MindOS Learning Manual: Probabilistic-Reasoning State | A High Percentage Can Still Answer the Wrong Probability Question

Wait, What?

“The test is 90% accurate” does not tell you there is a 90% chance the person has the condition after a positive result.

The percentage is real.

The conclusion can still answer the wrong question.

Probabilistic reasoning often fails not because the learner cannot multiply or divide, but because several different probabilities are silently treated as interchangeable.

the probability of evidence if a hypothesis is true;
the probability the hypothesis is true given the evidence;
the base rate before new evidence arrives;
the chance of two events occurring together;
the chance of at least one event occurring;
the probability conditional on a selected subgroup.

MindOS therefore asks:

What exact probability is being asked for, what population does it live inside, and what information is allowed to update it?

Quick Answer

Probabilistic-Reasoning State is the learner operation of identifying the target uncertainty, defining the reference population, preserving base rates, distinguishing conditional from inverse probabilities, representing nested sets or frequencies clearly, updating with new evidence, and checking whether the final probability answers the original question.

The RFE is:

Can the learner represent uncertainty so that the denominator, conditioning information and comparison set remain visible enough to prevent a numerically correct answer to the wrong question?

Owned Learning Operation

PROBABILISTIC-REASONING STATE = name target probability → define reference class → preserve base rate → translate information into compatible representation → distinguish P(A|B) from P(B|A) → update using evidence → verify denominator → express uncertainty → change representation → transfer.

This is not Bolt measurement calibration. Bolt asks what a score, estimate or statistical result justifies believing about performance. MindOS owns the learner’s reasoning operation while studying probability and uncertainty.

It is also not Causal-Reasoning State. Probability can describe uncertainty without establishing why an outcome occurs. A highly predictive association may remain non-causal.

And this is not simply a Mathematics content page. The operation appears in Science experiments, health claims, risk communication, source evaluation, everyday decisions and statistical arguments.

The Denominator Is Part of the Meaning

Consider:

90% of students who passed attended revision;
90% of students who attended revision passed.

These statements may sound similar. They condition on different groups.

The first asks:

Among PASSERS, how many attended revision?

The second asks:

Among REVISION ATTENDEES, how many passed?

Changing the denominator changes the probability question.

Six Probabilistic-Reasoning Failure States

1. Base-Rate Neglect

The learner focuses on the new evidence and ignores how common the target outcome was before that evidence arrived.

2. Inverse Fallacy

The learner treats P(evidence | hypothesis) as if it were P(hypothesis | evidence).

3. Denominator Drift

The learner begins with one reference group and silently switches to another during the calculation.

4. Percentage Without Population

A percentage is reported without enough information to know what population or time period it describes.

5. Representation Dependence

The learner can solve the problem only when it appears as a familiar probability tree or frequency table.

6. Precision Theatre

The learner gives 73.42% even when the data or assumptions do not justify that level of certainty.

Natural Frequencies: Make the Nested Sets Visible

Suppose a condition affects 1 in 100 people.

A test detects 90 of every 100 people who truly have the condition, but 5 of every 100 people without the condition also test positive.

Percentages can be hard to integrate mentally.

Translate to a population of 10,000:

100 have the condition;
about 90 of those test positive;
9,900 do not have the condition;
about 495 of those also test positive;
total positive tests ≈ 585;
positive tests that are true positives ≈ 90.

The learner can now see that the target probability is approximately:

90 / 585 ≈ 15%

The point is not the medical example itself. The point is representational: natural frequencies can make nested sets and denominators more visible.

The MindOS Probabilistic-Reasoning Protocol

Step 1 — Write the Target Probability in Words

Before calculation, complete:

I am trying to find the probability of ______ among ______.

The second blank is often the denominator.

Step 2 — Define the Reference Population

Who or what could possibly be counted?

Do not calculate until the universe of cases is clear.

Step 3 — Preserve the Base Rate

Ask what was known before the new evidence arrived.

A rare event can remain uncommon even after moderately strong evidence points toward it.

Step 4 — Convert to a Useful Representation

Possible representations:

natural frequencies;
two-way table;
tree diagram;
double tree;
unit square;
set diagram;
equation.

The best representation is the one that preserves the nested structure with the least unnecessary transformation.

Step 5 — Label Every Conditional Direction

Write:

P(A | B) = probability of A GIVEN B
P(B | A) = probability of B GIVEN A

Do not allow them to remain visually similar without verbal interpretation.

Step 6 — Update With Evidence

Determine which cases remain possible after the evidence.

Then calculate within that updated set.

Step 7 — Audit the Answer

What is the denominator?
Did I answer the requested conditional direction?
Did I preserve the base rate?
Is the result between 0 and 1 or 0% and 100%?
Does the answer change sensibly if the base rate changes?
Have I expressed more precision than the inputs deserve?

Step 8 — Change the Representation

After solving with a frequency table, solve a related case with percentages or a tree.

If the reasoning vanishes when the representation changes, the scaffold may be carrying too much of the operation.

Worked Example: Mathematics

In a school, 30% of students take a particular elective. Of those students, 40% join a competition. Of students who do not take the elective, 10% join the competition.

Question: A randomly selected competition participant is chosen. What is the probability that the student takes the elective?

A learner who calculates 40% has answered P(competition | elective), not P(elective | competition).

Using 1,000 students:

300 take the elective → 120 compete;
700 do not → 70 compete;
190 compete in total;
120 of the 190 competitors take the elective.

