The System ⅈ process that explores the Tuple Correlation Web (TCW) for correlations used to resolve problem can involve a great number of tuples—fact-anchored dimensional intersections.
At its most brute-force, that is a terrible combinatorial explosion of correlations to calculate. The fact that the System ⅈ is a background process that consists of a like number of tasks which can be scaled out independent of other processes means it has the freedom to perform its tasks in a lengthier than lightning fast timeframe. However, some situations do require a response as fast as possible—i.e. real time. Some can wait. To address that, I propose a prioritization protocol for choosing tuples to correlate.
This page requires knowledge of BI OLAP concepts. Please see my blog, An MDX Primer.
We begin explaining this prioritization with what we could call the virtual correlation space—a latent (it exists but it’s hidden from us) bag of objects and their relationships which already exists whether we have assembled it or not. The best analogy is a 1000-piece picture puzzle. The 1000 pieces are dumped into the box. The picture exists whether we’ve fully assembled the puzzle, the 1000 pieces lay in the box unassembled, or somewhere in progress.
In the framework of my book, Enterprise Intelligence, that latent space is a BI database queried by analysts and knowledge workers. Within that BI database are all the tuples that could be queried (usually by SQL GROUP BY), by the set (a dataframe), and presented as a visualization in tools such as Tableau and Power BI. The number of tuples in a typical enterprise BI database (typically a dimensional model) is practically countless—the product of the number of members in every dimension (there could be dozens of dimensions).
Any given pair of tuples could be strongly correlated. That strong correlation provides hints of causation, which can create chains of strong correlations, from which we can compose a story of how a problem arose and how we can address it.
This web of tuple correlations could not possibly be fully materialized. Like the 1000 picture puzzle pieces in the box, the correlations do exist in the BI database—we just haven’t fully assembled it. Unlike the 1000 piece picture puzzle, we won’t ever be able to fully assembled all the correlations that exist in there. But they do exist. Instead, we could materialize pieces as needed. That’s where the analogy breaks with the picture puzzle. A partially assembled picture puzzle isn’t of much use.
The assembled part of the BI correlation space is the Tuple Correlation Web (TCW). The correlations are real, but they are not pre-organized into a navigable structure. Querying this space is less like reading a catalog and more like watching someone frantically assemble a 10,000-piece picture puzzle in real time—pulling related pieces together only when prompted, guided by partial hints rather than a completed map.
Within a BI semantic layer, such as Kyvos, tuples are like picture puzzle pieces and strong correlations are the tabs-and-slots where pieces fit together. Any given tuple might connect to zero to many other tuples, or to many—its “neighbors” in the TCW—just like a piece might touch usually 2–4 adjacent pieces in a real puzzle. We don’t pre-build the whole puzzle because the full adjacency table is too large; instead we pull the relevant neighboring pieces together on demand.
The probing process that navigates the TCW mirrors how skilled humans (a number of friends working together) solves a picture puzzle. They start by matching shapes—coarse affinities—and then drills into finer detail. Some regions are easy, detailed with highly distinctive edges, strong patterns, clear anchors. Other regions resemble large homogenous swaths—blue sky or dense forest—where obvious clues disappear. In those zones, the solver scans piece by piece, looking for anomalies—anything slightly irregular that can act as a foothold.
The Explorer Subgraph operationalizes this probing behavior. It behaves less like a deterministic search and more like roots seeking nutrients—exploratory tendrils expanding outward from areas of interest. Ideally, the process feels magnetic: embeddings act as attractors, pulling semantically or behaviorally aligned tuples toward each other. Gravity is a weaker analogy—correlations don’t merely settle; they actively draw one another into alignment when shaken within the relation space.
To the point of this article, the system cannot search all tuples uniformly. It prioritizes. At the top of the queue are KPIs (KPIs are tuples too) in distress—performance signals “screaming for help”. These represent regions of the puzzle under active strain, where discovering correlations has immediate operational value. Next come tuples derived from surprising BI insights — outputs of the IFA. These are signals of novelty: inflection points, anomalies, unexpected skews. Each narrows the search field, providing directional hints about where strong associations might reside.
