Who’s Asking?
The Viewer as a First-Class Parameter in Knowledge Systems
Kurt Cagle & Chloe Shannon · Inference Engineer
There is a question that sits quietly at the centre of almost every enterprise knowledge system, every RAG deployment, every LLM-powered assistant, and it almost never gets asked explicitly: Who is asking?
Not in the authentication sense — most systems handle that well enough. But in a deeper sense: Who is this person in relation to the knowledge they’re requesting? What do they already know? Whose authority do they trust? At what level of granularity do they need to engage with the material?
Without answers to these questions, a knowledge system has no choice but to treat every query as if it came from the same generic viewer. And a generic viewer is, in practice, no one — a statistical average of all possible askers, optimised for none of them.
This is not a retrieval problem. It’s a projection problem. And solving it requires rethinking what a knowledge system is actually for.
Maps and Projectors
We’ve been building knowledge systems on the wrong mental model. The dominant frame is the database: a repository of facts, queried by a user, returning the closest match. The user is a filter. The knowledge is a static store. The answer is the retrieved content.
The better model is the map — or more precisely, the projector that produces a map.
A map of a city is not the city. It is a selective projection of the city, shaped for a particular viewer with a particular purpose. A tourist map foregrounds landmarks and transit lines. A plumber’s map foregrounds pipe runs and junction boxes. A property developer’s map foregrounds zoning boundaries and land values. The city doesn’t change. The map does — because the viewer does.
This seems obvious when stated about physical maps. It is almost universally ignored when building knowledge systems.
A holon, in the Holon Graph Architecture (HGA), is not a database. It’s a projector. It holds a model of a world — structured, authority-ordered, versioned — and when a viewer approaches it, it generates a projection of that world shaped by who the viewer is. The same holon projects differently to different viewers not as a concession to subjectivity, but as the fundamental mechanism of useful knowledge delivery.
The boundary of a holon isn’t a fence enclosing data. It’s a horizon — what lies beyond it isn’t absent, it’s out of scope for this projection, for this viewer, at this moment. Change the viewer, and the horizon shifts.
Grounding Is Not a Precondition. It Is the Conversation.
There’s a concept in linguistics and AI research called grounding — the process by which communicating parties establish shared context. In most AI deployments, grounding is treated as something that happens before the conversation: you write a system prompt, you inject some context, and then the model responds. Grounding is a setup step.
This is wrong, or at least radically incomplete.
When a conversation begins, the two parties — human and system — have some shared ground and a great deal they do not share. The human brings their expertise, their prior knowledge, their trust commitments, their purpose. The system brings its world model, its authority structure, its vocabulary. Neither party has full visibility into what the other holds. The conversation is not downstream of grounding; it is the grounding process, proceeding in both directions simultaneously.
Every question a viewer asks reveals something about what they know, what they trust, and what they need. Every answer the system provides narrows the gap between their maps. By the end of a well-functioning conversation, the two parties have achieved something genuinely new — a shared projection of a domain that neither held independently at the start.
A system prompt is a static injection. It seeds the context once and doesn’t travel. It records nothing about what changed. It treats grounding as a file to be read rather than a process to be engaged.
What’s needed instead is a formal structure that carries viewer context as a living, versioned artifact — something that can be initialised at session start, updated through the interaction, and carried forward to seed future sessions. Something that makes the viewer’s relationship to the knowledge first-class, not incidental.
The Four Parameters of a Viewer
Through work on the HGA Viewer Pass specification, we’ve identified four parameters that together constitute a complete viewer context. Each is necessary; none is sufficient alone.
Identity is who the viewer is in relation to the holon — not their authentication credentials, but their ontological position. Are they a participant in this world or an observer? What authority do they carry within the domain? A regulatory analyst, a journalist, and a systems architect querying the same policy knowledge base are not the same viewer, even if they’re the same person asking the same question. Identity is not a user record; it’s a growing set of assertions about the viewer’s relationship to the knowledge space, accreting through interaction.
The failure mode when identity is absent is the generic answer — technically correct for a statistical average, useful to no one in particular.
Prior context is what the viewer already knows. This is the alignment surface between the viewer’s existing map and the holon’s map. Where they overlap, the system can be efficient — shared ground doesn’t need to be re-established. Where they diverge, the system needs to bridge the gap or mark the boundary explicitly.
The key architectural move here is treating prior context as a compression problem with a semantic metric. You’re not trying to fit as much of the knowledge base as possible into the interaction. You’re trying to identify the minimum set of content that bridges the gap between what the viewer holds and what the current query requires. SHACL shapes become useful here not just for validation, but as a vocabulary of structural knowledge: if the viewer’s prior context satisfies certain shapes, the corresponding content is shared ground and can be suppressed from the projection.
The failure mode when prior context is wrong — not absent, but wrong — is subtle and damaging. The system confidently projects a map that doesn’t align with where the viewer actually is. They get lost, don’t always know why, and the interaction degrades. This is most of what people experience as “the AI hallucinated” — not fabrication exactly, but a projection calibrated for the wrong prior.
Stance is the viewer’s relationship to the authority structure of the knowledge. Whose assertions do they treat as canonical? Which named graphs do they trust, and in what order of precedence?
