
decisions across complex workflows, recognising that the output of one agent can become the input of
the next and propagate errors throughout. The most recent literature on reasoning provenance goes
further, arguing that execution traces and state checkpoints are insufficient to explain why a system
chose an action, what it inferred from an observation, and what evidence supports its conclusion.

   These contributions confirm that provenance, traceability, audit, and reasoning representation are not
marginal problems but central requirements of reliable AI systems. They answer predominantly the
question: where does this output come from, and through what steps was it produced? The challenge
Virgili Studio encountered required an extension of this question. Not only: from where? But also: what
status may this output acquire within the organisation? The two questions are connected but do not
coincide.

     Reconstructing the provenance of a claim does not automatically establish whether it
     should be accepted as evidence, treated as interpretation, maintained as hypothesis, or
     excluded from the authoritative corpus.

   A knowledge graph can connect entities without establishing whether the relationships are permitted.
An archive can preserve versions without establishing which is currently active. A retrieval system can
recover a source without understanding whether it has been superseded. An agent can document its
reasoning without possessing the authority to transform that reasoning into organisational knowledge.
The problem of provenance needed to become a problem of epistemic authority.




3 Research Method


   The discipline described in this paper did not emerge from an intention to build a general theory of
organisational knowledge. It emerged from working on problems that, in their origin, appeared
unrelated.

   The research developed through an iterative process. Recurring patterns observed in the first projects
generated interpretive hypotheses. These hypotheses were progressively tested across subsequent

projects belonging to different domains — executive positioning, opportunity identification, investment
evaluation, advisory mandates, venture construction, brand development. Hypotheses that did not show
continuity were abandoned. Recurrences that continued to manifest independently of context were
progressively formalised until they constituted the principles of the observational discipline.

   The corpus analysed includes the complete professional patrimony of Alessio Virgili — 23 years of
operational experience across Lunaria Cashmere (founded 2003, international luxury knitwear, 100+
wholesale accounts across 15+ markets, retail network, full P&L, exit 2024) and Oratio (investor-backed
ultra-luxury menswear platform, built from inception, six runway seasons, complete product and
manufacturing architecture across eight Italian manufacturing districts). It includes the mandate
development work of Virgili Studio across advisory, product development, capital, and venture
domains. It includes the applied research conducted to build the three engines documented in this paper.

   The method is qualitative and applied. It does not claim to produce a general theory of decision
processes. It documents principles that showed structural continuity across a sufficient number of cases
and domains to warrant formalisation. The claim is not that these principles are universally correct. The
claim is that they are structurally sound for the class of problems they address, and that their application
produced concrete operational improvement — reduction of repetitive reconstruction, preservation of
evidence across successive outputs, legibility of professional authority across different buyer types and
mandate contexts.

   What remains to be validated through independent cases is the comparative claim: whether this
discipline produces better decisions than alternative approaches, and by what measure. This paper
presents the implemented solution and the reasoning behind it. It does not yet present the controlled
comparative study.




4 What the Problems Revealed


   Three structural findings emerged from the work across the three problems.


4.1 Entities are not capabilities

   The first finding was a distinction that had previously remained implicit: the companies, brands, and
platforms built over time are not capabilities. They are the contexts in which capabilities manifested.
Confusing entity with capability is the source of most translation failures in senior executive
representation.

   When a profile says 'Founded and led Lunaria Cashmere for 21 years,' it names an entity. The
capability that this entity evidences — product architecture, manufacturing governance, international
commercial scaling, category building, P&L ownership, succession and exit management — is a

different level of description, and it is the level the market actually purchases. The entity provides the
proof. The capability is what the buyer acquires.

   This distinction allowed the system to identify at least three capability clusters that the senior market
tends to buy separately: brand building, product architecture, and platform building. A profile strong in
all three is exceptional. A mandate that requires all three simultaneously is rare. Most mandates require
one primary cluster with secondary contributions from another. The system could now match not
profiles to job titles, but capability clusters to mandate structures.




4.2 Preserved information and available knowledge are not the same thing

   The second finding is a distinction between preservation and continuity.

   Preservation concerns the survival of information. A document can be archived, indexed, and made
accessible across decades.

   Continuity concerns the persistence of the capacity for response that the knowledge contains. It
requires that what was learned can still be understood, judged, and applied when the context changes.