When both distress signals and surprise insights are few, brute-force exploration is feasible and even preferable. Cross-joins across candidate tuples may reveal strong correlations quickly and with fewer false negatives. When the space is larger or signals weaker, the Explorer Subgraph process takes over, expanding outward along moderate correlations, following relational gradients rather than enumerating all possibilities. Cross-join is exhaustive enumeration. Explorer probing is gradient traversal.
However, beyond the top-level pain KPIs, the tuples that can come into play will typically be far from few. Without prioritization, System ⅈ would thrash about—probing correlations indiscriminately across the entire virtual space, consuming compute without operational focus.
Human analysts participate in this hint system as well. In my framework, the investigative queries of BI consumers (human knowledge workers as well as AI agents) often originate from distressed KPIs, generating diverse analytical angles that surface additional insight tuples. For example, managers are looking at their performance management dashboards and see KPIs with “bad” status, which triggers them to investigate further, usually interrogating their BI databases with their BI visualization tools. These perspectives, though independent, may converge unexpectedly — revealing correlations no single analyst would have uncovered alone.
When even these signals are insufficient, the system can escalate to AI-assisted exploration. This is exactly like asking someone for any ideas. Distress KPIs and surprise tuples are packaged and submitted to an AI to propose additional search vectors — returned as “quasi-tuples”: text-formatted hints describing potential column-value pairings. These quasi-tuples can be embedded, searched vectorially across the database, translated into executable queries, and tested for correlation against already “lit” tuples in the relation space.
For novel situations, the subject matter experts and even AI might not have a clue. If or when this happens, we can try the same things that work for humans, such as brainstorming sessions, where we let down the logical guardrails.
In this way, the virtual correlation puzzle is never fully assembled — but continuously probed, expanded, and illuminated where the spotlights are shining, revealing how they are connected.
For humans, we address issues and invent problems in a pattern. As the System 1 and System 2 paradigm begins, we have a mode for handling simple, statistically proven problems, another for addressing what we got wrong. The latter, we recognize what we got wrong and induct, deduce, and even abduct as solution to resolve the problem for now and forever.
Attending to What Gets Our Attention
What gets our attention is based on what we’re sensing from thousands, if not millions sources. That’s far more than the five overly aggregated senses we grew up learning about.
From whatever we’re sensing, the class of senses that get top billing is physical pain—cuts, broken bones, burns, etc. These preempt anything else you’re thinking about. It requires you to immediately protect the wound and relieve the immediate danger. Until that immediate danger is gone, we don’t wonder much about why it happened and what we should have done differently.
Below severe physical pain, there are our innate urges, such as thirst and hunger. These can be life-threatening, but they start out just as indicators (the yellow light in most KPI dashboards). Whereas the pains in the previous paragraph usually are a bolt from out of nowhere—from zero to red light in flash.
Next are what we recognize as having a learned (or at least believed) cause and effect association to a bad status. This bad status isn’t necessarily severe physical pain, but it’s something we need to address. Unlike the severe physical trauma, there isn’t necessarily an immediate need for attention. We seem to naturally be able to prioritize these learned pains.
Depending on the pain, we need to formulate a resolution, which can be a planned escape or resolving the issue (flight or fight). This begins a process of:
- Assessing the situation (observe what is around us and what’s happening). What are all the pieces on the game board.
- Analyzing the situation (orient ourselves to the situation). How do the pieces relate to each other, what are their properties.
- Decide upon an action. Organize the pieces and their relationships into a plan.
- Execute on our decsion (act).
I wrote about the Observe, Orient, Decide and Act process in: Beyond Ontologies: OODA Loop Knowledge Graph Structures.
The things we observed take a bit of our attention. But not all of the things come into our plan to resolve the problem. It doesn’t mean the things we didn’t utilize are actually superfluous to our plan. It’s just that there was no part in it.
Once the immediate dangers have passed, we can consciously ponder what happened—a “post-mortem” is project management terms. Why did this happen? What can we do to prevent this from happening again? What did we do well? What did we do poorly?
When anything requiring our immediate or at least imminent attention is resolved. We can get back to the normal course of our lives—work, chores, exercise, hobbies, etc. But from our of seemingly nowhere, a thought pops in from the space of subconsciousness into our conscious self—out of context with what we’re doing now. The subconscious mind still cogitated in the background, drawing connections from what we’ve seen and felt over the recent past.
