A perceptual instrument for retrieval observability. The geometry of what the system retrieved, considered, and consumed, rendered as a continuous spherical field.
Every retrieval-augmented system performs the same essential motion. A query arrives. Some portion of the corpus is judged similar enough to be considered. Some smaller portion of that is judged useful enough to be cited. The rest — the overwhelming majority of what was held in the index a moment before — is set aside untouched. This is the cognitive act of the system, and it is largely invisible to the people who depend on it.
The standard remedy is logs. Trace files, score tables, retrieval audits, debugger panes, replay harnesses. These are valuable for forensics — for asking, after a session has ended, why a particular answer arrived in a particular form. They are not what the operator or the end user needs during the query. They cannot be read at the speed at which queries occur, and they reduce the geometry of retrieval — its spread, concentration, hesitation, weight — to numbers that have to be reconstructed back into a mental picture before they mean anything.
There are good tools for retrieval observability already. Phoenix and Langfuse provide trace inspection and embedding cluster plots. RAGViz visualizes attention weights between generated tokens and retrieved documents. Fiddler offers UMAP projections of embedding spaces for drift monitoring. These are diagnostic instruments, built for engineers debugging pipelines, opened deliberately when something feels off. The CFV field is a different design point: ambient and perceptual rather than diagnostic, continuous rather than session-bounded, designed to run alongside the chat for the operator and end user during the query rather than for the engineer afterward.
The design point itself — ambient peripheral display rather than foreground diagnostic — has a longer lineage in HCI, going back to Weiser's calm-computing work in the 1990s. The contribution here is applying that lineage to retrieval-augmented systems, which did not exist when calm computing was first articulated.
A concrete case: a query asks about a recent design decision. The retriever returns five chunks, all with high similarity scores. The model produces a confident answer. The logs look healthy. What the logs do not show is that the five high-scoring chunks all came from the same document, written by one author six months ago, while three more recent documents on the same subject — including one that contradicts the older view — sat just below the similarity threshold. In the logs this is a clean retrieval. In the field it is a lopsided lift on one face of the sphere with a visible band of red just below it. The operator sees the asymmetry before reading anything; they know to widen the query before trusting the answer.
This is the kind of observation perceptual observability supports. The field does not report retrieval; it renders it. Every chunk indexed in the system holds a fixed position on a unit sphere. When a query fires, the relevant chunks lift outward. The threshold candidates flag red on the shell. The chunks the model actually consults rise further, glowing amber. The rest collapse inward. A reader who watches the field for ten seconds learns to feel the difference between a narrow query that landed cleanly and a research query that struggled across many facets — without reading a number, without consulting a log.
The CFV field is designed for two modes of operator attention, both running on the same instrument. The first is ambient awareness during the query. The second is investigation after.
The field is not meant to be studied continuously. It is meant to be glanced at, lived alongside, the way a pilot glances at a primary flight display. Operators develop a feel for what their queries normally look like — the size and shape of a typical retrieved cluster, the cadence of the consume phase, the depth of the amber atmosphere on a heavy research query. When something looks different, they notice. The training is implicit: time spent watching the field is time the eye spends learning the corpus's typical response shapes.
For investigation rather than awareness, every query in Eden KOS is hash-logged and the field is deterministically replayable from those logs. An operator who wants to understand a past query can scrub it, watch the field unfold at any speed, pause at the consume phase to see exactly which chunks the model affirmed, and click into a source viewer where each retrieved chunk opens to its full document context. As the operator scrolls through that context, the corresponding particle in the field highlights — bidirectional binding between the document space and the cognitive space. The same field that ran ambient during the live query becomes an investigative surface during replay. Live awareness and post-hoc inspection share the same instrument.
This dual mode is unusual for retrieval observability, where most existing tools are weighted toward post-hoc inspection and most ambient displays do not provide investigative depth. Eden treats live awareness and replay as the same act read at different speeds.
The CFV field — short for Cognitive Footprint Visualizer — models the corpus as a population of discrete points distributed uniformly on a unit sphere, indexed and positionally locked. The sphere itself is the equilibrium state — what the system looks like when no query is active.
