eduKateSG Learning Node Series · 0083

How Networked Improvement Communities Work | Let Many Schools Learn From One Shared Problem Without Copying One Solution

One school finds a better way to help students explain mathematical reasoning. Another school struggles with the same problem and never hears about it. A third hears about the first school, copies the visible routine, and gets a different result because the timetable, prior knowledge, teacher preparation and student population are different.

Education has no shortage of good ideas. It has a harder problem: learning reliably across variation.

A networked improvement community, or NIC, is one attempt to solve that problem. It coordinates people across schools, districts, research organisations or other sites around a shared improvement aim. Members build a common understanding of the problem, develop a theory of improvement, test changes in local settings, use measures designed for learning, compare what happens across contexts and feed the results back into the network.

The network does not ask every school to copy the same answer. It asks many schools to help discover which mechanisms work, for whom, under which conditions, and how the system must change to make improvement reliable.

The 50-Second Read

• A networked improvement community is an organised group working across sites on a shared, measurable improvement problem.
• Carnegie Foundation descriptions emphasise four characteristics: a well-specified aim; deep understanding of the problem and system; disciplined use of improvement science; and coordination that accelerates testing, refinement and integration across contexts.
• The network begins with the problem, not a predetermined solution.
• Variation across schools is information. It can reveal boundary conditions, hidden supports and mechanisms.
• A theory of improvement links the desired outcome to the system changes expected to produce it.
• Small tests make it possible to learn before scaling expensive changes.
• Shared measures should support learning, not merely accountability.
• The network hub coordinates evidence, methods, communication and cumulative memory; it should not become the only place where thinking occurs.
• Local adaptation is often necessary, but adaptations must preserve the mechanism being tested if the network is to learn anything coherent.
• Failure is useful when it is visible, interpretable and rapidly returned to the network.
• Network social structure matters: roles, relationships, trust, norms and identity determine whether technical methods become collective learning.
• The aim is faster, more reliable learning across diverse settings—not faster copying of one celebrated practice.

Canonical Owner Boundary

This page owns the networked improvement community mechanism: multiple sites learning together around a shared aim through system analysis, improvement science, common measures, local tests and coordinated knowledge accumulation. How Education Works | School Improvement remains the umbrella owner for whole-school change. How Continuous Improvement Works owns the broader logic of iterative improvement. How Professional Learning Communities Work owns recurring educator communities, usually within a school or professional group. How Research–Practice Partnerships Work owns the long-term bridge between research and practice organisations. A NIC may include RPPs and PLCs, but its distinctive owner boundary is coordinated improvement learning across multiple sites.

1. Start With One Shared Problem

Networks often begin by sharing practices: lesson plans, intervention programmes, dashboards, training materials. NICs begin more fundamentally: what problem are we trying to solve?

“Improve mathematics” is too broad. “Increase the proportion of Secondary 1 students who can independently translate unfamiliar word problems into correct algebraic representations by the end of Term 2” is closer to an improvement aim.

A shared problem gives different sites a common object of learning even when their local routes differ.

2. A Well-Specified Aim Creates Direction

A NIC needs an aim precise enough to guide decisions. Good aims often specify the population, capability or outcome, magnitude of change and time horizon.

The point is not management theatre. Without a clear aim, every improvement idea can look relevant and the network cannot tell whether activity is converging toward anything.

3. The Problem Must Be Understood as a System

If students fail to complete homework, the visible outcome may be influenced by task design, prior understanding, home conditions, platform access, feedback timing, competing workload, teacher follow-up and student beliefs about whether the work matters.

A NIC maps these relationships. It asks where the current system reliably produces the unwanted outcome and which conditions vary across sites.

4. Variation Is Not an Annoyance

Traditional implementation often treats variation as deviation from the model. Improvement networks treat some variation as evidence.

Why does the same routine work in School A but not School B? Is teacher preparation different? Do students enter with different prerequisite knowledge? Does one timetable allow repeated practice while another compresses the sequence? Does leadership protect common planning time?

Variation can reveal which conditions are part of the mechanism.

5. Build a Theory of Improvement

A theory of improvement connects the desired outcome to the system changes expected to cause it. It is a working map, not a decorative diagram.

If the aim is stronger independent problem representation, the theory might include teacher modelling of representation choices, deliberate comparison of problem structures, student explanation of method selection, mixed practice and delayed independent checks.

Each component becomes testable.

6. Driver Diagrams Make the Theory Operational

Improvement teams often use driver diagrams to connect the aim to primary drivers, secondary drivers and candidate changes. The value is not the template itself. It forces the network to state why a proposed action should affect the outcome.

If a change idea cannot be linked to a plausible driver, it may be attractive activity rather than improvement work.

