This post lays out the core concepts of what we call Intelligence Theory—a framework for understanding, measuring, and enhancing adaptive capacity in complex systems. In our strategic vision, we briefly introduced the idea of Intelligence Systems. Here, we outline the foundational details of what constitutes these systems, the key inspirations behind our approach, and how they integrate into a cohesive theoretical structure.
The System
An intelligence system is a complex adaptive system: a collection of interacting parts whose aggregate behavior cannot be reduced to any single component. What makes it adaptive is precise: it restructures itself in response to transactional feedback to preserve or increase its capacity to extract value from its environment.
Value, throughout this framework, means any state that increases the system's capacity for future transactions - revenue, knowledge, optionality, structural resilience. The term is deliberately broad because what counts as value is system-dependent. But it is not unbounded: value must be observable through transactions, or it falls outside the framework's scope.
Motivating Question: How does one determine whether a tool embedded in an intelligence system is genuinely increasing the system's adaptive capacity, rather than merely creating the appearance of adaptation?
Core Concepts
Transaction
An exchange across a system boundary that updates the state of both system and environment, producing a measurable information differential. This is the atomic unit. Everything else is built from transactions or describes patterns across them.
Adaptation
The system's response to changes in its environment or internal structure that preserves or restores its capacity to extract value through transactions.
Adaptation is what occurs between transactions — the system encounters a change and restructures. But adaptation alone is insufficient. A system that adapts to each new situation independently — relearning from scratch every time — is adaptive but not intelligent. Each full relearning cycle consumes resources (time, energy, capital) without building on prior structure. If adaptation never compounds, the system's cost of navigating novelty remains constant even as its transaction history grows. That asymmetry is the problem generalizability solves.
Generalizability
The degree to which structure extracted from one set of transactions retains predictive and causal validity across novel transaction contexts — the transferable component of learned coherence.
Generalizability converts adaptation from a recurring cost into a compounding asset. Without it, every new environment demands full relearning, and the system can never accumulate surplus capacity.
Intelligence
The system's capacity to maintain useful internal structure and extract value from an environment under constraints — finite energy, finite time, and incomplete information. Intelligence is not a behavior. It is a capacity. A system demonstrates intelligence not by responding correctly in familiar conditions, but by maintaining coherence when conditions change.
Why generalizability measures intelligence. The hardest constraint any system faces is novelty — conditions it has never encountered, where stored responses have no direct application. A system that performs well only in familiar territory is demonstrating memory, not intelligence. Generalizability — the ability to extract structure that transfers beyond the original context — is what separates systems that have learned the structure of their environment from systems that have merely cached responses to it.
This leaves edge cases. A deeply specialized system might maintain extraordinary coherence within a narrow niche without exhibiting broad transfer. Under this framework, such a system is adapted but not intelligent in the general sense. That is deliberate — the framework's purpose is to distinguish systems that will hold up under novelty from systems that will not.
Thrival
The regime in which a system extracts more from its transactions than bare persistence requires, and reinvests the surplus.
Every system has a persistence cost: the resource expenditure (time, energy, capital) needed to maintain current performance under stable conditions. In a startup context, this is the burn rate required to sustain current revenue and capability without growth or learning.
A system that merely survives adapts just enough to cover persistence costs. A system that thrives generates surplus and reinvests it — into new transaction types, broader model coverage, deeper causal structure. The distinction is not qualitative mood but quantitative trajectory: is the system's capacity flat or compounding?
The boundary between survival and thrival is not sharp. Systems oscillate — a startup thriving in stable conditions may drop to survival mode during a market shock, then recover. The measurement question is not "which regime is the system in right now?" but "what is the surplus trend over a meaningful time window?"
Emergence
Emergence is the phenomenon where complex, often unpredictable patterns, behaviors, or structures arise at a meso-scale (intermediate level) from the spatial-temporal-structural interactions of simpler components in a system, resulting in properties that cannot be fully explained by analyzing the parts in isolation.
In this framework, the relevant emergent property is system-level adaptive capacity. A single founder–tool interaction is a transaction. Whether that transaction contributes to adaptive capacity depends not on the transaction alone but on what surrounds it — what preceded it, what the founder does next, how the system responds to the outcome. Intelligence, as defined here, is not a property of any one interaction. It is a property of the network of interactions unfolding over time.
This has a direct consequence for measurement. Any metric that evaluates a single component in isolation — one tool's effect, one session's engagement, one founder's outcome — captures at most a fragment of the system-level property. Emergent capacity becomes visible only when the interactions between components are measured alongside the components themselves. This is why the structural ratio (introduced under Measurement Theory) requires layered approximation rather than a single indicator: no individual observation point has access to the property being estimated.
Dependency Chain
Adaptation is the process. Generalizability is the measure. Intelligence is the capacity. A system that responds only to familiar conditions is reactive. A system that generalizes across regimes is intelligent.
Transaction → Adaptation → Generalizability → Intelligence → Thrival → Emergence
Measurement Theory
The Core Problem
Consider a tool, program, or intervention embedded in a system. People use it. Some succeed afterward. The question is whether the intervention caused the success, or whether those users would have succeeded regardless. And even if the intervention contributed, whether it contributed in a way that lasts.
These are two distinct questions:
- Did the intervention have an effect? (Causal question)
- Is the effect real capacity or mere appearance? (Transfer question)
Everything in this section exists to answer them. The Bayesian framing that follows does not describe what the system is. It describes how one evaluates whether the system is working, specifically, whether an intervention is genuinely building adaptive capacity. This is measurement theory, not ontology.
