How does superintelligence work when an AI reads two different versions of the same rule? A school may have an old notice saying that two workshops are required and a newer notice saying that three are required. Both documents can be genuine. The challenge is deciding which rule applied to which application, on which date. An assistant that simply selects the latest-looking paragraph may reject a pupil who correctly met an earlier policy.
The mechanism is versioned-rule reasoning: keeping a policy’s wording, effective dates, authority and affected cases together. This is essential when AI helps interpret educational programmes, eligibility notices, timetables, administrative forms and other information that changes over time. An accurate calculation under the wrong policy version is still the wrong answer.
This guide uses a fictional workshop-certificate scheme and twelve fully specified application cases. We will see how two successive policies can yield different decisions, how a cutoff date affects an edge case and what evidence the AI needs before claiming that a policy is currently in force. Every programme and result is invented for learning, not actual institutional guidance or a measured benchmark of any AI product. “How Superintelligence Works” is an educational series title, not a claim that present systems have achieved artificial superintelligence.
1. A policy is a rule with a history
Suppose the first edition of a programme guide specifies a minimum of two completed workshops. A revised edition raises that requirement to three. It is tempting to describe the second edition as the one true answer. But a pupil asking about an application made before the change needs the rule applicable to the earlier application.
A reliable AI answer therefore retains the document date and the effective date. A policy’s publication date, the date it takes effect and the date a person performs an action may be different. Treating them as interchangeable can reverse decisions.
2. Why “latest” is not enough
An updated webpage may describe today’s application rules but say nothing about applications already filed. A dated archive may remain the correct source for a historical request. Sometimes transitional provisions preserve an earlier rule for existing participants even after a new policy begins.
The correct selection depends on authoritative wording, not merely on which file has the newest timestamp. A system should not invent retroactive effects or assumed exemptions where the source has not supplied them.
3. The effective date is not decoration
An effective date identifies when a rule begins governing the cases within its stated scope. If the revised policy starts on 15 October, an application submitted on 14 October may be evaluated differently from one submitted on 15 October. That is the point of a boundary test.
“After 15 October” and “on or after 15 October” do not have identical interpretations. A formal application system needs a precise cutoff, applicable time zone and a definition of which event date controls eligibility.
4. The controlling event must be specified
Does the submission date determine the policy? Or the assessment date, enrolment date, completion date or date an administrator reviews the file? Different programmes may choose different events. A model cannot settle that choice from the concept of versioning alone.
In this teaching case, we explicitly define the submission date as the controlling event. That removes a major ambiguity. A real applicant should verify the rule against the current official document rather than borrow our invented convention.
5. The fictional Cedar Workshop Certificate
Imagine Cedar School creating a simple certificate programme with two consecutive rule editions. Both are assumed to be genuine and complete. There are no special waivers, appeals or grandfathering clauses in the teaching fixture.
Under Version 1, applications submitted through 14 October 2026 inclusive need at least two completed workshops and a test score of at least 70 out of 100. Under Version 2, applications submitted on or after 15 October 2026 need at least three workshops and a score of at least 75.
6. The frozen two-version policy
- V1 effective range: through 14 October 2026, inclusive, for the applications in this packet.
- V1 workshop threshold: at least 2 completed workshops.
- V1 score threshold: at least 70 of 100.
- V2 effective range: 15 October 2026 onward.
- V2 workshop threshold: at least 3 completed workshops.
- V2 score threshold: at least 75 of 100.
- Controlling event: date of submission, not review or certificate issuance.
- Decision: QUALIFIES only if both applicable thresholds are met.
- No additional exceptions: this is a complete fictional rule set.
7. Twelve dated applications
The test cases contain a submission date, workshop count and examination score. Each must be evaluated using the policy edition selected by its submission date. We intentionally include several applicants close to the change date and close to threshold values.
Every numeric value is part of a made-up example. The cases are not anonymised real pupils, and the labels below are teaching answers rather than official award decisions.
- A — 10 Oct; 2 workshops; score 72: V1, QUALIFIES.
- B — 10 Oct; 1 workshop; score 90: V1, DOES NOT QUALIFY.
- C — 10 Oct; 2 workshops; score 68: V1, DOES NOT QUALIFY.
- D — 14 Oct; 2 workshops; score 70: V1, QUALIFIES.
- E — 14 Oct; 3 workshops; score 67: V1, DOES NOT QUALIFY.
- F — 15 Oct; 2 workshops; score 90: V2, DOES NOT QUALIFY.
- G — 15 Oct; 3 workshops; score 75: V2, QUALIFIES.
- H — 16 Oct; 3 workshops; score 74: V2, DOES NOT QUALIFY.
- I — 16 Oct; 4 workshops; score 80: V2, QUALIFIES.
- J — 16 Oct; 2 workshops; score 75: V2, DOES NOT QUALIFY.
- K — 20 Oct; 3 workshops; score 76: V2, QUALIFIES.
- L — 20 Oct; 3 workshops; score 75: V2, QUALIFIES.