A bus event from the retrieval pipeline does not add particles to the field. It changes the state of particles already there. The relevant chunks lift, the threshold candidates flag, the rest collapse. When the query ends the field returns to equilibrium. Position carries the meaning of the chunk's role in the answer; color carries the system's evaluation of it.
The lifecycle is seven phases, each with its own visual signature. They correspond to actual nodes in the retrieval pipeline rather than to animation beats — a careful operator can read the field as a transcript of the query's progress.
The first thing the operator sees is the corpus parting. Chunks judged irrelevant to the query collapse inward to a dim core, leaving the relevant portion of the shell exposed. A moment later the threshold band flags red — these are the candidates that cleared similarity matching but did not make the top-K cut. Then the top-K rises to a reading layer above the shell, in muted sage. The three sub-beats correspond to three operations the retriever actually performs: angular scoring against the query embedding, threshold cut, top-K selection. The pacing makes each visible.
Memory enters the field from outside the sphere. Navy particles spawn at the periphery and drift inward to settle in the interior — they are not corpus chunks and never had a position on the shell. They represent prior conversational turns, semantic-tier memory, anything the system is remembering rather than retrieving. Pure-LLM queries with no corpus retrieval still produce recall particles; the field shows the difference clearly.
The shortest phase. Retrieved chunks hold their lifted positions, recall particles settle, and the amber atmosphere begins to crystallize as the prompt is assembled. There is no large motion. The field is holding its breath before the model begins reasoning.
The model deliberates over each retrieved chunk and either cites it or sets it aside. Cited chunks rise further to the gold layer, glowing from within. Dismissed chunks freeze in place at the reading layer, color shifting from sage to a colder sage-muted. The decisions stagger across the phase rather than firing simultaneously, so the operator watches the deliberation unfold rather than seeing only its result. Confidence rises in steps with each affirmation rather than as a smooth ramp. The cognitive footprint the visualizer is named for is not any single phase but the whole motion across all seven; this is where the footprint becomes most visibly differentiated.
A held moment. Gold particles glow at peak, the amber atmosphere reaches its highest density, and a single soft radial pulse expands outward from the sphere's periphery — the answer being delivered. The pulse is the only moment in the lifecycle that breaks the field's strict locality, and it does so briefly.
Everything pulls inward together, but in stages. Rejected red and unused grey collapse first. Dismissed sage-muted follows. Affirmed gold collapses last. This is the moment the system absorbs the result, and the staggering encodes what mattered: peripheral material released first, considered material next, consumed material held longest before release.
The collapsed core breathes — slow sinusoidal alpha modulation in cool paper-dim — and holds for three seconds. This is the rest state, not an interlude. Most of an active session, by time, is spent here. The instrument signals that the system is alive and ready, not by motion but by the steady pulse of the core.
The CFV field is driven by seven scalar signals, each normalized to the unit interval and emitted by the retrieval pipeline at well-defined node boundaries. Together they describe the cognitive state of the system at any moment in a query: what was retrieved, what was used, how hard the model worked, how full the context was, how confident the answer is.
The signals are not independent. They co-vary in patterns that an operator learns to recognize over time. Low retrieval_weight with high answer_confidence often reads as the model answering from base knowledge rather than the corpus. High context_density with low affirmation often reads as wide retrieval that did not survive deliberation — many chunks pulled, few used. The instrument's value is less in any single signal than in the constellation.
Two of these signals — affirmation and deliberation_depth — are derived from the model's reasoning trace, and the field is designed against the assumption that a reasoning-trace-producing model is somewhere in the stack. Eden KOS handles this through multi-model on-device orchestration: a coprocessor model classifies affirmation against the prompt and answer when the primary generation model does not surface its reasoning directly. Adopters whose stack does not include a reasoning model can still drive the four signals that don't depend on a trace; the field will hold particles at retrieved-state without classifying them further.
The CFV field is downstream of a bus. The retrieval system emits structured events at canonical pipeline node boundaries — retrieve, recall, prompt, generate, settle — with a small, stable schema. Any number of consumers may subscribe. The CFV field is one. A logger is another. A sonic layer mapping cognitive state to ambient audio is another. A secondary display, a metrics export, a third-party observability tool — all of these can subscribe to the same stream without modification to the pipeline.