One Intervention, One Outcome
A founder enters the system carrying some probability of success based on background, knowledge, and market conditions. This is the baseline. An intervention is delivered: a tool, coaching, a network connection. The founder's probability of success changes. The difference between probability with the intervention and probability without it is the effect:
Effect = P(success | intervention applied) − P(success | no intervention)
Both sides of this equation can never be observed directly and simultaneously — the founder either received the intervention or did not (similar to Schrodinger’s cat). But this difference is the target. Every measurement approach in what follows is an attempt to approximate it.
Structural vs. Cosmetic Effects
When an impact is observed, it is critical to understand its nature. For our purposes, we focus on two types of effect:
Cosmetic effect. The intervention makes the founder look stronger to evaluators. The pitch deck is polished, the answers are rehearsed, the vocabulary is right. The founder raises money, and the metric registers success. Then the market shifts, and the founder has no basis for response, because the tool provided outputs, not understanding.
Structural effect. The intervention changes how the founder reasons. The founder understands why certain framings work, how to stress-test assumptions, when a model is breaking down. The founder raises money as well, but when the market shifts, adaptation follows, because what was built was a mental model, not a document.
Both effects look identical in familiar conditions. They diverge under novelty.
Structural Ratio = Effect in unfamiliar conditionsEffect in familiar conditions
Ratio near 1: The effect transfers. Real capacity was built.
Ratio near 0: The effect disappears outside the original context. Nothing was built.
This is the single most important diagnostic in the framework. Everything else supports estimating it.
Multiple (Synergistic) Interventions
Real systems do not deliver just one intervention. A founder might receive a planning tool, training on how to use it, and introductions to investors: three interventions, layered.
The naive approach, measuring each effect independently and summing them, is almost always wrong. Interventions interact.
Training may change how the founder uses the planning tool. Instead of copying its output, the founder iterates, challenges assumptions, rebuilds. Training alone did not produce that outcome. The tool alone did not produce it. The combination did.
Each interaction has its own structural ratio. The training-plus-tool combination might produce an effect that transfers well to new contexts, even if neither component alone transfers at all. Or the reverse: each component transfers independently, but their interaction is brittle and context-specific.
This diagnostic reveals whether the system is genuinely functioning as a system or is simply a bundle of unrelated parts.
How the System Learns Over Time
The structural ratio captures whether an intervention is building real capacity at a point in time. But the system delivering those interventions also changes — it accumulates outcome data, encounters new founder types, discovers broken assumptions. How it responds determines whether it improves or merely persists.
Pattern 1 — Calibration. The system observes outcomes and adjusts within its existing frame. It learns which interventions work better for which profiles, refines matching, and tunes recommendations.
Example. An accelerator notices that technical founders benefit more from business model coaching than product coaching. It shifts allocation. The next cohort performs slightly better. The system became more precise about a relationship it could already see.
Calibration is necessary but limited. It can only improve at things it already knows to measure.
Pattern 2 — Expansion. Sometimes outcomes do not say "the estimates were slightly off." They say "the wrong things are being measured entirely."
Example. The same accelerator begins accepting hardware startups. Its model, built on software economics, predicts that shorter development cycles lead to better outcomes. For hardware founders, this is meaningless. The errors are not noise around a correct model. They point to missing variables: supply chain complexity, manufacturing lead times, capital intensity curves that do not exist in software.
The correct response is not to tune harder. It is to change what the system can represent: new variables, new categories, new intervention types.
The difference in measurement terms: a steady structural ratio across new contexts means calibration is working. A rising structural ratio means expansion is working, the system is learning to see more, not just see better. Most systems do the first well and the second poorly. That is the difference between accumulating experience and accumulating intelligence.
Three Layers of Approximation
The structural ratio cannot be measured directly. That would require observing the same founder with and without the intervention, in both familiar and unfamiliar conditions. Since only one of these realities is ever available, we approximate through three progressively stronger layers of evidence:
Layer 1 — Is anything happening at all? Behavioral signals from inside the tool. The cheapest and weakest evidence. This layer examines how users interact with the tool: do they extract outputs or reshape them? Do they iterate, stress-test, restructure? This cannot confirm that capacity was built, but it reliably indicates when it was not. A necessary filter, not a sufficient one.
Layer 2 — Did anything transfer? Users are given a problem the tool does not solve, in an adjacent domain, same reasoning structure, different surface content, no tool access. Performance above baseline suggests something transferred. The challenge is calibrating distance: too similar measures memory, too different measures something unrelated.
Layer 3 — Did anything last? Cohorts are followed over months or years, through conditions neither users nor system designers anticipated. This is the closest approximation to the true structural ratio, because the novelty is genuine rather than designed. It is the most expensive, the slowest, and the only layer that measures what actually matters.
What's Next
This completes the theoretical layer. We now have a vocabulary, transaction, adaptation, generalizability, intelligence, thrival, emergence, and a measurement apparatus: the structural ratio, its three-layer approximation, and the calibration/expansion distinction, sufficient to describe what an intelligence system is, to metricize its key processes and parameters at both the micro-scale and the meso-scale, and to analyze and better understand its inner workings.
In the next post, we move from Intelligence Theory to Intelligence Engineering, the applied layer, where we explore how to design, implement, and optimize these systems in a real-world scenario.