   These two conditions do not coincide. A procedure can be perfectly documented and prove
insufficient when reality presents an unforeseen configuration. A method can be described in every
detail and not enable a new operator to recognise when to apply it, when to suspend it, or when to modify
it. A decision can be preserved without preserving the judgment that had connected the circumstances to
the choice made.

   The difference appears clearly in high-competence domains. A medical protocol does not eliminate
the need for clinical judgment. A musical score does not contain the interpretation. A technical
specification does not exhaust the craftsman's capacity to respond to variations in the material. A
strategic model does not automatically determine which problem should be addressed first.

   Generative AI makes this tension more visible. It can enormously expand the capacity to use what has
been codified, but it does not eliminate the distance between available content and the specific reality to
which it must be applied. It can produce a plausible response even when that distance has not been
recognised. Its linguistic fluency can therefore conceal exactly the point at which judgment becomes
necessary again.




4.3 Production is not the same as accumulation

   The third finding concerns the conditions under which knowledge becomes cumulative rather than
episodic.

Most strategic, creative, and entrepreneurial work produces episodic knowledge. A project is
analysed, a decision is taken, a document is delivered, and the team moves to the next problem. Part of
the learning remains in the memory of the people involved, part in files, part in the informal language
with which the project is remembered. When a similar problem arises, the organisation rarely restarts
from the actual previous patrimony. It restarts from a reconstruction of it.

   This reconstruction is costly and imperfect. It depends on who is present, what they remember, which
version of the documents can be recovered, and how they retrospectively interpret the experience.

   Knowledge becomes cumulative when it satisfies four conditions: it must survive the occasion that
generated it; it must be reusable without being separated from the conditions that limit its validity; it
must preserve its provenance and its epistemic status; and it must increase the capacity to handle future
cases, not only document past ones.

   This principle changes how project results are evaluated. An opportunity not pursued can have
produced a verified method, a qualified network, an evidence, a capability, or sector intelligence. A
completed mandate can have produced a transferable proof, a decision structure, or a more precise
understanding of the limits of a model. An error can have produced an exclusion rule that prevents
similar errors in the future.




5 The Infrastructure Built


5.1 The CV Engine

   The CV Engine is the operational response to the first problem. It is not a system that automates CV
writing. It is a system that builds a verifiable representation of professional authority and transforms it
into different outputs without having to reconstruct the person from scratch each time.

   The engine is organised around a layered architecture. The Capability Pattern separates entities from
capabilities and maps the recurring combinations across the full professional corpus. The Value Creation
and Evidence Library catalogues, for each capability, the documented evidence — outcomes with
sources, metrics with context, transformations with conditions. The Problem-to-Proof Matrix connects
the problems a buyer might be trying to solve with the proofs the corpus can provide. The Asset
Narrative Library translates verified content into buyer-specific language for different mandate types.

   The output layer produces documents calibrated for different interlocutors: an executive CV for a
specific mandate, a CPO profile for search firms, an Operating Partner snapshot for private equity, a B2B
document for advisory contexts, a response to a confidential mandate. The content is not rewritten each
time. The selection, organisation, and translation change. The connection to verified evidence does not.

The engine also tracks outcomes. The Outcome Tracking Framework connects candidacies,
responses, shortlists, interviews, mandates, and compensations to the way the profile was represented.
Over time, this produces the first real dataset: not a framework about how the market works, but
evidence of how it actually responded to specific configurations of the profile. Translation Failure,
Depth Failure, Domain Failure, Visibility Failure, Timing Failure — these categories become
empirically populated rather than theoretically defined.




5.2 The Opportunity Engine

   The Opportunity Engine is the operational response to the second problem. It intervenes before the
opportunity is already a definable object, at the point where most evaluation processes do not yet have a
method.

   The engine is structured around an Observation Library — not a list of queries or trends, but a
collection of observation protocols by domain. The function of observation is not to immediately
produce a concept but to record a friction, a behaviour, a change of context, or an unused capacity in a
sufficiently disciplined way that it can be compared and subsequently developed.

   From observation, the system moves to the Requirement Engine: an observation acquires relevance
when it can be translated into a need or requirement, when a subject exists for whom the friction is real,
when it is possible to understand in which context it manifests, and when conditions emerge that allow an
episode to be distinguished from a pattern.