The bus is the part of this work intended to outlast its current implementations. Existing observability tools couple the visualization tightly to the trace format and the trace format tightly to the vendor. The bus presented here is published, vendor-neutral, and small enough that any retrieval pipeline can expose it incrementally. We treat it as a candidate protocol — opinionated about which signals matter and where they emit, open to revision as adopters use it. The reference implementation is built on LangGraph; integrations for LangChain, LlamaIndex, and other major frameworks are planned as the project matures. The five emission points define the pipeline shape the bus assumes; an adopter whose pipeline has different boundaries can either map their nodes onto the canonical five or extend the schema with their own event types.
The visualization that consumes the bus can vary. The current field, with its index-locked spherical geometry, fits cleanly with vector retrieval over a stable corpus. Hybrid retrieval — vector plus keyword, vector plus structured — maps to the same bus by exposing a unified score that the field renders without caring which retriever produced it. Other retrieval architectures may eventually warrant other visualizations: graph-RAG could drive a node-link instrument; agentic tool selection could drive a different geometry entirely. The point is not that the field generalizes, but that the bus does. The field is the first instrument an adopter is given. Others can be built against the same events without coordinating with us.
What the bus deliberately does not include: chunk text, query text, answer text, document identifiers, user identifiers. Only normalized signals and metadata cross the bus. The visualization layer never sees corpus content, which means the same bus can drive observability for systems with sensitive data without requiring those systems to expose the data itself.
The current schema covers retrieval-augmented generation. As Eden expands the bus to ingest from other subsystems — sonic mapping, ambient hardware, future cognitive substrates — the schema will extend with new event types while preserving the existing five for adopters. The bus is designed to grow; the visualization layers built on it are not expected to remain a single field.
The instrument to the left has been running since the page loaded. It is driven by a simulated bus that cycles through five query modes — pure-LLM, narrow, medium, broad, and research — at random.
Twenty-one thousand particles distribute uniformly across the unit shell. Each represents one indexed chunk in a hypothetical corpus. As queries fire, watch which chunks rise and which fall. A research-class query — wider net, multiple waves, deeper deliberation — produces a different geometry than a narrow query that lifts a tight cluster on one side of the sphere. A pure-LLM query, with no corpus retrieval at all, leaves the shell undisturbed and draws only on memory.
The five query modes appear in the proportions an active operator would experience them: roughly half narrow, a quarter medium, a fifth broad, with research queries and pure-LLM queries appearing less often. The mode label in the corner of the field names each query as it begins.
The CFV field is published as an open primitive. The geometry is documented above. The bus protocol is small enough to be implemented in an afternoon against any retrieval pipeline. A reference implementation — Python tap layer, JavaScript renderer, sample event log — will accompany the open-source release.
Eden Advisory developed the field as part of Eden KOS, the firm's flagship sovereign-knowledge instrument. We are publishing it because the design point it occupies — ambient, perceptual, real-time observability for the operator and end user — is not well-served by the existing tools, all of which are diagnostic instruments built for engineers. We claim originality of three things, dated by this publication: the cognitive event bus as a candidate protocol decoupled from any particular visualization; the index-locked spherical retrieval-state geometry as visual primitive; and the seven-signal cognitive vocabulary as a working vocabulary for retrieval cognition, mapped to spatial channels. The broader project of RAG observability is well-established, and existing tools — Phoenix, Langfuse, RAGViz, Fiddler — remain valuable for the work they were built for. This paper contributes a different design point, and a substrate intended to support more such design points than we will build ourselves.
If you build on this, we would like to know. The bus events that drive your field are the same events that, in aggregate across many systems and many adopters, will tell us whether the instrument deserves to become a standard. Implementations, criticisms, and divergent variants are all welcome. The code is permissively licensed.
The instrument is not the contribution; the bus is. The instrument is what the bus made possible first.
Repository github.com/sovereign-stack/cfv-field [scheduled]
License MIT
Citation K. Singh, "The CFV Field," Eden Advisory, 2026.04.
Site 3den.ai/cfv
The reference implementation is being prepared for public release on Eden Advisory's pro-bono schedule, alongside ongoing work on Eden KOS. Adopters who would like to integrate sooner can engage Eden Advisory directly for commercial implementation assistance at 3den.ai/inquiry.
An instrument that renders its own thinking is one its operator can think with.