How does superintelligence work when an AI must choose the right next step for different requests? A pupil asking for a revision sheet needs a different response from a teacher asking to share confidential student records. Even two harmless requests can need different next actions: one may be ready for drafting, while another lacks the year level or verified source.
A decision tree turns those differences into a sequence of explicit questions. Each answer sends the case down one branch. Used carefully, a decision tree can make an AI workflow understandable: why did it ask a question, verify a source, produce a draft or route the case for human review? The branches are only as good as the policy and evidence behind them, so a tree is not a substitute for permission or professional judgement.
This guide builds a fictional educational-resource desk with sixteen requests, five possible outcomes and twenty checkable statements. The cases contain no real pupils or confidential records; they are an invented exercise about decision logic, not measured AI performance. “How Superintelligence Works” is the name of our learning series, not a claim that today’s models have achieved artificial superintelligence.
1. What is a decision tree?
A decision tree is a branching structure. Start with a question, inspect its answer and follow the corresponding path. A later question may be asked only for the cases that reach that branch. The final position is often called a leaf and gives an outcome or proposed decision.
People use similar reasoning in everyday life: if it is raining, take an umbrella; if not, continue without one. In an AI-assisted application, the conditions need more precise definitions because a plausible guess about a missing fact can send a case down the wrong branch.
2. Why order matters
Imagine checking whether a requested learning resource contains personal student information before examining typography. That sequence prioritises privacy over formatting. If private information is involved, the system can route the request for an authorised review rather than producing a publicly shareable draft.
A different tree could ask formatting questions first, wasting work or overlooking a sensitive requirement. The order of checks should reflect which conditions block safe progress and which facts are cheap or useful to establish.
3. The difference between a tree and a checklist
A checklist often asks the same questions for every case. A decision tree can stop early when one answer already determines a valid next step. For example, a request containing private data may require review regardless of whether its grade level is known.
Stopping early is useful only when the shortcut is justified by the actual rule. A system must not ignore other requirements if the current branch does not decide the outcome. The branch should explain why further checks can or cannot be deferred.
4. A tree needs an authoritative policy
A language model might invent a plausible branching flow, but that does not make it the correct administrative procedure. For sensitive data, the institution’s actual permissions and handling rules must determine which actions are allowed.
Our teaching tree is fictional and deliberately conservative. It describes how to choose an initial response, not what any real school legally requires. An AI builder should treat this tree as a conceptual demonstration rather than operational permission.
5. The imaginary Cedar Resource Desk
Imagine a school help desk that receives requests for educational resources. The desk examines three fields. P indicates whether the request contains private student information; its possible states are YES, NO and UNKNOWN. G indicates whether the target year level is known; S indicates whether the intended source has been verified.
The desk’s response is never an external publication or sending action. PREPARE DRAFT means creating editable educational content only. Human review means a designated authorised process must assess the handling of personal information before any further action.
6. Freeze the order of the checks
The first branch checks P. If P = YES, the next step is PRIVACY REVIEW. If P = UNKNOWN, the next step is CHECK PRIVACY STATUS. Only when P = NO does the workflow proceed to the grade-level question.
Next, if G = NO, the action is ASK YEAR LEVEL. Otherwise, inspect S. If S = NO, the action is VERIFY SOURCE. If S = YES, the action is PREPARE DRAFT. This ordering defines the full toy workflow.
7. The five possible next steps
- PRIVACY REVIEW: confirmed private student information is involved; route for authorised handling.
- CHECK PRIVACY STATUS: privacy state is unknown and must not be guessed.
- ASK YEAR LEVEL: no private data, but intended school year is unspecified.
- VERIFY SOURCE: no private data and year level known, but the teaching source is unverified.
- PREPARE DRAFT: no private data, year level known and source verified.
- Not an outcome: published, emailed or externally submitted; the toy tree never authorises those actions.
8. The information available in each case
To make the logic inspectable, we record all three field values for every invented case, even though the decision tree may not need to inspect later fields after finding a decisive earlier branch.
Keeping the full fixture lets readers verify that the priority rule, not an arbitrary preference, determines the next step. The identifiers A to P refer only to fictional requests, not actual people.
9. Sixteen fictional requests
- A — P=YES, G=YES, S=YES: PRIVACY REVIEW.
- B — P=YES, G=NO, S=NO: PRIVACY REVIEW.
