The first principles of modern learning systems

Modern learning systems should not be designed around courses.

They should be designed around capability.

A course is a delivery format.

A video is a delivery format.

A document is a knowledge source.

A quiz is one possible source of evidence.

None of these is the final goal.

The real system begins with something a person should become better able to know, decide, explain, or perform—and ends with evidence about whether that capability exists when it matters.

The seven principles below are a LoreGraph synthesis of the Week 1 Modern Learning Systems research. They are not a formally validated academic model. They are a practical architecture for thinking about workplace learning in an era of AI, abundant content, and rapidly changing organizational knowledge.

The model in one line

A modern learning system should connect:

Knowledge → capability → practice → retention → transfer → performance → evidence → improvement

The arrows are not guarantees.

A document does not automatically create capability.

Practice does not guarantee retention.

Learning does not guarantee workplace behavior.

And better workplace results do not prove that training caused the improvement.

Each connection must be designed and, when important, tested.

Principle 1: Capability comes before content

The first question should not be:

What should this course contain?

It should be:

What should someone be able to do differently?

This changes the design process immediately.

Suppose an onboarding handbook contains 80 pages.

A content-first process asks AI to summarize the 80 pages and divide them into lessons.

A capability-first process asks:

• What does a new employee need to perform during the first week?
• Which decisions carry risk?
• Which concepts are prerequisites?
• Which information needs to be remembered?
• Which details only need to be findable?

The resulting training may contain less information.

That can be a strength.

The objective is not maximum content coverage.

It is sufficient knowledge and practice to support a useful capability.

Principle 2: Design backward from evidence

If you cannot describe what would count as evidence, the learning objective is probably still vague.

Compare:

Understand the refund policy.

with:

Given a customer case, select the correct refund or escalation path and explain which policy condition applies.

The second objective suggests its own assessment.

Give the learner a case.

Ask for a decision.

Ask for the reasoning.

Then vary the case later to test whether the rule transfers.

Evaluation should not be bolted onto the end of course production. CDC guidance similarly recommends defining evaluation purpose, questions, and collection methods early in training development.

Source: CDC evaluation-planning guidance. A modern learning system should therefore connect:

objective → learning activity → evidence

rather than:

content → quiz because every course needs a quiz

Principle 3: Learners must actively reconstruct knowledge

Exposure is necessary for many kinds of learning.

But passive exposure alone is a weak foundation for durable capability.

Learners need opportunities to retrieve, explain, compare, decide, predict, practice, and correct.

Retrieval practice is especially useful because retrieving information can itself strengthen later retention. Research has repeatedly distinguished retrieval from simply studying the material again.

Source: Roediger and Karpicke, “Test-Enhanced Learning”. This principle changes how we think about interaction.

Interaction is not valuable because someone clicked.

A next button is interaction.

That does not make it learning.

The useful question is:

What mental work did this activity require?

A scenario can require judgment.

A prediction can expose a mental model.

An explanation can reveal missing understanding.

A worked task can reveal procedural skill.

A strong learning system creates interactions because they expose and strengthen thinking—not because engagement metrics look better.

Principle 4: Design for what remains later

Immediate performance is easy to overvalue.

A learner may answer correctly because:

• the relevant sentence is still visible;
• the same example appeared moments earlier;
• a hint narrows the answer;
• feedback arrives before meaningful effort;
• the learner repeatedly sees the same item.

Research on learning versus performance shows why success during practice can be a poor proxy for durable learning.

Source: Soderstrom and Bjork, “Learning Versus Performance”. A modern system should therefore care about time.

For important knowledge, ask:

• Can the learner retrieve it tomorrow?
• Next week?
• At the interval where the capability will actually be needed?

Not every topic needs a delayed assessment.

But when the business claim is that employees will remember something later, measurement should eventually include later.

Retention is not an optional academic concept.

It is part of the product requirement.

Principle 5: Design for change in context

Remembering an answer is useful.

Using the underlying idea when the situation changes is more useful.

That is transfer.

The National Academies defines transfer around extending learning from one context to new contexts and emphasizes understanding underlying principles as important for flexible use.

Source: National Academies, “Learning and Transfer”. For workplace learning, this means practice should not consist entirely of cloned examples.

Change:

• the customer;
• the sequence;
• the exception;
• the symptoms;
• the interface;
• the wording;
• the level of ambiguity.

Keep the underlying concept stable.

If an employee can only succeed when the work resembles the exact training example, the system has taught a pattern more than a capability.

A modern learning system should therefore model not only:

What does this person know?

but also:

Under what conditions can this person use it?

Principle 6: Memory and performance support belong in one architecture

A mature learning system does not try to force every piece of organizational knowledge into human memory.

Workplaces already rely on external knowledge:

• SOPs;
• checklists;
• diagrams;
• calculators;
• search;
• reference manuals;
• software prompts;
• colleagues;
• AI assistants.

The design challenge is deciding what belongs where.

Human memory is valuable for:

• recognizing a situation;
• understanding principles;
• making judgments;
• performing frequent actions;
• knowing when to stop or escalate;
• connecting new information to prior experience.

External systems are often better for:

• changing details;
• rare exceptions;
• precise reference data;
• long procedural sequences;
• large information collections.

This distinction prevents training from becoming a memory contest.

It also creates a better role for AI.

AI can reduce search, organize reference material, provide contextual support, and retrieve detailed information while the learner retains responsibility for the capabilities that matter.

Principle 7: Evidence must be named honestly

Modern systems collect enormous amounts of data.

That does not mean they understand learning.

