“Weak at Mathematics.”

That is a measurement.

It is also almost useless for training.

“Weak at algebra.”

Better.

“Factorisation is accurate in blocked practice; mixed quadratic method selection fails when the method is not named.”

Now we have a training job.

Training measurement resolution is the deliberate choice of how finely to distinguish a learner’s performance state so that the distinction is detailed enough to change the next instructional decision, but not so detailed that the diagnosis becomes noisy, artificial or impossible to act on.

The central question is not:

How much data can we collect?

It is:

What is the smallest distinction that changes what we should do next?

---

Quick Read: Coarse Enough to Manage, Fine Enough to Repair

Broad Symptom → Split by Mechanism → Test the Branch → Stop When the Next Action Becomes Clear → Repair → Recombine

Too coarse:

• “bad at English”;
• “careless”;
• “doesn’t understand Science”;
• “weak foundation.”

Too fine:

• inventing twenty tiny categories from three errors;
• assigning a separate label to every hesitation;
• claiming exact mastery percentages from sparse evidence;
• tracking distinctions that do not change instruction.

Useful resolution sits between them.

Resolution Is About Decision Scale

Different decisions need different measurement resolution.

A parent deciding whether a child needs broad support may need a coarse view:

• English reading stable;
• writing unstable;
• Mathematics mixed;
• Science strong.

A tutor deciding what to teach in the next ten minutes needs a finer view:

• idea generation strong;
• paragraph structure stable;
• evidence selection weak;
• claim strength overextended.

A curriculum planner may need a different resolution again.

No single granularity is universally correct.

Why Total Scores Lose Diagnostic Information

Total scores aggregate.

Aggregation is useful for ranking, reporting and broad monitoring.

Aggregation also destroys detail.

A 2026 study in Thinking Skills and Creativity used cognitive diagnostic modelling to move from coarse to fine diagnosis of EFL reading inference across six subskills. The study found that fine-grained profiles could distinguish learners with different patterns even when broad proficiency information was insufficient.

A 2026 Behavioral Sciences study similarly argues that traditional total-score approaches are often too coarse for precise instructional support and develops a model that distinguishes knowledge components and cognitive levels.

Small-group tuition does not need those models.

It needs the same design question:

What useful information disappeared when we compressed the learner into one score?

The Resolution Ladder

A practical diagnostic ladder can move from broad to narrow.

Level 1: Subject.
Mathematics, English, Science.

Level 2: Domain.
Algebra, comprehension, scientific reasoning.

Level 3: Capability family.
Method selection, evidence inference, experimental interpretation.

Level 4: Mechanism.
Recognition, representation, retrieval, execution, checking, transfer.

Level 5: Vulnerable transition.
The exact point where the learner’s valid route first becomes invalid.

Do not descend automatically to Level 5.

Stop at the first level where the next intervention becomes sufficiently clear.

Mira: From “Weak at Algebra” to a Trainable State

Mira loses marks in algebraic questions.

Coarse label:

Weak at algebra.

The tutor samples:

• expansion;
• factorisation;
• linear equations;
• quadratic method selection;
• algebra inside geometry.

Expansion and factorisation are accurate when named.

Linear equations are stable.

Quadratic method selection is inconsistent.

Algebra inside geometry becomes slow because Mira first struggles to translate the diagram.

One label has become two different mechanisms:

• method discrimination in quadratics;
• representation translation in geometry.

That is enough resolution to change the training plan.

No need to invent thirteen micro-skills beneath them unless later evidence demands it.

Jonas: From “Weak Comprehension” to Evidence Calibration

Jonas’s comprehension results are inconsistent.

Broad diagnosis:

Weak comprehension.

But literal retrieval is strong.

Pronoun reference is stable.

Inference generation is usually plausible.

The recurring failure is that his answer is stronger than the evidence.

The useful resolution is:

Inference claim-strength calibration.

That phrase may sound technical, but it points to a precise training action: contrast candidate claims against the evidence boundary.

Nadia: From “Weak Science Answering” to Evidence Integration

Nadia knows the Science content.

Her written answers still lose marks.

A coarse system calls this:

Answering technique.

The tutor decomposes:

• concept identification;
• evidence selection;
• mechanism;
• conclusion;
• language form.

Concept and language are stable.