7. Measures Should Help the Network Learn

High-stakes accountability measures are often too slow or coarse for rapid learning. NICs need measures close enough to the change to show whether the system is moving.

• Outcome measures ask whether the desired result improved.
• Process measures ask whether the important mechanism is happening.
• Balancing measures ask whether improvement in one area creates damage elsewhere.

A homework intervention that raises completion by doubling teacher workload needs a balancing measure. A writing routine that improves rubric scores but reduces independent planning may need one too.

8. Measurement for Learning Is Different From Measurement for Judgment

If every data point becomes a performance ranking, sites learn to protect themselves. Weak results disappear, implementation problems are softened and measures become targets.

Improvement measures need enough psychological and organisational safety that teams can say, “This did not work here, and here is what we think happened.”

9. Small Tests Reduce the Cost of Being Wrong

A school system does not need to roll out every idea across fifty schools before learning anything. A change can be tested in one class, one department or one short cycle, then refined before wider use.

Small tests are not automatically weak. They are useful when their purpose is learning about design and implementation rather than making a final causal claim.

10. Plan–Do–Study–Act Is a Learning Loop, Not a Form

Plan what will change and what you expect. Do the test. Study the evidence against the prediction. Act by adopting, adapting or abandoning the change.

The cycle loses value when teams fill in PDSA paperwork after the event. The prediction must exist before the result if the test is to discipline reasoning.

11. The Network Hub Coordinates Without Owning All Intelligence

A central hub may maintain measures, facilitate learning sessions, synthesise results, support methods, document changes and connect sites. This role is essential.

It becomes dangerous when local teams wait for the hub to interpret every signal. A NIC should distribute learning capability while preserving enough central coordination to make findings comparable.

12. Local Adaptation Needs a Boundary

One school may need a five-minute routine; another may need ten. One may deliver it through paper, another digitally. Surface variation can be harmless.

But if one site removes the retrieval step from a retrieval intervention, the network is no longer testing the same mechanism. Sites need clarity about what is adaptable and what is an active ingredient.

This is where the NIC connects to How Implementation Fidelity Works: coherence requires protecting the causal core without forcing identical surface behaviour.

13. Standardise the Learning Language Before Standardising the Practice

Networks need common definitions: what counts as the target outcome, what counts as implementation, what constitutes a test, and how context is recorded.

Without common language, two sites can report “success” while measuring different things. Shared definitions create a comparison surface without erasing local context.

14. Context Data Should Travel With Result Data

A result without context tempts imitation. A result with context invites explanation.

Record the class level, prior knowledge, staffing, schedule, dosage, adaptations, support received and unusual events that may matter. The goal is not exhaustive documentation. It is enough context to understand why a result may or may not travel.

15. Failed Tests Are Network Assets

If only successful tests are shared, the network develops a distorted map. Sites repeat failed ideas because the failure history vanished.

A useful failure report says what was attempted, what was expected, what happened, how implementation differed and what the team would test next. Failure becomes cumulative knowledge rather than local embarrassment.

16. Learning Needs a Fast Return Path

A test that finishes in March but is not discussed until December has lost much of its improvement value. NICs require communication channels that return interpretable evidence while decisions are still live.

That may mean short learning huddles, shared run charts, structured test reports, cross-site calls or a searchable knowledge repository.

17. Run Charts Make Time Visible

A single before-and-after number hides the path. Time-series displays such as run charts can reveal whether performance changed after an intervention, whether improvement is stable, whether the system was already moving and whether gains disappear during holidays, staffing changes or examination periods.

The chart does not establish causality by itself. It helps the improvement team reason about sequence and variation.

18. Heterogeneity Is a Scientific Opportunity

If five sites improve and three do not, the network should resist averaging everything into one headline. The difference is the interesting part.

Which sites had stronger implementation? Which had different learners? Which changed another part of the system at the same time? Which adaptation preserved the mechanism? Variation can reveal boundary conditions that a single-site success story cannot.

19. Social Structure Is Part of the Intervention

Technical routines alone do not create a NIC. People must trust one another enough to share incomplete work, admit failure and expose local constraints. They need roles, norms, relationships and a sense that the network’s problem is genuinely shared.

Recent Carnegie work explicitly treats the social structure of NICs as part of the architecture: scientific-professional learning requires both technical improvement methods and a community capable of using them.

20. Strong Ties and Weak Ties Do Different Work

Close collaborators can develop deep shared understanding. More distant network connections can introduce different ideas and expose the group to contexts it would not otherwise see.

A healthy network needs enough strong ties for trust and enough weak ties for novelty.

21. Improvement Coaches Build Local Learning Capacity

Sites may need support to define aims, select measures, run small tests, interpret variation and document learning. Improvement coaches can help without dictating local solutions.