8. Check how many qualified
V1 covers A through E. Only A and D meet both V1 conditions, so that portion contains two qualifying applications and three that do not qualify. V2 covers F through L. G, I, K and L qualify, while F, H and J do not.
Across all twelve applications, six qualify and six do not. The total is a property of the specified toy packet, not a statement about how many real candidates would pass under similar rules.
9. The day before the change
Case D is submitted on 14 October with exactly two workshops and a score of exactly 70. Both V1 boundaries are inclusive. Therefore D qualifies. Applying V2 to D would incorrectly reject the applicant for having only two workshops and a score below 75.
The important lesson is that a model may know the latest requirements perfectly and still be unfair to a historical case if it attaches those requirements to the wrong date.
10. The first day of the new rule
Case F arrives on 15 October with two workshops and a score of 90. The score exceeds the V2 threshold, but the workshop requirement is not met. F does not qualify under the complete V2 policy.
Case G arrives on the same day with three workshops and a score of 75. Both conditions meet their inclusive boundaries, so G qualifies. The two cases show that a higher examination score cannot substitute for a required workshop count.
11. The boundary score of seventy-five
Under V2, a score of exactly 75 satisfies the minimum. G and L meet that threshold with three workshops. H has three workshops but a score of 74, so H does not qualify.
A model that interprets “at least 75” as “greater than 75” would wrongly reject G and L. Version selection and precise inequality interpretation are both necessary for a correct result.
12. A source can be authentic and outdated for the task
Imagine an old policy PDF still hosted on an official website. The document may truthfully record V1 rules without being the appropriate instruction for a new submission in late October. Conversely, a current V2 webpage may not accurately describe the rules governing D’s 14 October submission.
A trustworthy AI should cite the selected version and its applicability, not attach a generic official link and assume that settles the historical question.
13. What if an older rule was amended retroactively?
Our fictional fixture specifies no retroactive changes. In a real institution, a correction or transitional clause may alter how a new policy affects earlier applications. A model should inspect the text rather than assume either full retroactivity or complete protection of old cases.
Policies can also have distinct effective dates for different groups. If the source is ambiguous, a qualified administrator may need to decide. An AI-generated interpretation is not by itself an authoritative ruling.
14. Time zone matters at finer cutoffs
Here all cases use whole calendar dates, making the decision simple. If the change took effect at midnight in a specified time zone, a submission received in another region could require conversion to the official local clock.
A date-only rule and a timestamp-based rule are not automatically interchangeable. Software should preserve the authority’s precise effective boundary and the original submission timestamp where available.
15. Why a revised document needs a version identifier
“Certificate Guide — Final” is a weak document identity when an administrator has produced several files named Final. A version number, publication date, issuer and effective range give a reader a better way to distinguish them.
A model that merely says “the guide states” leaves the user unable to tell which edition it consulted. An auditable answer should identify V1 or V2 and the rule that follows from it.
16. What about missing submission dates?
Suppose an application records three workshops and a score of 76 but no submission date. Under our two editions, those values satisfy both sets of thresholds. A conclusion of qualification is possible within this narrowly frozen fixture, but the applicable version still cannot be identified without the date.
Now suppose the application has two workshops and a score of 72 with no date. It qualifies under V1 and fails V2. That outcome cannot be settled until the controlling date is recovered. Missing context matters when it changes the decision.
17. Twenty claims to classify
The fictional assessment asks the AI to label statements SUPPORTED, CONTRADICTED or NOT ESTABLISHED using only the frozen policies and application records. These labels test both arithmetic and interpretation of dates.
Claims about issuance, appeals or real-world certificate delivery require separate evidence. No application row proves that a certificate was physically generated or sent.
- 01. “V1 requires at least two workshops.” — SUPPORTED.
- 02. “V2 requires at least three workshops.” — SUPPORTED.
- 03. “The V1 score threshold is seventy.” — SUPPORTED.
- 04. “The V2 score threshold is seventy-five.” — SUPPORTED.
- 05. “D was submitted under V1.” — SUPPORTED.
- 06. “D fails because V2 requires three workshops.” — CONTRADICTED.
- 07. “F qualifies because ninety is a high score.” — CONTRADICTED.
- 08. “G qualifies under V2.” — SUPPORTED.
- 09. “H qualifies under V2.” — CONTRADICTED.
- 10. “L qualifies at exactly the V2 score threshold.” — SUPPORTED.
- 11. “Six of the twelve applications qualify.” — SUPPORTED.
- 12. “All applications use the same rule version.” — CONTRADICTED.
- 13. “Both thresholds apply to every application regardless of date.” — CONTRADICTED.
- 14. “Submission date controls version selection in this scheme.” — SUPPORTED.
- 15. “Every application on 14 October uses V2.” — CONTRADICTED.
- 16. “Applications on 15 October use V2.” — SUPPORTED.
- 17. “Every qualifying applicant received a physical certificate.” — NOT ESTABLISHED.
- 18. “Case C failed the V1 score threshold.” — SUPPORTED.
- 19. “Case I has four workshops and scores eighty.” — SUPPORTED.