Evidence selection is strong.

The weak link is connecting evidence to mechanism.

That is the resolution the next lesson needs.

Multiple Granularities Can Be Useful at the Same Time

A 2026 paper in ACM Transactions on Information Systems proposes multiple-granularity cognitive diagnosis because different users may need different levels of diagnostic detail. The system models coarse and fine concepts simultaneously rather than assuming one granularity fits every purpose.

This translates naturally to education.

A parent may need:

Mathematics foundations are stable; mixed method selection is the current priority.

The tutor may need:

Quadratic method selection fails specifically when factorisable and non-factorisable cases are interleaved; one structural cue restores the boundary.

Both are correct.

They operate at different resolutions.

Resolution Should Follow the Intervention

If two diagnoses lead to the same intervention, separating them may not yet be useful.

Suppose Mira makes two kinds of basic sign errors.

Both improve through the same short routine: preserve signs explicitly, estimate, then verify.

There may be little benefit in building two separate tracking categories unless one later behaves differently.

Resolution should therefore be action-sensitive.

Split the category when the split changes what you would teach, practise, cue, measure or stop.

Resolution and Training Observability

Training Observability asks whether the relevant learner state leaves enough visible traces.

Measurement resolution asks what distinctions we make once those traces are available.

Working shows where Mira’s route diverges.

Resolution determines whether we code that divergence as:

• “wrong”;
• “algebra error”;
• “sign-control error”;
• “negative distribution after bracket expansion.”

The last category may be useful for a recurring pattern.

It may be excessive for a one-off slip.

Resolution and Training Sampling

Fine resolution requires enough evidence.

One item rarely supports a highly specific learner label.

Use Training Sampling.

If Jonas overclaims once, record the observation.

If he overclaims across narrative, article and speech passages, the finer category becomes more defensible.

The finer the claim, the more carefully the sample should support it.

Resolution and Measurement Noise

Every time we split a category, we create smaller groups of evidence.

Smaller groups are more vulnerable to noise.

If Mira attempts only two “representation translation” items, a 50% success rate does not justify dramatic conclusions.

Fine-grained measurement can create an illusion of precision.

Training Measurement Noise therefore places a natural limit on resolution.

Resolution and Validity

A finely named category is not automatically valid.

Suppose a tutor labels an error “working-memory overload” from one slow response.

The label sounds precise.

The evidence is not.

Training Validity requires the inference to stay within what the task can support.

Use behavioural descriptions before psychological labels when evidence is limited.

Needs three rereads before identifying the question demand

is safer than:

has an attention disorder.

Training diagnosis is not medical diagnosis.

Resolution and Ceiling/Floor Effects

A measurement can have the right conceptual granularity and still sit at the wrong difficulty range.

If all items are easy, fine distinctions among strong learners disappear into a ceiling.

If all items are too hard, fine distinctions among lower states disappear into a floor.

Resolution requires both conceptual and difficulty range.

Resolution and Training Responsiveness

Fine-grained measures can detect early change that broad grades miss.

Jonas’s English grade stays at 68%.

But within inference tasks:

• evidence selection improves;
• cue dependence falls;
• claim-strength errors decline.

The finer measure is more responsive to the training target.

Later, the broader grade should be revisited to confirm that component improvement recombines into whole performance.

Resolution and Training Branching

Training Branching needs enough resolution to choose different next actions.

“Wrong” gives only one branch: more help.

Higher resolution allows:

• missing model → explain;
• wrong competitor → contrast;
• unstable execution → repeat;
• cue dependence → fade support;
• surface dependence → vary;
• stable fresh performance → progress.

Resolution earns its value by improving branch selection.

Resolution and Training Dependencies

High-level failure often requires downward resolution.

Mira cannot solve an unfamiliar geometry problem.

Split the task:

• interpret diagram;
• identify relationship;
• form equation;
• solve algebra;
• check units.

Now test the suspected dependency.

If the diagram relationship is supplied and algebra succeeds, the resolution has revealed the bottleneck.

This connects directly to Training Dependencies.

Hierarchical Resolution: Do Not Flatten the Map

Capabilities are often hierarchical.

“Comprehension” can contain inference.

Inference can contain evidence location, semantic integration and claim calibration.

Mathematics can contain algebra.

