How well can you reason, remember and spot patterns?
Work through a set of short thinking challenges involving words, shapes, patterns, memory and quick visual checks. Read carefully, have a genuine attempt, and let the questions adjust as you go.
Work through a set of short thinking challenges involving words, shapes, patterns, memory and quick visual checks. Read carefully, have a genuine attempt, and let the questions adjust as you go.
Enter your age, choose a length, then begin when you are ready. Most people can leave the advanced settings alone.
Use an earlier withdrawal code to delete that session’s saved research record. New tests do not upload answers or timings.
Read the question carefully before choosing.
Use number keys or select an answer.
This page shows how you went across the different sections.
Questions adapt during the test, so percentages correct cannot be compared directly across people or sections. Response time excludes invalid timings and breaks.
The score uses an IQ-style scale with 100 at the project reference centre. This version includes lower-range and high-ceiling questions, but it is not yet based on representative age norms.
These experimental model estimates are not age-normed. Question difficulty changes during the test, so percentages correct should not be compared between sessions or areas.
The test does not jump to conclusions after one answer. It begins with a balanced set, then adds a few questions where it still needs a clearer picture.
Every area receives several questions, using more than one type of task.
Correct answers generally lead to harder questions. Incorrect answers lead to a more suitable level. Untimed reasoning is not rewarded simply for being fast.
The final few questions are used where the program is least certain, rather than being added at random.
The complete paper explains the question maker, adaptive selection, Bayesian score updates, composite calculation, IQ-style conversion, uncertainty range, worked example and present limits of the model.
The conversion makes the result easier to read. It does not create representative age norms or turn the project into a clinical test.
Because the item difficulties have not yet been calibrated on a large sample, their influence is deliberately reduced and the displayed ranges are kept broad.
A basic paper on question generation, adaptive testing, Bayesian score updates, composite scores and the experimental IQ reading.
ACRA v0.7.6 is a teacher-built, browser-based cognitive assessment created by Mr. Sheedy. It contains 60 procedural question families across verbal reasoning, visual–spatial reasoning, fluid reasoning, working memory and quick visual checking. It maintains a model distribution for each area, updates that distribution after every response, combines the area estimates with bounded precision weights, and converts the final model estimate to an IQ-style scale. Response time is kept separate from ordinary reasoning accuracy, except in the dedicated timed section. The score is experimental: the question parameters, age comparisons and measurement range have not been established from representative samples.
ACRA measures five areas. Each area has several different task families so that a result is not based on one repeated puzzle type.
| Cognitive area | Families | Examples | Main evidence |
|---|---|---|---|
| Words and verbal reasoning | 11 | Analogies, definitions, sentence meaning, categories, verbal deductions and rule chains | Correct and incorrect responses |
| Shapes and space | 11 | Rotation, reflection, paper folding, cube nets, overlays, compound transforms and hidden grid rules | Correct and incorrect responses |
| Patterns and logic | 20 | Matrices, equation systems, multi-stage number patterns, hidden operations, constraints and code-breaking | Correct and incorrect responses |
| Working memory | 10 | Forward, backward, ordered, spatial, paired, dual-stream and path recall | Correct and incorrect recall after a controlled presentation |
| Quick visual checking | 8 | Symbol search, pair comparison, rule checking, code lookup, ordered-pair search and comparative counting | Accuracy within a standard response window |
Two batteries are available. The full assessment uses all five areas. The reasoning-only assessment uses verbal, visual–spatial and fluid reasoning. If the timed processing section is incomplete or non-standard, the five-area score is withheld and the program reports the reasoning estimate instead.
| Mode | Core questions per area | Maximum per area | Extra verification budget | Typical full-battery range |
|---|---|---|---|---|
| Quick | 5 | 8 | 5 | 25–30 scored questions |
| Standard | 7 | 12 | 12 | 35–47 scored questions |
| Extended | 10 | 18 | 25 | 50–75 scored questions |
Every mode begins with a balanced core. The program first gathers a minimum amount of evidence from every selected area and requires several different task families. It then uses a limited set of adaptive verification questions where uncertainty remains high. If an area estimate approaches the lower or upper tail, the program also requires extra evidence from questions designed for that part of the range.
Each question family is a small rule-based generator. It does not select a question from a short fixed list. It builds a new item by changing the underlying rule, values, symbols, shapes, orientations, clue combinations, sequence length, distractors and answer order.
A session begins with a secure random seed. That seed drives a deterministic pseudo-random generator. As a result, the assessment can create varied questions while still retaining enough metadata to reproduce and inspect a particular item later.
