Cognitive Factory

Overview

Every company is building a factory, and The Knowledge Factory described its operating system: an organization that turns evidence and intent into reusable capability. This essay examines how that factory thinks.

Cognition is not a model subscription or a collection of agents. It is how the organization represents context, checks it against action, and learns from outcomes. Concretely, the factory's cognition has four parts:

• Graph context exploration asks relational questions — which evidence motivated this capability, which decisions depend on this assumption — instead of treating context as documents containing matching words.
• Executable context turns institutional knowledge into contracts: a definition becomes a schema, a customer promise becomes an evaluation, an observed failure becomes a regression case.
• The compounding loop returns every outcome to the system as evidence that updates context and evaluation, so capability grows through use.
• A diagnostic build order says what to build first: expose evidence, name distinctions, build evaluation before scaling generation, instrument outcomes, and give teams authority inside the new boundaries.

The diagnostic for how far this cognition reaches is the cognitive light cone — how much of the relevant domain the system can observe, interpret, affect, and learn from. This essay is where the organization implements the machinery that advances it down the light cone.

The factory also needs two human-governed disciplines to keep this cognition purposeful: ontology maps the domain, and strategy chooses direction. Both are developed in their own essays, and this essay closes by connecting them.

Relationship to the Series

This is the sixth essay in the sequence. The Knowledge Factory introduced the operating system: an organization that turns evidence and intent into reusable capability. The Ontology Factory mapped that system's semantic infrastructure. This essay examines how the factory thinks — the cognition that makes the map useful, the learning compound, and the light cone that measures how far that cognition reaches. The strategy discipline that chooses where the factory should act is covered inside The Knowledge Factory.

Extending Loop and Graph Engineering

This essay extends two established engineering ideas. Loop engineering builds feedback systems in which outcomes return as evidence that updates context and evaluation; graph engineering builds traversable relationships among people, concepts, systems, evidence, decisions, and outcomes. The cognitive factory is both of those — and it extends both with ontology and cognition:

• Ontology makes a loop or a graph checkable rather than plausible: stable terms, boundaries, invariants, and evidence rules that let the machinery be validated instead of admired. Without ontology, the graph degenerates into named edges and the loop into dashboards.
• Cognition is what the machinery is for. The cognitive light cone measures how much of the relevant domain the system can observe, interpret, affect, and learn from, and every loop and graph in this essay exists to advance it.

The sections that follow are the implementation of that extension: graph context and executable context (the graph made semantic), the compounding loop (the loop made systemic), and the scorecard and build order that decide how far the cognition reaches.

1. Graph Context Exploration

Graph context is a signature concept, not a generic knowledge-graph pitch. Most organizational search treats context as documents containing matching words. Graph context exploration asks relational questions:

• Which customer evidence motivated this capability?
• Which definition of conversation applies in this service?
• What decisions depend on this assumption?
• Which failures caused this evaluation to exist?
• Which teams, systems, and metrics will a change affect?
• Where does the current model conflict with observed behavior?

The figure above is a live proposition graph: pick a claim in the explorer to follow its relationships. You can open the same graph in the right-side tool drawer.

Explore the graph →

The graph may be implemented through links, metadata, schemas, code dependencies, event lineage, or a graph database. The product requirement is traversable relationships with provenance — not a particular storage engine.

2. From Documents to Executable Context

Documents remain important, but the factory needs context that can guide and check action:

This is how institutional knowledge becomes productive capital rather than a larger pile of prose.

3. The Compounding Loop

The factory's return comes from a loop:

The loop compounds only when the organization captures corrections. More AI output without retained learning is throughput, not a knowledge factory.

4. The Cognitive Light Cone Scorecard

The cognitive factory is where an organization implements the machinery that advances it down the light cone. This essay makes an organizational adaptation of Michael Levin's cognitive light cone: it uses the metaphor as a diagnostic for how much of the relevant domain a system can observe, interpret, affect, and learn from. Three systems in increasing reach:

• LLM: works from supplied context without its own harness. Humans select the evidence, state the goal, and evaluate the response.
• Agent: combines an LLM with tools, memory, and bounded workflows. Humans establish its objective, permissions, evaluation, and escalation boundaries.
• Knowledge factory: connects agents to organizational data, context stores, operational signals, evaluations, and feedback loops. Humans systematize the inputs, govern how evidence is interpreted, and remain accountable for the values and decisions propagated through the system.

Expanding a system's cognitive light cone increases what it can coordinate; it does not by itself authorize the governing values it applies.

5. What Companies Should Build First

Not everything at once. A diagnostic order:

1. Identify the decisions or workflows with repeated context loss and review burden.
2. Expose the customer and operational evidence behind them.
3. Name the domain distinctions and invariants required for safe delegation.
4. Build evaluation before scaling generation.
5. Instrument outcomes and connect them back to decisions.
6. Give teams authority inside the new boundaries.
7. Measure whether capability, learning speed, and customer outcomes improve — not only whether token or labor costs fall.

6. The Two Factory Disciplines

The knowledge factory needs two human-governed disciplines. Ontology Factory asks how humans map the domain so models and teams share the right entities, relationships, constraints, and evidence. The strategy discipline — covered in The Knowledge Factory — asks how humans choose direction through narrative, empathy, opportunism, memory, and systematic feedback.

Ontology makes the factory coherent. Strategy makes it purposeful.

Sources

• Michael Levin, “Technological Approach to Mind Everywhere: An Experimentally-Grounded Framework for Understanding Diverse Bodies and Minds” (2022). Introduces the cognitive-light-cone framework adapted as an organizational diagnostic in this essay.
• Patrick Lewis and colleagues, “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks” (2020). Establishes retrieval from explicit non-parametric memory and identifies provenance and knowledge updating as design problems.
• Shunyu Yao and colleagues, “ReAct: Synergizing Reasoning and Acting in Language Models” (2023). Demonstrates an agent pattern that interleaves model reasoning, tool actions, environmental observations, and plan updates.
• Noah Shinn and colleagues, “Reflexion: Language Agents with Verbal Reinforcement Learning” (2023). Tests the use of retained task feedback in episodic text memory to improve subsequent agent trials.
• W3C, “PROV-O: The PROV Ontology” (2013). Defines a standard model for representing provenance among entities, activities, and agents.
• National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (2024). Supports lifecycle governance, context mapping, measurement, documentation, and human accountability for generative-AI systems.
• DORA, Google, 2025 State of AI-assisted Software Development Report. Supports the claim that effective AI adoption depends on the surrounding organizational system and capabilities rather than model access alone.