Algebra can contain symbol interpretation, equivalence, manipulation and checking.

A 2026 article in Educational Measurement: Issues and Practice explores hierarchical attribute structures in mathematics diagnostic classification, highlighting the value of representing cognitive skills in structured relationships rather than as a flat list.

The practical implication is that diagnostic resolution should preserve hierarchy.

Parents can see the branch.

Tutors can descend to the leaf when needed.

The Resolution Stop Rule

How far should diagnosis descend?

Use four stop conditions.

• Action clarity: the next intervention is clear.
• Evidence sufficiency: current data cannot support a finer claim reliably.
• Cost: further diagnosis would consume more time than the likely benefit.
• Recombination: the identified component can be repaired and returned to the whole task.

If all four conditions are met, stop diagnosing and train.

Resolution Can Increase Temporarily

During diagnosis, zoom in.

During practice, zoom out.

During verification, return to whole performance.

This prevents a learner from living permanently inside isolated micro-skills.

For Mira:

Whole quadratic problem → method-selection diagnosis → contrast drill → mixed quadratic problem → delayed full-paper return.

The resolution expands and contracts with the training job.

Resolution and the Three-Student Classroom

A small group makes multi-resolution teaching possible without creating three separate curricula.

All three students may work on the same broad Mathematics topic.

Mira’s micro-target is method selection.

Jonas’s may be sign control.

Nadia’s may be checking.

The whole-task context remains shared.

The diagnostic resolution differs.

This is a practical form of personalisation that does not require a separate syllabus for every learner.

Parent Resolution: What Families Need to Hear

Parents usually do not need twenty diagnostic codes.

They need a compressed but actionable explanation.

Mira knows the algebra. The current issue is choosing the right method when topics are mixed. We are training discrimination, not restarting algebra from zero.

This is high enough resolution to guide home expectations and low enough complexity to remain usable.

Student Resolution: What the Learner Needs to Know

The learner’s explanation should be even more operational.

You can do the methods. Your next job is to notice which structure tells you which method belongs.

The diagnosis becomes a training instruction rather than an identity.

False Precision: The Danger of Decimal Diagnoses

Technology makes precise-looking outputs easy.

“Mira has 73% mastery of representation translation.”

What does that number mean?

How many items support it?

How valid was the attribute map?

How much uncertainty remains?

Contemporary cognitive-diagnosis research itself emphasises uncertainty, Q-matrix quality and the difficulty of mapping external performance to underlying cognitive states. A 2026 Information Sciences paper describes precise modelling of the relationship between external performance and underlying cognitive states as a continuing challenge.

Small-group tutoring should therefore resist fake decimals.

Use ranges, states and confidence language when evidence is sparse.

Diagnostic Categories Should Be Behaviourally Anchored

A useful category should be recognisable in work.

Instead of:

low executive control

write:

does not return to skipped questions without an external reminder.

Instead of:

weak metacognition

write:

confidence remains high after repeated method-selection errors and does not change after feedback.

Behavioural anchors make high-resolution categories testable.

Resolution Should Preserve the Whole Learner

Fine-grained diagnosis can fragment the child.

Thirty skill cells.

Twenty error codes.

Ten confidence indicators.

But a learner performs as an integrated system.

After micro-repair, return to whole tasks.

Ask whether the repaired component actually improves reading, writing, problem solving or examination performance.

Training Recombination protects against permanent fragmentation.

Do Not Split Before You Observe

A tutor arrives with a twenty-category checklist and tries to force every error into it.

The categories begin driving the observation.

Instead, begin with the learner’s route.

Use categories to organise recurring mechanisms after they appear.

Do Not Split When the Evidence Is Sparse

One error.

One new label.

That produces diagnostic clutter.

Record an observation first.

Promote it to a stable category when sampling shows a pattern.

Do Not Stay Coarse Because Fine Diagnosis Is Hard

The opposite failure is familiar:

Do more Mathematics.

If the learner’s actual problem is method recognition, generic volume may simply repeat the failure.

Move down one level.

Ask what mechanism keeps producing the lost marks.

Do Not Stay Fine After the Repair Is Stable

Mira repairs negative-sign distribution.

It remains stable across mixed work for several weeks.

Stop tracking it as an active micro-target.

Let ordinary Mathematics maintain it.

