Cognitive Factory

Overview

The factory's ability to act depends on the world it can make sense of. A team may execute an impeccable workflow while overlooking the reason the work exists.

The Knowledge Factory develops triggers, agent DAGs, and the engineering of feedback loops. This closing essay examines the capabilities those mechanisms serve: how far the factory can see, what it can understand, and which experiences can change its next decision.

The AI Factory Series

1. Vision and Values
2. Understanding and Bottlenecks
3. Truth and Inference
4. The Knowledge Factory
5. Ontology Factory
6. Cognitive Factory — you are here

1. How Far Can the Factory Make Sense?

Michael Levin's cognitive light cone concerns the spatial and temporal reach of the goals and events a system can measure, model, and try to affect. His framework examines diverse forms of agency; applying it to an organization is the adaptation proposed here. It supplies no validated intelligence score for an AI factory. Levin, 2019.

Consider a particular customer-onboarding workflow. Can it connect a failed request to a customer's interrupted project? Can it retrieve the decision made three months ago, distinguish it from today's priority, and anticipate a tradeoff that will appear next month? Can the responsible team intervene and later determine whether its explanation held?

These questions describe a capability profile. An agent that can edit many files may have a narrow understanding of the consequences. A system that retrieves a million documents may still miss the one decision that governs the present case.

Use the profile to measure a configured factory against named tasks. Keep observable reach, understanding, and authorized influence separate. More access does not itself establish competence or grant permission to act.

2. Choose Signals That Explain Progress

A factory needs a reason to attend to a signal. FranklinCovey's The 4 Disciplines of Execution distinguishes lag measures of results from lead measures chosen for their expected relationship to those results and the team's ability to influence them. That gives us a practical question: which controllable behavior do we expect to improve the outcome? FranklinCovey.

For the onboarding case, use an explicitly hypothetical measurement plan:

The proposed lead measure must earn its place through evidence. Counting trials can reward superficial activity; an early-arriving metric is not automatically a useful lead measure. Keep the definition, cohort, baseline, observation window, owner, and expected relationship visible. If the practice improves while the customer outcome does not, revisit the relationship.

A scoreboard becomes informative when it connects effort with consequence. A threshold used to start work belongs to the factory's execution policy. Cognition asks what the observation means, which competing explanations remain, and whether the goal or measure needs revision.

3. A Second Brain Holds History and Current State

Tiago Forte's Building a Second Brain offers a practice for keeping knowledge outside immediate memory so it can support later work. His PARA method organizes material around projects, areas of responsibility, resources, and archives. Here, that is a starting point for organizational memory, extended with ownership, record state, provenance, and relationships. Forte's Second Brain and PARA.

The factory needs both a history of what happened and a view of what currently governs the work:

A past priority is evidence about an earlier decision; it is not automatically a current instruction. An experiment that failed in one customer segment does not settle a different segment's case. A superseded ADR can explain the code without authorizing a new implementation.

State therefore travels with the record. Preserve identities and revision history, connect replacements to what they supersede, and give someone responsibility for freshness. Conflicting records should expose a question for the owner rather than quietly collapse into a confident summary.

4. Make Context Discoverable by Convention

This memory can remain in the systems that maintain it: a project tracker, research store, decision log, or document workspace. Copying their entire contents into a source repository creates another body of material to keep current.

The repository should keep the implementation, contracts, and decisions that must be reviewed with the code. An ADR governing a code boundary may belong beside that code. Changing roadmaps, raw research, and historical project material can remain with their owners elsewhere.

Give people and agents a small, predictable entry point. A repository's existing agent instructions or a linked context index can identify:

• the project and domain identifiers used to look up relevant records;
• the authoritative locations for plans, experiments, ADRs, and priorities;
• the owner, access method, and fields used to judge a record's state; and
• when each source should be consulted.

This is a proposed convention, not a new documentation repository. The entry point routes discovery; it does not replicate the organization's memory.

At the destination, each record needs a stable identity, type, scope, owner, status, effective or updated date, and links to evidence and related records. Summaries help choose what to open. Source links let the reader recover the reasoning and verify that the summary still applies.

5. Retrieve the Context the Decision Needs

Forte's progressive summarization preserves access to underlying detail while making the useful parts easier to find. We can adapt that principle into a bounded retrieval practice for people and agents. Progressive Summarization.

Start with the question and its scope, then navigate from a small index to selected records and their evidence:

1. Identify the affected project, domain, decision, and time period.
2. Inspect record summaries and state to choose relevant sources.
3. Open the few records that can support or contradict the proposed explanation.
4. Follow their evidence or supersession links when a consequential detail is missing.
5. Assemble the relevant excerpts with provenance, dates, uncertainty, and a retrieval budget.

For the activation drop, the first packet might contain the current onboarding priority, the release's experiment, one governing ADR, and the measurement definition. An earlier experiment enters only if its conditions bear on a live hypothesis. The team can expand its inquiry when those records reveal a gap.

A retrieval limit needs a visible stopping condition. If missing information could change the decision, mark the gap and seek it or escalate; do not treat the budget as evidence that the search was complete. If access is unavailable, preserve that uncertainty. Retrieved material informs the work within existing authority; an archived instruction does not silently acquire new authority.

6. Experience Must Change the Next Judgment

Return to the two teams. Finding the experiment may reveal that the activation drop was an anticipated tradeoff. Finding its outcome may show that the tradeoff failed. Finding a current priority may explain why the same tradeoff is now unacceptable. The useful capability is relating those records to the present decision.

Retain the comparison between expectation and result, the explanation it changed, and who accepted the revision. Promote durable findings into the appropriate definition, test, decision, or operating practice. Keep the fuller experimental history linked and discoverable.

That division gives the factory a working memory for today's task and a larger organizational memory it can consult. Neither requires every fact to be permanently present in the model's context.

7. Evaluate the Factory's Cognitive Reach

Use representative cases with known relevant evidence to test whether the configured system can:

• find the current priority and distinguish it from a superseded one;
• connect a present symptom to an applicable earlier experiment or decision;
• preserve a credible alternative explanation and identify missing evidence;
• trace the claims in a recommendation to their sources; and
• revisit the recommendation when a later outcome contradicts it.

Record retrieval misses, stale-state mistakes, unsupported claims, and review effort alongside task success. Compare the same tasks under different context and memory arrangements. These are proposed evaluations of this factory's sensemaking, not a universal cognition benchmark.

The temporal reach matters: how far back can the system recover applicable experience, and how far forward can it formulate consequences that someone will actually evaluate? Its reach across teams matters too: can one group's learning inform another without erasing the difference between their contexts? A larger capability profile is useful when it improves those judgments.

The series ends with an organization that can choose direction, test claims, distribute understanding, execute work, maintain shared meaning, and make sense of experience. Its second brain gives the next decision a history without requiring every decision to carry the whole archive.

Sources

• Michael Levin, “The Computational Boundary of a ‘Self’: Developmental Bioelectricity Drives Multicellularity and Scale-Free Cognition” (2019). The organizational capability profile is this article's adaptation.
• FranklinCovey, “5 Ways to Narrow Your Sales Focus in Times of Uncertainty”. Lead and lag measures; the onboarding measures are illustrative.
• Tiago Forte, Building a Second Brain.
• Tiago Forte, “The PARA Method”. The record conventions proposed here extend a personal knowledge-management practice.
• Tiago Forte, “Progressive Summarization: A Practical Technique for Designing Discoverable Notes”. Selective agent retrieval is an adaptation, not an evaluated result of this source.
• W3C, “PROV-O: The PROV Ontology” (2013). A reference vocabulary for provenance; it does not prescribe this storage design.