The coach’s long-term success is not permanent dependence. It is a site that becomes increasingly capable of running disciplined improvement itself.

22. Data Support Is Infrastructure

A NIC cannot learn quickly if every site spends weeks cleaning incompatible spreadsheets. Shared definitions, lightweight data pipelines and accessible visualisation reduce the cost of each cycle.

But data infrastructure should follow the improvement question. A sophisticated dashboard that measures the wrong construct is an expensive distraction.

23. Spread Is Not the Same as Scale

Spread means a practice appears in more places. Scale means a system can support reliable performance across those places.

A routine that succeeds in three enthusiastic pilot schools may fail when expanded because coaching capacity, materials, leadership attention or protected time do not scale with it. NICs study the support system as well as the practice.

24. Network Learning Should Change the Theory

The theory of improvement is not sacred. If repeated tests show that an assumed driver does not move the outcome, the model should change.

A network that keeps the same theory while collecting contradictory results is performing improvement rather than learning.

25. Equity Must Be Inside the Aim

An average improvement can hide widening gaps. If high-attaining students benefit while those with weaker prior knowledge fall further behind, the network has not solved the same problem for everyone.

Disaggregate outcomes where ethically and statistically appropriate. Include learners and communities in problem definition. Ask whether the burden of an intervention falls unevenly on teachers, families or students.

26. Data Governance Must Match Network Scale

Multiple organisations create additional risk. Who can see site-level results? Which data can be pooled? Can results be published by school? How are small groups protected? What happens when a member leaves the network?

Clear governance makes honest participation safer.

27. Cross-Domain Comparison: Aviation Safety Reporting

Aviation improves partly because incidents and near misses can be studied across many flights rather than being treated as isolated local events. Repeated patterns reveal system vulnerabilities that no single crew could see.

A NIC can create a similar learning advantage. One school’s implementation failure may be idiosyncratic. The same failure across six schools points toward a system condition worth redesigning.

28. Cross-Domain Comparison: Distributed Software Systems

Large software services run across many machines and environments. Engineers need observability: common signals that show what happened while preserving enough context to diagnose local failure.

Education networks need the same balance. Common measures make comparison possible; contextual information explains why sites differ. Standard signals without local context create false simplicity. Local stories without common signals create incomparable anecdotes.

29. Cross-Domain Comparison: Public Health Improvement

Public health systems often coordinate surveillance, local interventions and learning across jurisdictions. An outbreak or behaviour pattern may look different by place, but common definitions and rapid communication let local observations contribute to a wider response.

A NIC similarly turns local educational variation into shared intelligence without pretending every site is identical.

30. Example: The Secondary Mathematics Representation Problem

Eight schools observe that students can solve equations once formed but struggle to translate unfamiliar contexts into equations. The network agrees on a common aim and a short representation measure.

Sites test different changes: contrastive examples, diagramming routines, explicit quantity-language mapping and student explanation of variable choices. The network tracks which mechanisms improve transfer and under which prior-knowledge conditions.

After several cycles, the network does not simply announce the “best worksheet.” It develops a stronger theory: successful sites repeatedly require students to externalise relationships before algebraic manipulation, and the effect weakens when prerequisite ratio reasoning is unstable.

31. Example: Attendance After a School Transition

A network of schools wants to reduce repeated absence during the transition into a new level. One site tests personalised family contact, another transport support, another rapid follow-up after the second absence.

Shared measures show that the same intervention does not work equally everywhere. Transport support matters in one cluster; relationship continuity matters more in another. The network refines the problem theory instead of forcing a universal programme.

32. Example: Feedback That Students Actually Use

Several schools have invested in teacher feedback, yet students often read comments without revising work. The NIC shifts the outcome from “feedback given” to “student response produces a demonstrable improvement.”

Sites test dedicated revision time, reattempt requirements, shorter actionable feedback and exemplar comparison. Because the outcome and process measures are shared, the network can identify which combinations improve uptake without increasing teacher workload beyond sustainable levels.

33. The Carnegie NIC Model

The Carnegie Foundation has been central to the education NIC movement. Its framework describes networked improvement communities as characterised by four features: a well-specified common aim; a deep understanding of the problem, the system that produces it and a theory of improvement; discipline through improvement science; and coordinated organisation that accelerates development, testing, refinement and integration across contexts.

The significance is easy to miss. A NIC is not simply a professional network plus PDSA cycles. It is an architecture for cumulative scientific-professional learning across variation.

34. Technical Core Plus Social Architecture

Later Carnegie work has paid increasing attention to the social structure required for that architecture: roles, relationships, norms, identity and network organisation. Improvement science can specify how teams learn from tests, but social structure determines whether participants will expose uncertainty, share weak results and act on knowledge generated elsewhere.