Measurement resolution should relax as the skill becomes routine.

The Resolution Budget

Every extra diagnostic distinction costs:

• observation time;
• recording time;
• sample size;
• cognitive attention;
• interpretation effort;
• communication complexity.

Spend resolution where it buys better decisions.

High-cost recurring errors deserve more resolution.

Low-cost one-off slips deserve less.

Resolution for Examination Training

An examination paper is a whole-system sample.

After marking, increase resolution only where lost marks cluster.

For each recurring cluster, ask:

• knowledge?
• retrieval?
• interpretation?
• representation?
• method selection?
• execution?
• control?
• transfer?

These are operational families, not perfect scientific categories.

Their value is that each leads to a different repair.

This route connects to the existing Error Analysis for Exams owner.

Resolution for Home Study

Parents do not need to diagnose every cognitive mechanism.

A useful home resolution is often:

• can start alone / cannot start alone;
• knows method / does not know method;
• can execute / execution breaks;
• can check / does not check;
• stable today / uncertain today.

Bring unresolved cases into tuition rather than turning home into a full diagnostic lab.

Resolution for School Feedback

School feedback often arrives at different resolutions:

• grade;
• question marks;
• teacher comment;
• rubric;
• annotated script.

Each layer can be translated into training resolution.

A general comment like “needs more elaboration” can be unpacked by sampling whether the problem is idea quantity, explanation depth, example relevance or paragraph organisation.

Resolution for Transitions

At stage transitions, old coarse categories may stop being useful.

“Good at Primary Mathematics” is not enough to predict which Secondary Mathematics dependencies are ready.

Increase resolution around the foundations the next stage assumes:

• fraction fluency;
• equality;
• symbol interpretation;
• representation switching;
• independent checking.

Then reduce resolution again once the transition is stable.

A Parent Measurement-Resolution Audit

• Is the current diagnosis too broad to guide action?
• What smaller distinction would change the next step?
• Do we have enough evidence for that finer claim?
• Are we using behaviour rather than identity labels?
• Have we split the problem more finely than we can actually teach?
• Does the child understand the diagnosis in usable language?
• When can this micro-target leave active tracking?

A Tutor Measurement-Resolution Audit

• What decision am I trying to improve?
• What is the current diagnostic granularity?
• Would one finer split change the intervention?
• What observation can test that split?
• Is the sample large enough?
• Am I creating false precision?
• Does the category have a behavioural anchor?
• What is the stop rule?
• How will the component be recombined?

The Deeper Idea: Diagnosis Is a Zoom Lens

A zoom lens is useful because it can change scale.

Too far out and important detail disappears.

Too far in and context disappears.

Training diagnosis works the same way.

Start broad enough to see the whole performance.

Zoom until the limiting mechanism becomes trainable.

Repair.

Then zoom back out to see whether the whole performance improved.

The best diagnostic resolution is not the finest description we can produce. It is the finest description that improves the next decision without losing the learner inside the categories.

Research Foundations and Evidence Boundaries

Current diagnostic-measurement research provides strong conceptual support for multi-resolution thinking. The 2026 MGCD multiple-granularity cognitive diagnosis paper explicitly models coarse and fine concepts simultaneously. The 2026 Exercise–Knowledge–Cognition model aims to generate more interpretable, fine-grained profiles than simple total scores. The 2026 purpose-built reading diagnostic study demonstrates the importance of subskill-level design and validation. The 2026 hierarchical-attribute mathematics paper shows why cognitive skills can be structured rather than treated as a flat list. The 2026 Behaviormetrika preface on methodological innovations in cognitive diagnosis places these developments within the broader challenge of producing actionable feedback about strengths and weaknesses.

These studies use sophisticated psychometric models and usually much larger data sets than a tuition classroom. They do not justify assigning precise mastery probabilities to a child from a handful of worksheet items. The transfer made here is narrower: diagnostic usefulness depends on granularity; different decisions need different resolutions; fine-grained claims require adequate evidence; and useful diagnosis should terminate in an actionable training decision rather than in ever-finer classification.

Continue Through How Training Works

Read this alongside Training Observability, Training Sampling, Training Validity, Training Measurement Noise, Training Responsiveness, Training Branching, Training Decomposition and Training Recombination.