How People Learn: The Principles Behind the System

Every part of the system is pedagogy-driven: it is built on what research says about how people learn, and the same principles hold for a 13-year-old learning English and an adult learning to sell or to work safely.

This page is the learning-science half of the evidence base. The K12 essays cover the system around learning (what gets measured, who checks the educator’s judgement, where tools fit). This page covers how learning works, and it applies to schools, academies and corporate learning and development (L&D) alike. The Corporate essays take the same principles into training at work.

Each principle below has three parts: what it is, what the research says, and where it lives in the system, with a K12 and a corporate use case. A principle only appears on a part that actually uses it, and every claim carries a source.

1. Retrieval Practice

Pulling knowledge out of memory strengthens it more than reading it again.

Research. Learners who tested themselves remembered more a week later than learners who restudied the same text (Roediger and Karpicke, 2006). A meta-analysis of practice testing found a reliable benefit across ages and settings, including over restudying (Adesope and colleagues, 2017), and a major review rated it one of the two highest-utility techniques (Dunlosky and colleagues, 2013).

In the system. Quiz block library, assignment builder, adaptive practice engine, correction loop.

K12 A learner recalls the third conditional in a gap fill a week after the lesson, instead of rereading their notes.

Corporate A sales rep answers two objection-handling questions on their phone each week, instead of rewatching the module.

Read more. One Brain, Two Settings

2. Spaced Practice

The same practice spread over days and weeks is remembered far longer than practice crammed into one session.

Research. A meta-analysis of hundreds of comparisons found spaced practice beat massed practice, with the best gap growing as the test date moves further away (Cepeda and colleagues, 2006). The effect holds for working professionals: urology residents who got spaced questions by email retained more than those who did not (Kerfoot and colleagues, 2007, US and Canada).

In the system. Assignment builder (schedules review into later assignments), adaptive practice engine (brings an item back just before it is likely to be forgotten). The correction loop could feed it, so errors return later; that link is planned, not built.

K12 An error from September's essay comes back as a short exercise in October and again before the exam.

Corporate Health and safety refreshers arrive as short spaced checks across the year, replacing the single annual session.

3. Interleaving Similar Items

Mixing things that are easy to confuse teaches learners to tell them apart. Brand-new items are still best learned one block at a time.

Research. A meta-analysis found interleaving helps most when the items look alike, and helps little or not at all for learning single words (Brunmair and Richter, 2019). Learners tend to believe blocked practice works better even when interleaving has helped them more (Kornell and Bjork, 2008), so the system chooses the mix rather than leaving it to the learner.

In the system. Assignment builder, adaptive practice engine (uses the error tags to mix items that are commonly confused).

K12 Past simple and present perfect questions appear mixed in one set, so the learner has to decide which form fits.

Corporate A set of near-miss scenarios mixes a slip hazard, a trip hazard and a manual-handling risk, so the worker learns to tell them apart on site.

4. Scaffolding That Fades

Novices need structure and worked examples. As expertise grows, the same support gets in the way and should be withdrawn.

Research. Instruction that helps beginners can lower the performance of more experienced learners, which is known as the expertise reversal effect (Kalyuga and colleagues, 2003). Fading worked-example steps one at a time moves learners smoothly from studying to solving (Renkl and Atkinson, 2003). This matters most in corporate training, where learners arrive with very different levels of experience.

In the system. Adaptive practice engine (support level set by performance, not by age or job title), help agent (guiding questions get fewer as the learner improves), assignment builder (partly worked examples for beginners).

K12 A weaker writer gets a paragraph frame for an FCE essay; a stronger one gets only the task.

Corporate A new hire sees a fully worked discovery call; an experienced rep goes straight to a live-style scenario with no model answer.

5. Feedback Built on a Map of Common Errors

Mistakes in any subject follow patterns. Feedback that names the pattern, and sets the right idea against the wrong one, works better than a right-or-wrong mark.

Research. Wrong intuitions are not wiped out by teaching. University students were slower and less accurate on statements where intuition and science disagree (Shtulman and Valcarcel, 2012), and science professors showed the same effect (Shtulman and Harrington, 2016). The old idea is held back, not replaced. So feedback has to contrast the two explicitly. Maths already has a catalogued misconception map, and tools built on such maps have trial evidence behind them (ASSISTments, 43 schools in Maine, 0.18 SD; see the K12 essay The Tool Was Never the Hard Part).

In the system. Quiz block library (every wrong answer is tagged to the error it points to), AI grading agent (names a grammar-map point for each error), correction loop, adaptive practice engine.

K12 The system's 2,243-item grammar bank is the error map for English; "the people is" is tagged as a plural-noun agreement error, not just "grammar".

Corporate A sales error map lists the ways reps go wrong (pitching features before finding the need, discounting too early, skipping the objection). Health and safety has its own (assuming someone else has isolated the power).

6. Learner Reasoning Before System Explanation

Learners who attempt, judge or explain their own work before they are told the answer learn more from the answer when it comes.

Research. Generating an answer is remembered better than reading it (Slamecka and Graf, 1978). Trying a problem before instruction improved conceptual understanding in a meta-analysis of productive failure (Sinha and Kapur, 2021), though only when good instruction follows soon after. Self-assessment against clear criteria supports self-regulated learning (Panadero and colleagues, 2017) and sits at the heart of formative assessment (Black and Wiliam, 1998).

In the system. Pre-marking check (the learner highlights their own work and gives reasons before the AI marks it), adaptive practice engine (confidence prediction before the answer is shown), help agent (attempt before hint).

K12 Before submitting an essay, the learner finds each content point in their own text and flags the sentence they are least sure of.