   Only then does the system construct an opportunity hypothesis, connect it to evidence, compare it
with available capabilities, potential buyers, channels, acquisition models, foreseeable objections, and
development conditions. The resulting sequence — observation, requirement, hypothesis, evidence,
evaluation, decision — differs from both an idea generator and a normal investment screening system. It
does not presuppose that the opportunity exists in complete form. It makes explicit the path through
which an observed friction becomes an object sufficiently defined to be subjected to judgment.

   The engine also preserves rejected opportunities: the reasons for exclusion, the missing evidence, the
conditions that could make them relevant again, and the capabilities required to develop them. A
negative decision contributes to the system's knowledge rather than disappearing with the end of a
conversation.




5.3 The Virgili Core Engine — a canonical knowledge kernel

   The Virgili Core Engine (VCE) is the architectural response to the third problem and the common
foundation beneath the first two.

The decisive architectural step occurred when the CV Engine, the Opportunity Engine, and the
broader analytical work of Virgili Studio were no longer treated as autonomous systems. The question
became: what functions remain necessary regardless of domain? Every application needed to be able to
register sources and their status; collect preliminary observations; transform those observations into
candidate objects; distinguish evidence, metrics, and projections; create relationships between objects;
validate the coherence of the graph; apply policies; preserve versions; register audits; read and export
governed views; and produce outputs without allowing generation to modify the canonical knowledge.

   From this convergence emerged the VCE: a governed canonical knowledge engine in which the
repository constitutes the only authoritative engineering memory. The conversation history is not
authoritative. The kernel is defined as domain-independent — its function is not to know about knitwear
or operating partners or investment opportunities, but to govern how any knowledge in any domain
acquires, maintains, and can lose its authority.

   The kernel's canonical lifecycle covers three primary object types: METRIC (a quantity defined
through an identifiable criterion), PROOF (an evidence that supports, qualifies, or contradicts a claim,
with a traceable connection to a source), and PROJECTION (a description of a future outcome or
possible transferability, explicitly labelled as such). The relations between these objects — which
PROOF supports which claim, which METRIC underlies which PROJECTION — are governed through
a validation layer that prevents unsupported relationships from entering the canonical corpus.

   The architectural decision to keep Capability, Buyer, Market, and Mandate outside the initial kernel
was deliberate. These categories were important for the applications but not yet stable enough to be
structurally encoded. Rather than transforming still-discussed categories into rigid structures, they were
maintained outside the certified core, pending approved domain packages. This boundary prevents the
kernel from becoming an indiscriminate container for all taxonomies produced during research.

   The generation process operates subordinate to this architecture. AI can assist in synthesis, in
proposing fields, in creating candidate relationships, in preparing review notes. Every model output
remains draft assistance until a human accepts it in a candidate package and the package passes the
kernel's boundaries. The system records the model used, references to the inputs, the output, and the
status of human acceptance.

   A critical principle governs the boundary between generation and canonisation: documents produced
for external use do not become canonical sources simply because they were approved or published. To
re-enter the system, they must pass through a new governed intake process. This rule prevents the system
from learning from its own formulations and from the most recent version of a text acquiring authority
over the source from which it derives.




6 What This Contribution Is and Is Not

The construction documented here does not claim to have invented provenance, knowledge graphs,
executive assessment, or opportunity scanning. All of these domains exist and are developed by
institutions, universities, technology companies, and international consulting practices. The contribution
is their operational configuration around a starting point that most enterprise architectures do not
address.

   Enterprise knowledge architectures typically begin from organisations that already possess data,
taxonomies, roles, processes, and decision criteria. The problem is to make them interoperable,
accessible, and usable by AI systems.

   Virgili Studio began from a prior condition: the experience existed but was not yet represented as
governed knowledge. The capabilities had been exercised but had not been separated from the
companies and roles in which they manifested. The evidence was distributed across CVs, documents,
conversations, projects, numbers, and personal memory. The ideas existed, but a chain was missing that
distinguished observation from thesis and thesis from opportunity. The method existed in practice but
had not been rendered independent of any particular domain.

   The result of the work is not a more elegant formulation of known principles. It is their transformation
into an operational sequence:

A career into a corpus of authority, capabilities, and evidence.

The corpus into differentiated market outputs.

Observations and frictions into evaluable opportunities.

The repetition of the method into a domain-independent kernel.

Generation into a process subordinate to canonical knowledge, rather than the reverse.