This is why a network can possess all the right templates and still fail to become a learning community.

35. Failure Mode: Networking Without an Aim

Schools meet, share presentations and leave inspired. Nothing accumulates because the network has no common problem, measure or testable theory.

Connection is valuable, but connection alone is not networked improvement.

36. Failure Mode: Solutionitis

The network begins with a solution—an app, programme, protocol or training package—and defines the problem in whatever way justifies adoption.

Improvement reverses the order: understand the problem, build a theory, then select or design changes that address the relevant drivers.

37. Failure Mode: Dashboard Theatre

The network has beautiful graphs but cannot explain which process generated the movement. Sites optimise metrics while the underlying learning problem remains.

Measurement should sharpen the theory, not replace it.

38. Failure Mode: Forced Uniformity

Every site is required to use the same script, schedule and materials despite different conditions. Variation is treated as noncompliance.

The network loses one of its greatest scientific assets: the opportunity to learn how context changes performance.

39. Failure Mode: The Hub Becomes a Bottleneck

All data interpretation, facilitation and decision-making flow through a tiny central team. As membership grows, learning slows and local capability remains weak.

A mature hub standardises what must be common and deliberately distributes improvement skill.

40. Failure Mode: Success-Story Contagion

A charismatic school presents a dramatic result. Other sites copy the visible practice before the network understands the baseline, implementation support or context. The anecdote outruns the evidence.

Celebrate results after asking what mechanism, measure and conditions produced them.

41. Failure Mode: Network Overload

Too many meetings, forms and reporting requirements consume the time needed to improve teaching. The network becomes another layer of work rather than a way to make work better.

Coordination has a cost. Every network routine should earn its place by improving learning speed or reliability.

42. A Practical NIC Cycle

1. Define one consequential shared problem.
2. Specify a measurable aim and time horizon.
3. Map the system currently producing the outcome.
4. Study variation across sites before prescribing a solution.
5. Build a working theory of improvement.
6. Identify primary and secondary drivers.
7. Select a small number of candidate changes.
8. Agree common outcome, process and balancing measures.
9. Define what local context should be recorded.
10. Run small tests with explicit predictions.
11. Study results quickly.
12. Share failures as well as successes.
13. Compare patterns across sites.
14. Update the theory when evidence contradicts it.
15. Clarify which parts of a change are adaptable and which are essential.
16. Expand only after the support system is understood.
17. Build local improvement capability as the network grows.
18. Preserve the test history and decision rationale.
19. Keep asking what works, for whom, under which conditions.

43. Missing-Node Scan: Questions a Mature NIC Should Ask

• Can every member state the shared aim precisely?
• Is the network studying a problem or promoting a solution?
• Does the theory of improvement explain why each change should matter?
• Are measures frequent enough to support learning?
• Are balancing measures protecting against hidden costs?
• Can sites report weak results without fear of ranking or embarrassment?
• Does context travel with outcome data?
• Are common definitions strong enough for comparison?
• Can the network distinguish adaptation from loss of the active ingredient?
• Are local teams becoming better at improvement, or more dependent on the hub?
• Does the network inspect heterogeneity rather than averaging it away?
• Are students and communities represented where the problem affects them directly?
• Does the network know which supports will be required at larger scale?
• Can members find the history of earlier tests and failed ideas?
• Does the theory change when evidence changes?

44. Authoritative Research Starting Points

Useful starting points include the Carnegie Foundation’s Why a NIC?, which sets out the four defining characteristics; its work on Getting Ideas into Action: Building Networked Improvement Communities in Education; and newer Carnegie work on the social architecture of networked improvement communities. A 2025 article, Catalyzing Scientific-Professional Learning Communities, further examines how network hubs can coordinate diverse schools, districts and participants around shared problems and theories of improvement.

These sources place NICs at the intersection of improvement science, professional learning and network organisation. Their central promise is not that a network automatically knows more than a school. It is that disciplined coordination can make many local tests accumulate into shared knowledge faster than isolated improvement efforts.

45. The Return Path

At the start, eight schools share the same frustrating sentence: students understand examples in class but fail unfamiliar mathematics problems independently.

One year later, the network knows much more. It has learned that representation skill, not manipulation, is the common bottleneck; that contrastive examples help when students explain structural similarities themselves; that the routine fails when prerequisite ratio reasoning is unstable; that teacher modelling must fade; and that sites need common planning time to maintain the sequence.

No school discovered all of this alone. No school simply copied another. The network turned variation into evidence and evidence into a more reliable model.

Networked improvement communities work when many sites stop behaving like separate anecdotes and begin behaving like a coordinated learning system.