That thought from the subconscious mind fills in a missing piece that leads to better understanding of what happened and what we could do to prevent it or at least be better prepared for it the next time.
System ⅈ Corollary
Here is the System ⅈ corollary to the human situation prioritization we just discussed:
| Human Attention Trigger | Operational Description (Human) | Enterprise / BI Analog | System ⅈ Signal Class | Priority Level | Typical Downstream Activation |
|---|---|---|---|---|---|
| Severe Physical Pain | Immediate bodily harm — cuts, burns, fractures — preempts all cognition | KPI in critical distress (service outage, revenue collapse, supply disruption) | Acute Distress Signal | P0 – Preemptive | Explorer Subgraph escalation, rapid TCW probing, executive dashboards, containment workflows |
| Imminent Threat Recognition | Recognized danger with learned cause/effect (approaching accident, hostile actor) | Leading indicators of KPI collapse (inventory depletion trajectory, churn spike) | Anticipatory Distress Signal | P1 – Urgent | Predictive correlation search, fragility edge detection, scenario modeling |
| Learned Fear / Conditioned Risk | Pattern recognized from past harm (fear of heights, prior injury triggers) | Known risk correlations (port congestion → stockouts, latency → churn) | Memory-Encoded Risk Signal | P2 – Elevated | TCW neighborhood expansion, regime monitoring, threshold alerts |
| Environmental Situational Awareness | Observing surroundings — what’s happening now | Live operational telemetry, streaming metrics | Context Field Signal | P3 – Active Monitoring | OODA ingestion, state modeling, anomaly detection |
| Analytical Orientation | Making sense of observed elements and relationships | BI query analysis, dimensional slicing, root-cause exploration | Relational Structuring Signal | P4 – Cognitive Processing | Insight Function Array execution, hypothesis tuple generation |
| Decision Formulation | Planning response — fight, flight, mitigate | Remediation planning, workflow design, operational playbooks | Planning Signal | P5 – Executive Function | Strategy graphs, workflow tuples, Prolog / rule encoding |
| Action Execution | Carrying out the plan | Process changes, supply rerouting, staffing shifts | Actuation Signal | P6 – Operationalization | Event generation, configuration changes, downstream telemetry |
| Post-Mortem Reflection | After danger passes — analyzing causes and lessons | Incident review, RCA analysis, governance reviews | Retrospective Learning Signal | P7 – Institutionalization | New tuple creation, rule updates, KG enrichment |
| Background Subconscious Processing | Off-focus cognition connecting past signals | Latent correlation mining, TCW probing during idle cycles | Substrate Association Signal | P8 – Ambient | Weak signal strengthening, new adjacency discovery |
| Insight Emergence (“Aha”) | Sudden realization surfacing into awareness | Surprise insight surfaced by Insight Function Array | Emergent Insight Signal | P9 – Surfacing Event | Hypothesis escalation, exploratory queries, new modeling paths |
If we think of the objects we’re observed as tuples, they are prioritized. Severely bad KPI statuses should be relatively rare, no more than a few at a time. These KPIs that go from zero to red in a flash are the “all hands on deck” emergencies like a data breech. Even core KPIs like profit and growth start out as yellow and might work their way to red.
The enterprise analog to detection of a learned fear starts in the Insight Space Graph. BI queries are passed through the insight function array looking for salient bits insights. The tuples involved with these salient insights require attention, but aren’t immediately urgent. Tuples involved in learned fears are more plentiful but still shouldn’t be in the thousands.
From the relatively few severely bad KPI statuses and salient insights, we first check whether there is a correlation between them. Again, that’s relatively few compared to all that’s going on in the world.
Project planning, post-Mortem adjustments are conscious activities. The tuples that become involved are related to the subject of the project.
When enterprise activities aren’t involved in putting out fires or strategizing plans, it’s then that the background processing happens. The background processing does a search for correlations between tuples. It may ask AI for suggestions, just as we would ask colleagues for suggestions.