   The distinction most important to state explicitly: the existing literature on provenance seeks
primarily to reconstruct from where an output came. The Virgili Core Engine seeks also to govern what
status that output may acquire within the organisation. These questions are connected but do not
coincide, and the second has received less systematic attention.




7 Transferability


   The three problems documented here are not specific to the context in which they arose. They belong
to a class of challenges that recurs across any organisation whose primary competitive asset consists in
the capacity to produce, evaluate, and act on knowledge.

The discipline is transferable not because it can be replicated exactly in different contexts, but
because it addresses a structural problem — the distinction between observation, knowledge, decision,
and generation — that appears in any organisation that must build reliable representations of complex
realities.


For executive search and leadership advisory

   The evaluation of a candidate does not consist in the simple collection of information. It requires the
construction of a representation of professional authority, acquired capabilities, evolutionary potential,
and coherence relative to a specific organisational context. The same failure modes documented for
Alessio Virgili's profile — Translation Failure, Depth Failure, Domain Failure, Visibility Failure —
apply to every complex profile that does not correspond to a standard functional path. The CV Engine
provides a model for how capability, evidence, and context of transferability can be kept structurally
separate, producing different outputs for different buyers from the same verified corpus.


For private equity and operating partners

   Investment decisions are made on the basis of representations built progressively. The Opportunity
Engine addresses the phase before the object to be evaluated exists as a company or business case — the
point at which most due diligence processes do not yet have a disciplined method. The Kernel provides a
framework for connecting an investment hypothesis or a value creation thesis to the evidence, metrics,
capabilities required, and conditions that support its validity, while keeping projections explicitly
labelled as projections.


For venture builders and innovation teams

   The chain from observation to requirement to opportunity to evidence to decision provides a structure
for the phase that most innovation processes manage informally — the point at which a signal becomes a
concept and a concept claims to be an opportunity. The discipline does not eliminate intuition, which
remains essential in discovery; it prevents intuition from automatically becoming an investment thesis
without passing through a verification discipline.


For advisory firms and consulting practices

   Each mandate produces documents, analyses, and recommendations. The question the Kernel
addresses is whether the knowledge produced in one mandate is available — in verified, provenance-
carrying, epistemically labelled form — for a subsequent mandate. Most practices answer this question
informally, through the memory of people who were present. The governance discipline provides an
architecture in which knowledge produced in one context can be reused in another without losing its
connection to the conditions under which it was generated.

For any knowledge-intensive organisation

   The broader principle is applicable wherever organisations must build reliable representations of
complex realities to support decisions that matter. The question it poses is simple and demanding: does
the knowledge your organisation produces survive the project that generated it? Can it be interrogated
again without reconstructing the person who originally held it? Can its status — observed, verified,
inferred, projected — be read by someone who was not present when it was produced?

     When the answer to these questions is no, the organisation produces episodically. When
     the infrastructure exists to answer yes, it accumulates. The difference, compounded
     across projects and years, is substantial.




8 Discussion


   The literature most directly relevant to this work clusters around three areas: knowledge
management, provenance in AI systems, and executive capability assessment.

   The knowledge management tradition — from Nonaka and Takeuchi's distinction between tacit and
explicit knowledge, through Davenport and Prusak's working knowledge frameworks, to more recent
work on organisational learning — has long recognised that codification does not exhaust knowledge.
What the present work adds is an operational distinction between the preservation of information and the
continuity of the capacity for response: a dimension that becomes critical precisely when generative
systems are used to process and extend codified knowledge at scale.

   The AI provenance literature — NIST's AI Risk Management Framework, Microsoft Research's
VeriTrail, PROV-AGENT, and the emerging field of reasoning provenance — addresses the traceability
of outputs through multi-step generative workflows. The present work shares this concern but extends it
from the question of where an output came from to the question of what epistemic status it may
legitimately acquire. This extension is not merely terminological. It requires a different architectural
response: not a logging system but a governance system, with explicit lifecycle management for
canonical objects, policy enforcement at the boundary between observation and canonisation, and
structural separation between generation and authority.

   The executive capability assessment literature — including Spencer Stuart's movement toward skills-
based architectures and Korn Ferry's work on learning agility and capability transferability — has shifted
attention from role-based to capability-based candidate representation. The present work is built, in part,
from the same intuition, but from the executive's side rather than the search firm's side. The result is an
architecture that makes capability, evidence, and context of transferability structurally explicit — not as
an assessment tool applied from the outside, but as a self-representation system built from the inside.