Before a question is shown, the program checks that it contains the required fields, has between two and six answer choices, has one valid correct-answer position, has no duplicate options, contains finite difficulty and timing values, and does not render broken values such as undefined or NaN.
Some task families receive additional checks. For example, a code-breaker must leave one possible code, an ordering puzzle must leave one valid order, a logic-matching item must leave one complete matching, and a function machine must isolate one rule. If a generated item fails, the program discards it and tries again.
The finished item is given a structural signature. A question with the same signature cannot appear twice in one session. The selector also avoids the three most recent families used in the same cognitive area when the bank allows it. Correct-answer positions are balanced separately for two-, three-, four-, five- and six-option questions so that the correct answer does not repeatedly appear in the same location.
The assessment does not immediately chase the user’s score after one answer. It first gathers balanced evidence. Within each area, however, question difficulty gradually moves towards the current estimate.
The first two valid questions in an area are aimed broadly around model levels of approximately −0.90 and +0.15. This gives the program an initial sense of whether to move down, remain near the centre, or move upwards.
After the opening questions, the target becomes the current estimated ability for that area. Candidate questions are generated slightly below and above that point. Near the tails, the selector favours dedicated lower-floor or high-ceiling families and may target as low as approximately −3.2 or as high as approximately +3.3 on the project scale. This prevents the assessment from becoming locked to one exact difficulty.
Each candidate receives a project utility score:
Here, Ii is an approximate information score, Vi is a bonus for an unused family, Nf is the number of times that family has already appeared, Ti is a small penalty for an unusually long untimed item, bi is item difficulty, t is the target, and ε is a small random term.
The tail bonus applies to eligible extended-range candidates. Information is calculated as I(θ) = [P′(θ)]² / [P(θ)(1−P(θ))] for the guessing-adjusted response model described below. It reflects the model's assumed item parameters, not measured precision.
The program sorts the candidates and randomly chooses from the best few. This keeps the test adaptive without making every session identical or repeatedly exposing the single highest-ranked item.
An area is ready to finish when it has completed the required core questions, used enough different families, and either reached its provisional uncertainty target or reached its maximum question count. The whole assessment also has a limited verification budget. Standard and Extended modes reserve enough of this budget to confirm low-tail and high-tail performance rather than stopping as soon as a score first crosses 80 or 140.
The program does not award fixed points. It asks: How likely was this response at each possible ability level?
Each cognitive area begins with a standard normal prior centred at zero. The program represents possible ability values, written as θ, on a grid from −4.20 to +4.20 in steps of 0.05. Before any answers, values near zero have more probability than extreme values.
| Symbol | Meaning |
|---|---|
| θ | The user’s possible ability level in the relevant cognitive area |
| bi | The provisional difficulty of question i |
| ai | The provisional discrimination of the question; larger values make the response more informative near its difficulty |
| ci | The chance floor, normally 1/k for a question with k options |
| 0.72 | An evidence-tempering factor that reduces the influence of uncalibrated design parameters |
Suppose a five-option question has c=0.20, design discrimination a=1.20, difficulty b=0.40, and the model is considering θ=0.60. The effective discrimination is 1.20×0.72=0.864, giving:
At that ability level, a correct answer multiplies the existing probability weight by about 0.634. An incorrect answer multiplies it by about 0.366.
For every point θj on the grid, the program updates the weight:
The new weights are then normalised so that they add to one. A correct response to a difficult question usually shifts the distribution upwards more than a correct response to an easy question. An incorrect response to an easy question usually shifts it downwards more than an incorrect response to a difficult question.
| Response pattern | Typical effect |
|---|---|
| Correct on an easy item | Small or moderate upward update because success was already likely |
| Correct on a hard item | Larger upward update because success was less expected |
| Incorrect on an easy item | Larger downward update because failure was less expected |
| Incorrect on a hard item | Small or moderate downward update because failure was already likely |
After the update, the area estimate is the weighted mean of the grid:
The posterior standard deviation describes the remaining spread:
Smaller values of sd indicate a more concentrated estimate. They do not prove that the item bank itself is valid; they only describe certainty within the current project model.
ACRA calculates two possible composites:
An area with lower posterior uncertainty can contribute slightly more, but no area is allowed to dominate. First, the program calculates:
The value 0.18 is a design-uncertainty allowance. Each raw precision is divided by the average raw precision, then limited to the interval 0.75 to 1.25:
The final weights are:
This means a well-measured area may receive somewhat more influence, but the composite remains broadly balanced.