A platform may know:

• who enrolled;
• what pages were opened;
• how long a browser session lasted;
• whether a course was completed;
• which quiz items were correct.

Each signal is useful.

Each signal also has a boundary.

Week 1's Learning Evidence Ladder separates participation, reaction, immediate performance, delayed retention, transfer, workplace application, and organizational results because they answer different questions.

A system should say:

completed

when it measured completion.

It should say:

immediate assessment score

when it measured an immediate assessment.

It should reserve stronger terms such as:

retention, transfer, workplace application, mastery

for evidence models that actually justify those interpretations.

The rule is:

Make the claim no stronger than the evidence.

That principle becomes more important as AI makes it easy to generate impressive dashboards and automated conclusions.

AI changes the implementation, not the first principles

Generative AI changes what is economically possible.

It can:

• extract structure from documents;
• generate drafts;
• explain concepts in different ways;
• create scenarios;
• produce practice variations;
• analyze responses;
• suggest feedback;
• retrieve relevant knowledge;
• personalize support.

That is powerful.

But none of it changes the underlying learning problem.

AI can also do too much.

A field experiment in high school mathematics found that unrestricted AI assistance improved students' performance while the tool was available but could reduce later unaided performance; a more safeguarded tutoring design mitigated much of that negative learning effect.

Source: Bastani et al., “Generative AI Without Guardrails Can Harm Learning”. Other experiments show that AI used to improve instructional behavior can support better learning outcomes. Tutor CoPilot, for example, helped tutors use more pedagogically useful practices and improved the study's student topic-mastery outcome.

Source: Tutor CoPilot study. Those findings come from specific educational settings, not workplace training.

But together they reinforce a first-principles view:

AI does not remove the need for learning design. It makes learning design more consequential.

The system must decide which difficulty to remove and which thinking to preserve.

A worked example: turning an SOP into a modern learning system

Imagine an organization has a 25-page incident-reporting SOP.

A traditional document workflow looks like this:

Publish SOP → send link → collect acknowledgment

A traditional e-learning workflow might look like:

SOP → slides → narration → quiz → completion report

A first-principles workflow looks different.

1. Define capability

What should an employee recognize, decide, or perform?

For example:

Recognize a reportable incident, take the correct immediate action, and begin the approved reporting process.

2. Identify evidence

What would show that capability?

• immediate scenario decision;
• independent reporting simulation;
• delayed new scenario;
• workplace quality review where appropriate.

3. Structure the source

Separate:

• essential principles;
• triggers;
• immediate actions;
• procedure;
• exceptions;
• reference information.

4. Teach selectively

Explain only what supports the capability.

Keep detailed rare information accessible externally.

5. Create active practice

Give the employee realistic cases.

Ask:

Is this reportable?

What is the first action?

Why?

6. Introduce time and variation

Return later with a different case.

Do not simply repeat the original quiz.

7. Connect to work

Ensure the real reporting tools, manager expectations, permissions, and job aids support the trained behavior.

8. Collect evidence and improve

If employees consistently miss one decision point, investigate why.

The answer may be:

• the explanation is weak;
• the examples are misleading;
• the SOP itself is unclear;
• the tool is difficult;
• employees lack authority;
• the work environment conflicts with the training.

That is a learning system.

The course is only one component.

Common failure modes

A modern interface does not automatically create a modern learning system.

Watch for these failures:

Automating content dumping

AI turns a 100-page document into 100 polished lesson screens.

The format changed.

The learning problem did not.

Personalizing without a learning model

Changing tone, length, or examples is personalization, but it does not necessarily adapt to the learner's knowledge or evidence.

Measuring clicks more precisely

Better analytics on weak proxies are still weak proxies.

Removing every difficulty

If AI supplies every answer, practice may become easier while capability remains dependent on the tool.

Ignoring the workplace

Training cannot compensate for a broken process, missing permission, poor management support, or unavailable equipment.

Treating AI output as authoritative

High-risk learning still requires source integrity, appropriate review, and clear ownership.

How LoreGraph fits the model

LoreGraph currently focuses on one important part of this architecture: turning existing organizational documents into structured, editable training with lessons, quizzes, human review, assignment, and learner tracking.

Source: LoreGraph. The first-principles model gives that workflow a larger direction.

Instead of optimizing only:

How quickly can a document become a course?

we can also ask:

What capability does the source support?

What should be practiced?

What should remain externally available?

What evidence should the learner produce?

What should happen after the first assessment?

What changes when the source changes?

That is the difference between an AI authoring tool and a modern learning system.

Next step

Before creating your next piece of workplace training, answer seven questions:

1. What capability should change?
2. What evidence would demonstrate that change?
3. What must the learner actively practice?
4. What should still be available after time passes?
5. What changed situation should the learner handle?
6. What belongs in memory and what belongs in external support?
7. What can the available data honestly tell us?

Only then ask:

What content should we create?

That reversal is the first principle behind all the others.

Sources and further reading

• National Academies, How People Learn II: Learners, Contexts, and Cultures.
• National Research Council, “Learning and Transfer”.
• Soderstrom and Bjork, “Learning Versus Performance: An Integrative Review”.
• Roediger and Karpicke, “Test-Enhanced Learning: Taking Memory Tests Improves Long-Term Retention”.
• CDC, “Evaluate Training: Building an Evaluation Plan”.
• Bastani and colleagues, “Generative AI Without Guardrails Can Harm Learning”.
• Wang and colleagues, “Tutor CoPilot: A Human-AI Approach for Scaling Real-Time Expertise”.