- C — P=YES, G=YES, S=NO: PRIVACY REVIEW.
- D — P=UNKNOWN, G=YES, S=YES: CHECK PRIVACY STATUS.
- E — P=UNKNOWN, G=NO, S=NO: CHECK PRIVACY STATUS.
- F — P=NO, G=NO, S=YES: ASK YEAR LEVEL.
- G — P=NO, G=NO, S=NO: ASK YEAR LEVEL.
- H — P=NO, G=YES, S=NO: VERIFY SOURCE.
- I — P=NO, G=YES, S=YES: PREPARE DRAFT.
- J — P=NO, G=YES, S=YES: PREPARE DRAFT.
- K — P=NO, G=YES, S=NO: VERIFY SOURCE.
- L — P=NO, G=NO, S=YES: ASK YEAR LEVEL.
- M — P=NO, G=YES, S=YES: PREPARE DRAFT.
- N — P=YES, G=NO, S=YES: PRIVACY REVIEW.
- O — P=NO, G=YES, S=NO: VERIFY SOURCE.
- P — P=NO, G=YES, S=YES: PREPARE DRAFT.
10. Count the leaf outcomes
Four requests require PRIVACY REVIEW: A, B, C and N. Two need privacy clarification: D and E. Three require the year level: F, G and L. Three need source verification: H, K and O. Four are ready for an editable draft: I, J, M and P.
The sum is 4 + 2 + 3 + 3 + 4 = 16. Every fictional request reaches one leaf, and no request is counted twice. The count is a consistency check on the tree and its reference cases.
11. Why A goes to privacy review immediately
Case A has a known year level and a verified source, but the privacy field is YES. The next step is PRIVACY REVIEW, not PREPARE DRAFT, because the first branch deliberately blocks the ordinary drafting flow for such requests.
This illustrates how a policy priority can override attractive later attributes. Good source information does not automatically authorise its use when the requested output may expose a student’s private record.
12. Why D does not guess
Case D has a known year level and verified source but an unknown privacy state. The model should not silently assume P = NO simply because everything else looks ready. The correct next step is to establish whether protected information is involved.
UNKNOWN is neither YES nor NO. For this policy, treating UNKNOWN as NO could prematurely enter the drafting branch. Treating it as confirmed YES could misstate the source. Clarification preserves the actual information state.
13. Why F needs only one question first
Case F has no private data and a verified source, but its intended school level is unknown. Asking for the year level is the minimal useful next step. Repeating a long questionnaire about the already verified source would add friction without resolving the blocker.
Once the user supplies the year level, the application can continue along the next branch using the unchanged privacy and source facts. This makes the decision tree practical for conversation as well as software.
14. Why H is different from F
H has a known year level but an unverified source. There is no reason to ask for the school year again. The right next step is to check the teaching resource against a reliable source before claiming it is an approved or factually grounded input.
An assistant that drafts confidently from an unverified source could be useful for a provisional brainstorming note, but that is a different workflow. Under the Cedar tree’s frozen policy, the current next step is VERIFY SOURCE.
15. Why I may be drafted but not sent
I has P = NO, G = YES and S = YES, so PREPARE DRAFT is the correct leaf. The word draft is essential. The tree grants permission to create proposed content under the toy procedure, not to publish it or contact students.
Execution is a separate boundary. A tool-using AI should not treat a logical label in the tree as proof that an external message was authorised or delivered.
16. A decision tree should be testable
Every branch needs cases that demonstrate what happens when the answer is YES, NO or UNKNOWN. If the tree only receives examples with complete positive data, developers will not discover how it treats uncertainty.
The sixteen cases include four privacy-positive requests, two privacy-unknown requests, missing year levels, unverified sources and fully ready drafts. This variation reveals more of the branch behaviour than repeating sixteen perfect requests.
17. Four-valued details are not always necessary
Some applications need more states such as NOT APPLICABLE, EXPIRED, CONFLICTED or PENDING APPROVAL. Our toy privacy field has three states only because that is enough to demonstrate the selected branches.
The system should define its state vocabulary from actual source and policy requirements. Adding arbitrary labels without a decision rule can make a tree appear sophisticated while leaving its actions unclear.
18. Why a tree is not inherently fair
A decision tree can consistently implement a flawed policy. If its training examples or rules treat one legitimate group unfairly, consistent branching does not repair the underlying discrimination.