Corporate A rep scores their own call recording against the sales rubric before the AI and the manager give theirs. The gap between the two scores is itself an actionable insight.

7. Task-Focused Feedback, With a Human in the Loop

Feedback works when it is about the work and what to do next. Feedback about the person (praise, blame, ranking) often does nothing, or harm.

Research. In a large review, over a third of feedback interventions made performance worse, most often when the feedback drew attention to the self rather than the task (Kluger and DeNisi, 1996). Feedback about the task and the process of doing it is the most effective kind (Hattie and Timperley, 2007), and feedback overall has a medium effect on learning (Wisniewski and colleagues, 2020).

In the system. AI grading agent (comments on the text, never the learner), correction loop, educator review layer (an educator approves every piece of AI feedback before the learner sees it, which also meets the EU AI Act’s human-oversight rules).

K12 "Your second paragraph doesn't answer the question about cost", never "You're a weak writer".

Corporate "You offered the discount before the client raised price", never a league table of the team's close rates.

8. Help Needed as a Measure of Learning

How much help a learner needs to get something right says as much about their progress as whether they get it right.

Research. Dynamic assessment, which grew out of Vygotsky’s zone of proximal development, measures learning by the support a learner needs. It is established in second-language teaching (Poehner, 2008), and a review found it predicts later achievement (Caffrey and colleagues, 2008). Using it inside an AI help agent to track progress is new ground and still to be tested.

In the system. Help agent (the twin-question model: it never works on the set task, gives the answer only to a similar question, then sends the learner back), tracking layer (records the help level and compares it with the learner’s own past).

K12 In October a learner needs four guiding questions to fix a tense error; by December they need one.

Corporate A new manager needs three prompts to run a return-to-work conversation in a scenario; a month later they need none.

9. Learning Analytics on What Learners Do, Compared With Their Own Past

The most useful data is direct behaviour (attempts, corrections, revisions, time on task), read against each learner’s own baseline rather than the class or team.

Research. Simple behaviour records predict who is struggling: in Chicago, ninth graders who were on track on credits and course failures were 3.5 times more likely to graduate in four years (Allensworth and Easton, 2005, US). Evidence suggests self-reports are skewed by the group a learner compares themselves with, known as reference bias (West and colleagues, 2016, US), and comparing a learner with their own past avoids it. No trial yet shows that such records improve outcomes or reduce educator bias. They give the educator an outside check, and whether that check works is still to be tested.

In the system. Tracking layer, educator dashboard and academy dashboard (flag, never gate: a flag prompts a conversation and never blocks a learner).

K12 The dashboard flags a learner whose revisions have dropped to half their usual rate, without comparing them to the class.

Corporate A learning and development manager sees which competencies a team keeps needing help with, as an actionable insight for the next programme. Individual records are never fed into performance reviews.

10. Motivation, Relevance and Transfer

People keep learning when they have real choices, see themselves getting better and can see why it matters. At work, learning only counts if it shows up on the job.

Research. Self-determination theory links motivation to autonomy, competence and relatedness (Ryan and Deci, 2000), and leaders who support autonomy get more motivated staff (Slemp and colleagues, 2018). Whether training transfers to the job depends heavily on the workplace: support from managers and peers and a chance to use the skill soon (Blume and colleagues, 2010). Satisfaction scores (“did you enjoy it?”) barely predict learning or transfer (Alliger and colleagues, 1997), yet they are what most learning and development teams report.

In the system. Adaptive practice engine (real choices at set points), educator dashboard (progress shown against the learner’s own past, which builds a sense of competence). For corporate use, a transfer step is new: practice set on the learner’s real tasks, and a manager check-in on whether the skill was used.

K12 A learner chooses which of two essay prompts to write on, and sees their own error rate fall over the term.

Corporate After a module on discovery questions, the rep's next three real calls are reviewed for those questions, so the measure is behaviour on the job rather than a smile sheet.

What the System Leaves Out

Some popular ideas have weak evidence, so no part of the system is built on them. Two essays go into these: Good Design, Not Learning Styles for schools and One Brain, Two Settings for work.

Learning styles. There is no good evidence that matching teaching to a learner’s preferred style helps (Pashler and colleagues, 2008). No onboarding quiz, no “visual learner” labels.
Growth-mindset scales and interventions. The average effect on achievement is very small (Sisk and colleagues, 2018), and the scales are open to reference bias.
Inferred emotion. The system never guesses how a learner feels from their face, voice or typing. The EU AI Act bans emotion recognition in schools and workplaces (Article 5).
Corporate folklore. The 70-20-10 split has never been tested as a rule, and smile-sheet scores are not evidence of learning (see principle 10).
Gamification in feedback. Streaks and rewards stay in the learner practice products only. Leaderboards are never used, and nothing from gamification feeds the data an educator or manager sees.

Where Each Principle Lives

Marking and Feedback is built and live in the demo; the other parts are in testing or planned. The same map applies in K12 and corporate settings, and only the content and the error map change.

Principle	Quiz Blocks	Assignment Builder	Adaptive Practice	Marking and Feedback	Learner Help Agent	Educator and Learner Views
1. Retrieval practice	●	●	●	●
2. Spaced practice		●	●	planned
3. Interleaving		●	●
4. Fading scaffolding		●	●		●
5. Error map	●		●	●
6. Reasoning first			●	●	●
7. Task-focused feedback, human in the loop				●	●
8. Help level as a measure					●	●
9. Analytics against own past						●
10. Motivation and transfer			●			●

Marking and Feedback covers the pre-marking check, AI grading with educator review, and the correction loop.