Where this work departs most clearly from existing frameworks is in the insistence that generation
must be structurally subordinate to canonisation, not equivalent to it. The risk in most generative
workflows is not that the model produces incorrect outputs — it is that correct-sounding outputs are
treated as authoritative without having passed through the conditions that constitute authority. The
Kernel enforces this boundary architecturally rather than editorially, which is the difference between a
rule and a structure.




9 Limits of This Research


   This working paper documents a discipline developed within Virgili Studio's operational context.
The following limits apply.

   The research corpus is qualitative and single-organisation. The patterns identified emerged from a
specific professional history and were tested across a range of mandates and opportunity contexts within
that history. They have not been validated through independent organisations, controlled experiments, or
comparative studies against alternative methods.

   The Kernel V1.0 is implemented, tested, and certified as a technical baseline. Its architectural
contracts have been translated into code and passed through 174 tests. This demonstrates that the
architecture exists and its primary contracts have been verified. It does not demonstrate comparative
superiority over alternative knowledge governance approaches.

   The CV Engine has produced differentiated outputs and a substantially more rigorous corpus than the
starting documentation. It has identified real errors — narrative contaminations, overlaps between
entities and capabilities, insufficiency of evidence, visibility failures. Whether the capability-and-
evidence representation increases the quality of interlocutors, shortlists, mandates, and compensation
over time remains to be measured against a sufficient dataset of real candidacies.

   The Opportunity Engine has produced maps, rankings, and selection frameworks useful for internal
action. Whether the observation-requirement-evidence chain produces better investment or development
decisions than predominantly intuitive selection requires systematic comparative study across a
sufficient number of cases.

   The limits are stated not to diminish the contribution, but because the discipline documented in this
paper is founded on the principle that claims must be traceable to their evidence and projections must
remain visible as projections. Claiming comparative superiority without the evidence to support it would
be an instance of exactly the failure mode the Kernel is designed to prevent.

10 Conclusion


   Organisations are not entering a period characterised by scarcity of content. They are entering a
period characterised by scarcity of distinction.

   It will become increasingly easy to produce a synthesis, a strategy, a recommendation, a market
thesis, or a professional representation. It will become more difficult to establish what, within that
production, merits trust; what derives from a source; what has been inferred; which decision modified
the patrimony; which formulation is only a more effective way of communicating knowledge that
already existed.

   The three problems from which this work arose appeared different. The first concerned the possibility
of making a complex professional authority transferable without reducing it to a chronology or
transforming it into a narrative construction lacking evidence. The second concerned the possibility of
transforming observations and intuitions into evaluable opportunities without confusing the plausibility
of the solution with the existence of the problem. The third concerned the possibility of using generative
systems in a cumulative process without allowing generation to silently rewrite the knowledge on which
it depended.

   The common response was to separate four functions: observe, know, decide, generate.

   Observing means recording reality without attributing a definitive explanation to it prematurely.
Knowing means elevating some information through provenance, verification, relationship, and
governed acceptance. Deciding means applying judgment to a specific configuration and preserving its
reasons. Generating means translating the available patrimony into a useful form, without automatically
acquiring authority over it.

   This separation does not necessarily slow organisational intelligence. It makes it cumulative.

   The CV Engine, the Opportunity Engine, and the Virgili Core Engine are three applied responses to
this same structural challenge. They constitute an implemented and sufficiently formalised solution to be
subjected to real validation. Whether the principles they embody are transferable to other organisations
encountering the same structural problem — the management of knowledge across a growing body of
generative output, the representation of complex authority, the identification of real opportunities before
they have been defined — is the question this working paper poses to its readers.

     The primary limit of generative systems is not their capacity to formulate responses. It is
     the tendency of organisations to mistake production for knowledge. An organisation does
     not know more because it generates more documents. It knows more when what it
     observes can become reliable without becoming rigid; when what it decides preserves its
     connection to the reasons that made it possible; when what it produces does not cancel
     the difference between reality and representation; when every new work leaves a
     patrimony that survives its immediate result. It is in this continuity — and not in the
     volume of outputs — that knowledge becomes infrastructure.

Virgili Studio · virgilistudio.com · July 2026 · Working Paper — First Release · Not for distribution without permission