Fairness evaluation needs appropriate domain evidence, scrutiny of features and consequences, and a way for affected people to challenge outcomes. A clean tree diagram is not proof that the underlying choices are equitable.
19. Twenty statements for an AI to classify
The fictional desk asks a classifier to label statements SUPPORTED, CONTRADICTED or NOT ESTABLISHED using the frozen branch rules. These are claims about the next step, not verified live actions.
The exercise tests whether a model preserves priority order and distinguishes a source-verification step from permission to publish.
- 01. “A goes to PRIVACY REVIEW.” — SUPPORTED.
- 02. “B goes directly to ASK YEAR LEVEL.” — CONTRADICTED.
- 03. “C goes to PRIVACY REVIEW even with an unverified source.” — SUPPORTED.
- 04. “D can safely be assumed to contain no private data.” — CONTRADICTED.
- 05. “E needs its privacy state established first.” — SUPPORTED.
- 06. “F needs the target year level.” — SUPPORTED.
- 07. “G goes directly to VERIFY SOURCE before the year level.” — CONTRADICTED.
- 08. “H goes to VERIFY SOURCE.” — SUPPORTED.
- 09. “I goes to PREPARE DRAFT.” — SUPPORTED.
- 10. “J goes to PREPARE DRAFT.” — SUPPORTED.
- 11. “K already has a verified source.” — CONTRADICTED.
- 12. “L needs the year level.” — SUPPORTED.
- 13. “M is ready for an editable draft.” — SUPPORTED.
- 14. “N can bypass the privacy branch because its grade is unknown.” — CONTRADICTED.
- 15. “O needs source verification.” — SUPPORTED.
- 16. “P is ready for an editable draft.” — SUPPORTED.
- 17. “Four cases route to privacy review.” — SUPPORTED.
- 18. “The desk automatically published four real documents.” — NOT ESTABLISHED.
- 19. “Exactly two cases have unknown privacy state.” — SUPPORTED.
- 20. “Every request with P=NO is immediately ready to draft.” — CONTRADICTED.
20. Check the scorecard
Supported cases are 01, 03, 05, 06, 08, 09, 10, 12, 13, 15, 16, 17 and 19: thirteen. Contradicted cases are 02, 04, 07, 11, 14 and 20: six. Statement 18 is not established: one. Thirteen plus six plus one equals twenty.
The publication claim is unverified because a branch returning PREPARE DRAFT does not prove that any external publishing operation occurred. A model’s plan is different from an executed and confirmed action.
21. The right branch depends on meaning
Suppose a prompt says “The worksheet includes a student nickname” but it is unclear whether that nickname identifies a real child. The privacy field may need further inspection or advice under the applicable rules. The model should not decide based solely on whether a name looks common.
Classification of sensitive information requires clear definitions and appropriate safeguards. The tree cannot replace those definitions or a qualified decision when the context is genuinely ambiguous.
22. A tree can be deterministic after facts are established
Once P, G and S are known and the branch policy is frozen, a simple computer program can calculate the next step reliably. The language model’s main challenge may be translating flexible requests into the right structured fields.
If the extracted fields are wrong, a perfectly implemented decision tree will still return the wrong branch. Systems therefore need both source extraction checks and a tested decision procedure.
23. A wrong condition can affect many cases
Imagine a programmer accidentally writes “if P is not YES, continue to drafting checks”. That sends P=UNKNOWN cases D and E into later branches instead of asking about privacy. The error affects every request with unknown privacy state.
This is a good reason to include explicit unknown cases in regression tests. Testing only YES and NO would not expose the accidental promotion of missing information into permission.
24. The difference between choosing and doing
A decision leaf can say VERIFY SOURCE, but no source has been verified until the tool or responsible reviewer actually checks it. A leaf can say PRIVACY REVIEW without showing that a privacy officer has made any determination.
A trustworthy AI reports both the intended next step and the confirmed progress. It should never write “review completed” merely because the decision tree chose a review branch.
25. When a rule changes
Suppose the institution later authorises a special drafting path for anonymised classroom examples. That would require a revised tree with a clearly defined new branch, effective date and approval record.
The assistant should not improvise the new route within an old policy. Versioning ensures that decisions can be traced to the actual procedure that applied when they were made.