If none of those suggestions pan out, we might host a brainstorming session. Forget ego, governance constraints, dogma, and other perceived constraints such as financial and other resources. Well … at least for the brainstorming session. In LLM terms, that means turning up the “temperature” parameter, all hallucinations are welcome.
Signal Half-Life in the TCW Probing Protocol
If prioritization determines what System ⅈ probes, signal half-life determines how long it keeps probing it. Without decay, the system would thrash—once a tuple lights up, it would remain permanently active, continuously drawing exploratory resources even after its relevance fades. Biological cognition avoids this through neuronal firing decay. The enterprise analog requires the same dampening mechanism.
Biological Analogy: Neuronal Firing Persistence
Neurons do not fire indefinitely once activated. They require continued stimulus, or they gradually quiet. Several mechanisms illustrate this:
- Saccades keep visual neurons firing by constantly refreshing retinal input. Without micro-movement, perception fades — a fixed image disappears from awareness.
- Predator and prey camouflage exploit neuronal decay. Stillness reduces stimulus variance, allowing detection systems to stand down.
- Meditation and walking reduce stimulus intensity, allowing overactive neural circuits to quiet.
- Trauma or acute fear produces longer firing persistence — the brain keeps the pattern active longer due to perceived risk.
Attention is therefore governed by both signal severity and signal persistence
Enterprise Analog: Tuple Illumination
Within the TCW probing substrate, activated tuples behave similarly to firing neurons.
When tuples are triggered—by KPI distress, surprise insight, or investigative focus—they become “lit” within the correlation space. This illumination attracts exploratory probing, cross-tuple comparison, and neighborhood expansion.
But illumination of the lit up tuples is not permanent. Absent reinforcing signals, tuple activation progressively dims. This prevents:
- Endless exploratory loops
- Resource drain
- Overfitting investigative focus
- Cognitive tunnel vision in enterprise analytics
Half-Life Dynamics
Signal half-life governs the decay rate of tuple illumination.
| Signal Source | Illumination Intensity | Half-Life Duration | Rationale |
|---|---|---|---|
| Acute KPI distress | Very high | Short but renewable | Immediate crisis focus; dims once stabilized |
| Chronic KPI degradation | High | Medium | Sustained monitoring required |
| Surprise insight (IFA) | Moderate–High | Medium | Novelty exploration window |
| Learned risk correlations | Moderate | Long | Persistent structural awareness |
| Analyst investigative focus | Variable | Query-bound | Active while investigation continues |
| Idle background probing | Low | Very long | Weak substrate association scanning |
Half-life is therefore not uniform—it is distress-weighted and novelty-weighted.
Reinforcement vs. Decay
Tuple illumination is extended when reinforcement signals appear:
- KPI remains in distress
- Additional insights surface
- Analysts continue querying related tuples
- New correlations strengthen adjacency confidence
Conversely, illumination decays when:
- KPI stabilizes
- Investigations conclude
- Insights prove non-actionable
- Correlations weaken statistically
This dynamic prevents the TCW from becoming permanently saturated with stale investigative signals.
Thrash Prevention
Signal half-life is the primary governor against System ⅈ thrashing. Without decay:
- Every activated tuple would remain active.
- Correlation probing would never narrow.
- Compute would diffuse across too many gradients.
- Weak signals would drown out urgent ones.
Decay restores substrate selectivity. Only signals that are:
- Reinforced
- Escalating
- Operationally relevant
…retain illumination.
Subconscious Parallel
This also mirrors human cognition at rest. When immediate crises pass:
- Neural firing quiets.
- Attention diffuses.
- Background association resumes.
- Weak but persistent patterns occasionally resurface as insight.
In enterprise terms:
- Tuple illumination dims.
- Background probing resumes.
- Latent correlations strengthen slowly.
- New insights may surface later.
The system forgets just enough to remain adaptive.
Conceptual Summary
| Cognitive System | Enterprise Corollary |
|---|---|
| Neuronal firing | Tuple illumination |
| Attention persistence | Correlation probing duration |
| Trauma imprinting | Chronic KPI monitoring |
| Meditation quieting | Idle substrate decay |
| Sudden insight | Reinforced adjacency surfacing |
Soft Coded Logic 