So the target probability is 120/190, about 63%.

The important learning object is the denominator switch, not the arithmetic.

Worked Example: Science

A genetics question gives the probability of inheriting a trait under particular parental genotypes.

The learner must distinguish:

probability of genotype;
probability of phenotype given genotype;
probability of genotype given observed phenotype.

Those are different conditional objects. A Punnett square can support the reasoning, but the learner should later be able to explain which population each fraction counts.

Worked Example: English / Media Literacy

A headline says:

“80% of top performers use Strategy X.”

This does not tell us what proportion of Strategy X users become top performers.

The learner asks for the missing denominator and base rate before accepting the implied recommendation.

How Do We Know?

A 2017 meta-analysis in Psychological Bulletin reviewed 35 articles representing 226 performance estimates on Bayesian reasoning. Across two decades of research, people generally performed better when relevant statistical information was represented as natural frequencies rather than conditional probabilities. The review also found important moderators: visual aids and less computationally complex representations improved performance in both frequency and probability formats.

A 2025 study in Learning and Instruction compared four Bayesian-reasoning training courses with a control group in 515 law and medical students. All four training conditions improved performance relative to the no-training condition. A frequency-based “double tree” produced the strongest gains in that study, increasing correct performance from about 13% at pre-test to about 70% at post-test. Prior mathematical achievement interacted with some training formats, while the double-tree condition was comparatively robust across achievement levels.

These findings support instruction in representation and conditional structure. They do not show that one diagram should become a universal probability method.

Evidence Boundary

Much of the strongest evidence concerns Bayesian inference problems, particularly applied risk contexts. Probabilistic reasoning is broader than Bayesian reasoning, so the evidence should not be stretched to every probability topic.

Natural frequencies are often helpful, but the 2017 meta-analysis shows that representation effects have moderators. Visual structure, computational complexity, expertise and other design features matter.

The 2025 training study involved university students in law and medicine. Its results do not automatically establish the same effect sizes for Primary or Secondary students.

And a learner who performs correctly only with one trained diagram may have learned the representation rather than the underlying conditional reasoning.

The safe educational inference is:

Making base rates and nested sets visible—often through natural frequencies or well-designed visual representations—can substantially improve Bayesian reasoning, but durable learning requires the learner to preserve the conditional structure when the representation changes.

What This Does Not Prove

It does not prove that people should always use frequencies instead of probabilities.
It does not prove that a correct Bayesian calculation implies good causal reasoning.
It does not prove that uncertainty can always be reduced to one precise number.
It does not prove that one visualisation works equally well for every learner.
It does not prove that a learner who follows a trained template can recognise the same structure in a different context.

When Probabilistic Reasoning Is the Wrong Tool

When the task is deterministic and no uncertainty needs to be represented.
When the learner lacks the basic fraction, ratio or set knowledge needed to interpret the quantities.
When the central question is causal rather than probabilistic.
When the numerical inputs are themselves unreliable and Source Evaluation must come first.
When the learner needs to interpret a performance score or measurement uncertainty in an educational system—route to Bolt.
When the problem is simple enough that a complex Bayesian scaffold adds unnecessary load.

Scaffold Fade

Stage 1: tutor supplies a natural-frequency representation and labels the denominator.
Stage 2: learner converts percentages into frequencies independently.
Stage 3: learner chooses between table, tree, frequency or equation based on the problem.
Stage 4: learner solves the same structure under a different representation.
Stage 5: learner recognises conditional direction and base-rate structure directly, using external representations only when they genuinely improve reliability.

Immediate, Delayed and Transfer Checks

Target: can the learner state the requested probability in words?
Denominator: can the learner identify the correct reference group before calculating?
Inverse: can the learner distinguish P(A|B) from P(B|A)?
Base rate: can the learner predict how the answer should change if prevalence changes?
Delayed: can the learner reconstruct the nested sets later without the original diagram?
Transfer: can the learner solve an equivalent problem presented as prose, a table, a tree and percentages?

AI Boundary: AI Can Compute the Probability While Hiding the Denominator Error

AI can perform Bayesian calculations, draw trees and explain the result.

If the learner’s job is probabilistic reasoning, the tool should not choose the target conditional structure before the learner attempts it.

learner states the target probability in words;
learner identifies the reference group and base rate;
learner chooses a representation;
AI may challenge the denominator or provide one alternative representation;
learner checks the result;
AI closes;
learner solves a structurally equivalent new problem independently.

Correct computation from the tool is not automatic evidence that the learner understood which probability was being computed.

Teaching Guide for Parents, Tutors and Teachers

“Probability of what, among what?”
“What is the base population?”
“What did we know before the new evidence?”
“Which direction is the conditional statement going?”
“Can you turn the percentages into people out of 100 or 1,000?”
“What is your denominator now?”
“Would your answer change if the base rate changed?”
“Now solve the same structure without this diagram.”

MindOS Direction

If the learner cannot reason with ratios or fractions: repair the prerequisite mathematical representation first.

If the learner confuses association with cause: use Causal-Reasoning State.

If the numerical evidence comes from an unreliable source: use Source-Evaluation State.

If the learner can solve only the trained representation: use Representation State, Practice Variability and Transfer State.

If the probability is being used to interpret educational performance evidence: route the measurement claim to Bolt after the learner operation is complete.

MindOS rule: probability begins by preserving the question. Name what is uncertain, preserve the reference class, make the denominator visible, update only with relevant evidence, and prove the reasoning survives when the helpful representation changes.