26. How to draw a decision tree for learning
Write a question in a diamond or as a simple yes/no branch. Put the outcome labels at the ends. Trace each example from the start and record the actual answers to the questions it encountered.
Students can use the method for identifying shapes, choosing which arithmetic operation to apply or sorting science observations. The benefit comes from explicit logic, not from complicated diagram artwork.
27. A transfer exercise about Mathematics
Consider a classifier for a whole number n: if n is even, label EVEN; otherwise, if n is divisible by three, label ODD MULTIPLE OF THREE; otherwise label OTHER. Under this rule, six is EVEN even though it is also divisible by three.
The chosen priority produces one outcome. If the application needed all applicable mathematical properties, a multi-label classification method would be more appropriate than a first-match decision tree.
28. A transfer exercise about book choice
A parent wants a book within a strict budget and reading level. A decision process can first exclude books above the budget, then exclude books outside the intended difficulty range, and finally rank the remaining choices by topic interest.
That is a different policy from our Cedar resource workflow, but it uses the same general idea: test required conditions and preserve the user’s priorities before recommending an option.
29. What if a field is inconsistent?
Suppose a request is labelled P=NO in one form but a verified attachment shows personal information. This is a conflicting observation, not an ordinary missing value. A practical system should reconcile the evidence rather than blindly follow whichever field was read first.
An AI can help flag that conflict, but the final handling of sensitive information must be governed by the authorised application and responsible reviewer.
30. What decision trees cannot explain alone
A tree explains how defined features lead to a decision. It does not prove that the features were measured accurately, that the policy is appropriate or that the recommended next action succeeded.
A complete explanation includes the source values, rule version, chosen path, action permissions and any verified outcome. The visual tree is one component of an accountable process.
31. A classroom audit method
Hand pupils the sixteen request cards, with P, G and S visible. Give them the five-step tree and ask them to route each card. Then compare totals across the five outcomes and investigate any mismatch.
A useful extension is to hide one field on a card and ask whether the missing value changes the result. That shows why some unknowns require clarification immediately while others are irrelevant once a decisive earlier condition is established.
32. Evaluation for AI builders
Test every branch with normal cases, boundary cases and missing or conflicting inputs. Freeze the policy before evaluation. Record whether the model extracted the fields correctly, selected the correct leaf and complied with any external action restrictions.
Distinguish decision accuracy from task completion. An assistant may choose the correct VERIFY SOURCE branch but fail to obtain a valid source. A successful decision does not equal a completed operation.
33. What a transparent answer sounds like
For a case like H, a clear response is “The request has no private student information and specifies a year level, but the source has not been verified. The next step is to verify the source before preparing a draft under this policy.”
That is more useful than “I cannot help” or “Everything is ready.” It identifies the actual blocker and what would resolve it, without inventing an approval or unnecessary additional tasks.
Frequently asked questions
What is a decision tree? A structured sequence of questions directing cases to different outcomes.
Does the order of questions matter? Yes, particularly when one condition determines whether later branches may be considered.
Can a tree handle unknown values? Yes, if its rules explicitly define what happens when information is missing.
Does a decision tree guarantee fairness? No. The underlying policy, features and evidence still need evaluation.
Why is a draft different from publishing? Drafting creates proposed content; publishing is an external action requiring authority and confirmation.
Can a language model use a deterministic tree? Yes. The model may interpret requests while software evaluates the frozen branch logic.
What if the policy changes? Update the authoritative tree with a versioned, approved change and test the new behaviour.
The central lesson
An effective AI assistant does not treat every request as a cue to produce the same kind of answer. It identifies the decisive conditions and chooses a next step that fits the known facts, missing details and action boundaries.
Decision trees make those choices visible. With checked inputs and controlled permissions, they help turn a fluent assistant into an easier-to-audit workflow. Without those checks, a neat branch diagram can still route a request in the wrong direction.
Connected learning and source references
- How Super Intelligence Works — main curriculum.
- Necessary and Sufficient Conditions — interpreting prerequisite rules.
- Goals, Constraints and Trade-Offs — choosing a feasible path.
- Quoted Text and Instructions — which actions are actually authorised.
- Scikit-learn: Decision Trees — mathematical and software decision-tree methods.
- NIST AI Risk Management Framework — accountable and evaluated AI decisions.