The additional 0.18 term deliberately broadens the composite uncertainty because the question difficulties and section relationships have not yet been established empirically.
The composite model value is converted to a familiar display scale:
The number is rounded to the nearest whole point and its display is bounded to 55–160. The same transformation is used for individual area estimates. The limits and scale are project design choices, not established measurement limits or age norms.
A percentile can be calculated from an assumed standard normal curve:
This is a model percentile, not an observed age percentile from a representative population. The results screen does not report it.
The range combines posterior uncertainty with design allowances for mode, session quality and limited tail evidence:
| Component | Values | Purpose |
|---|---|---|
| Fmode | Quick 7.0; Standard 5.5; Extended 4.25 points | Design allowance for uncalibrated scores; not empirically estimated |
| Qquality | Standard 0; Caution 2.5; Limited 5.5 points | Widens the range when interruptions or poor response conditions reduce confidence |
| Ttail | 0, or 4.5 points when tail checks are incomplete | Design allowance when an extreme model result has limited tail evidence |
The calculation uses the normal 90% multiplier as a design convention:
The model also adds 4.5 points in quadrature when an extreme estimate has insufficient tail checks. Endpoints are rounded and limited to 55–160. Because the underlying item parameters and allowances are not calibrated, this is an engineering uncertainty range, not a validated confidence interval; its coverage is unknown.
Response time does not directly raise or lower verbal, visual or fluid-reasoning ability. A careful correct answer remains correct. A very fast correct answer receives no automatic bonus.
A response is marked as unusually rapid when it is faster than the larger of 850 milliseconds or 11% of that item’s expected time. This flag contributes to the session-quality classification only when rapid responses become common.
Working-memory items are shown through a controlled presentation. If the tab changes, the test is paused or the presentation is interrupted, that item is excluded and replaced. It does not count as evidence of low ability.
Processing items are intentionally timed, but version 0.6.1 gives substantially more time than the earlier build. Depending on the task, form and difficulty, most response windows are now roughly 13–30 seconds, with a small additional allowance for younger users. A failure to answer before the response window ends is treated as an incorrect response if the item was presented normally. Therefore, time influences this section through a task-specific time limit, not through a general speed bonus.
| Classification | Examples of conditions | Effect on the result |
|---|---|---|
| Standard session | Few interruptions, low exclusion rate and no strong rapid-response pattern | No additional quality uncertainty |
| Interpret with caution | Early ending, more than 10% excluded items, more than 20% invalid timing, more than 18% rapid responses, or more than four focus losses | Adds 2.5 points to the uncertainty calculation |
| Not enough clean evidence | Very little usable evidence or more than 30% excluded items | Adds 5.5 points to the uncertainty calculation and warns the user |
Importantly, session quality changes the width of the range, not the centre score. The program does not deduct IQ points merely because the user paused or changed tabs.
Suppose a Standard reasoning-only session finishes with the following three area estimates:
| Area | Estimate θ̂ | Posterior SD s | Raw precision | Final weight |
|---|---|---|---|---|
| Verbal reasoning | 0.40 | 0.55 | 2.986 | 0.277 |
| Visual–spatial reasoning | 0.70 | 0.50 | 3.541 | 0.328 |
| Fluid reasoning | 0.55 | 0.45 | 4.257 | 0.395 |
The bounded precision weights give:
The composite uncertainty is:
The experimental IQ-style reading is:
For Standard mode with a standard-quality session:
The normal-model percentile is approximately Φ(0.558)=0.711; it is not displayed as a population percentile.
The report should be read in this order:
The program does not use the historical ratio formula:
Age currently selects either a youth or adult question form. It does not mechanically increase or decrease the result.
A later formal age-normed model would first locate the person within the calibrated distribution for the same age and administration conditions:
That step requires a representative sample, age-specific or continuously modelled norms, stable item parameters, and evidence that the test behaves fairly across relevant groups.
The current model is transparent and internally coherent, but several parts remain provisional.
ACRA generates questions, checks them before use, maintains separate model distributions, and adapts difficulty and question family. Its model assumptions and scores require empirical testing. The experimental reading is calculated from the model estimate:
The model’s main strength is that it separates question difficulty, cognitive area, response accuracy, response time and session quality. Its main limitation is equally clear: the scale is not yet based on representative age norms or empirical item calibration. For that reason, the current output should be treated as a broad project estimate and a basis for further research, not as a formal psychological IQ result.