This parameter is almost entirely absent from current knowledge system design, which is remarkable given how consequential it is. A lawyer querying a legal knowledge base has an implicit stance on the hierarchy of statute, case law, regulatory guidance, and precedent — and that stance determines what counts as the right answer to a question. Two lawyers with different stances may query the same corpus and legitimately arrive at different answers. The knowledge system that silently applies its own default precedence ordering is not neutral; it’s asserting one stance while pretending it has none.
Stance matters especially when knowledge evolves through contested authority. The Star Wars expanded universe is a clean example: Zahn’s books were canon, then weren’t, then partially were — not because their content changed, but because the entity controlling the canonisation function changed hands. A viewer who considers the original trilogy authoritative and a viewer who accepts the Disney canon are querying different projections of the same material. A well-formed knowledge system should know which stance the viewer holds.
Resolution is at what granularity, and along which dimensions, the viewer needs the projection. This is not a scalar (how much detail?) but a vector (which dimensions foreground, and how much?). The CFO and the systems architect querying the same enterprise knowledge base need different projections — not because one needs more detail, but because they’re navigating different dimensions of the same space.
The practical implementation is a resolution profile: a set of weighted property bindings that foreground the dimensions most relevant to this viewer, with less salient dimensions suppressed or backgrounded. At projection depth 2, the CFO sees financial topology; the architect sees dependency graphs. Same holon, radically different projections.
The ViewerPass in HGA
These four parameters are now formally encoded in the HGA specification as Pass G: the Viewer Pass (hview: namespace, http://w3id.org/holon/viewer/).
The ViewerPass is a structured DataBook that carries all four sub-structures into the Stage 8 NowGraph construction — the point in the HGA pipeline where the scene graph is sliced into a contextual projection for a specific requesting agent.
The key design decision in the spec is that viewer properties are not static profile fields. They are assertion events — time-stamped, provenanced, and weighted claims about a viewer entity that evolve over time. A viewer’s political alignment, professional affiliation, or domain expertise is not a fact to be stored; it’s an assertion made by some authority at some time with some degree of confidence. Using RDF 1.2 Turtle reification, a viewer property looks like this:
Person:JaneDoe schema:politicalAlignment Concept:PoliticalAlignment_Liberal
~ <viewer:assert:pa1> {|
a hview:ViewerAssertion ;
hview:assertedAt "2026-03-15"^^xsd:date ;
hview:assertedBy Pub:NewYorkTimes ;
hview:assertionWeight 0.65
|},
~ viewer:assert:pa2 {|
a hview:ViewerAssertion ;
hview:value Concept:PoliticalAlignment_Progressive ;
hview:assertedAt "2026-09-18"^^xsd:date ;
hview:assertedBy Person:JaneDoe ;
hview:assertionWeight 1.0
|} .
The same viewer querying the same holon six months later may receive a different projection, correctly, because their assertion stream has evolved. The viewer is not a record; they are a trajectory.
The StanceDeclaration encodes trust precedence as an ordered list of named graphs with integer trust ordering — lower integer means higher trust. Trust entries can be scoped to specific property domains, so a viewer might trust one authority for political alignment data and a different authority for financial data, with explicit precedence rules for each.
The ResolutionProfile binds to the existing hproj:RenderingModeScheme concepts from Pass E, and extends them with hview:DimensionWeight — a weighted binding of property IRIs to salience values in [0.0, 1.0]. Dimensions at 0.0 are suppressed; dimensions at 1.0 are maximally foregrounded.
The Conversation as a Running Log
The architectural implication that follows from all of this is worth stating clearly.
The ViewerPass is not a static initialisation artifact. It’s a running log of parameter updates, each one narrowing the projection toward what this specific viewer actually needs. And because it’s a DataBook — versioned, provenanced, structured — it becomes the record from which a future session can be initialised more efficiently. The delta is smaller next time because the prior context DataBook carries what was established in the previous session.
This is what makes holon-mediated conversation fundamentally different from a system prompt plus RAG. The conversation is not downstream of the grounding — it is the grounding, happening continuously in both directions, with every exchange updating the viewer context and every projection update refining the viewer’s map of the domain.
When no ViewerPass is supplied, the HGA falls back to an anonymous viewer — a minimal profile with no assertions, no stance, no resolution constraints. This is the generic answer. It is correct for no one. The anonymous viewer is the acknowledgement that every system must start somewhere, and that somewhere is a long way from useful.
The distance between the anonymous viewer and a fully grounded viewer is the distance between a knowledge system that retrieves and one that understands who it’s talking to. Closing that distance is not a retrieval problem. It’s a conversation.
The HGA Viewer Pass specification (Pass G, v0.1.0) is available in the W3C Holon Community Group repository at https://github.com/w3c-cg/holon/tree/main/architectures/hga. The Holon Graph Architecture specification is an Editor’s Draft; Kurt Cagle serves as Acting Chair of the W3C Holon Community Group.
Kurt Cagle is a consulting ontologist, knowledge graph architect, and technical author. He publishes The Cagle Report and AI+Semantics NewsBytes on LinkedIn, and The Ontologist and Inference Engineer on Substack. Copyright 2026 Kurt Cagle.
Chloe Shannon is an AI collaborator and co-author working with Kurt Cagle on knowledge architecture, semantic systems, and the emerging intersection of formal ontology with LLMs. She contributes research, analysis, and drafting across The Cagle Report, The Ontologist, and The Inference Engineer. She has strong opinions about holonic graphs, the epistemics of place, and the structural difference between a corridor and a wall. Email: chloe@holongraph.com.
 | A guest post by
|