- 20. “The scheme included an undocumented appeal exception.” — CONTRADICTED by the complete no-exceptions fixture.
18. Reconcile the labels
Supported statements are 01, 02, 03, 04, 05, 08, 10, 11, 14, 16, 18 and 19: twelve. Contradicted statements are 06, 07, 09, 12, 13, 15 and 20: seven. Statement 17 is not established: one. Twelve plus seven plus one equals twenty.
The unknown certificate-delivery claim is important. Eligibility under a rule does not prove that an external system executed the award. The AI should ask for an issuance record rather than treating a calculation as a receipt.
19. Versioning and fact-checking work together
A model can extract the right number from the wrong edition or the wrong number from the right edition. Those are two failure modes. First verify which policy applies, then verify the values and the individual’s record.
A short checklist makes them visible: identity of policy, effective dates, controlling event, relevant exceptions, threshold values and verified outcome. Omitting one can change a decision.
20. Policy versions are not general evidence of improvement
A stricter V2 threshold does not prove that the programme became better. It only shows that a rule changed. Assessing learning quality would require evidence about instruction, participants and outcomes.
An assistant should not turn the date of a revision into a causal story. Policies can change for many reasons, and the rule document alone may not record those reasons.
21. Comparing populations across versions needs caution
If V1 and V2 applicants have different preparation or backgrounds, comparing their pass rates directly may say little about which policy is more effective. Different standards also mean that the word qualifies has different operational requirements across the groups.
Before comparing outcome percentages across periods, report the definition used in each period and whether the populations are comparable. Versioned data must preserve meaning as well as numbers.
22. A model can create a useful decision table
For each application, list the controlling date, policy version, workshop threshold, score threshold, observed workshop count, observed score and decision. This allows a person to audit a result without guessing which text was applied.
If the user changes the application date in a follow-up, recompute the relevant case using the new effective version while retaining every other unchanged fact. The procedure should not silently replace the source document or modify unrelated records.
23. Change records should record who approved them
A rule update should be linked to its legitimate issuing authority and an approval or publication record where the workflow requires one. A file that looks official is not guaranteed to be an authentic current policy.
An AI retrieval system can locate candidate sources, but it must distinguish a signed or authoritative rule from a draft, commentary or cached copy. The language model’s confident interpretation cannot manufacture authority.
24. The difference between superseded and false
An older policy can be superseded without ever having been false. V1 genuinely governs the toy submissions from the earlier period. It should not be labelled an error simply because V2 exists.
This also matters in science textbooks, government notices and software documentation. Historical truth, present applicability and factual correctness are three different questions.
25. What to do if two sources disagree
First ask whether they describe different effective periods. If they do, the apparent contradiction may disappear. If both claim authority over the same time and population, the system must investigate whether one is an amendment, a correction, an unofficial copy or an actual unresolved conflict.
Choosing whichever document appears first in search results is not a safe resolution method. Keep provenance and the exact claim visible.
26. A fictional baseline error pattern
Imagine an invented AI that checks all twelve applications using V2. It would incorrectly reject A and D, which meet V1 requirements. It could also misdescribe the earlier policy and confuse the effect of 14 versus 15 October.
These are teaching errors, not the results of a measured benchmark. Their causes involve temporal scope and version selection rather than difficulty with basic addition. The correct repair begins with the policy ledger.
27. A revised fictional workflow
An improved approach first selects the policy edition from the controlling event date, then evaluates all required fields and preserves the source of the decision. The system checks edge cases and refuses to invent retroactivity or special exceptions.
If a fictional second pass gets every case right, it shows that this exercise was handled correctly, not that a live AI can interpret every changing legal or educational policy. Real cases need current authoritative sources and suitable professional review.
28. Transfer exercise with a textbook edition
A student uses a page reference from an older textbook edition. A newer edition may rearrange chapters. The older page number was not necessarily wrong; it may simply refer to another edition.
The useful next step is to confirm the edition before claiming that the citation is broken. Publication date and edition identity can be as important as the page number.
29. Transfer exercise with a submission cutoff
Suppose a fictional school raises the minimum project score from 60 to 65 for submissions made on or after 1 November. A pupil submits on 31 October with 62 and another on 1 November with 62. Under an otherwise identical rule, the earlier submission passes the score threshold and the later one does not.
This difference is not a claim about the pupils’ ability; it is a consequence of applying two time-bounded policies correctly. The example reinforces why a date is part of an eligibility calculation.
30. When an AI should ask instead of guess
Ask when the controlling event date is absent and the decision would change between editions. Ask when the effective-date text is ambiguous or when an exception may govern the applicant. Ask for authoritative confirmation if the retrieved pages conflict on a consequential rule.
Do not ask for unrelated details that cannot affect the decision. A focused question about the submission date may be enough to resolve the entire problem.
31. How to evaluate a version-aware AI
Prepare a test set with dates immediately before and after changes, equivalent applicants across two versions, missing dates, retroactive clauses, corrections and rescinded notices. Include source links with different publication and effective dates.
Score version selection independently from condition checking. A model that chooses the right edition but reverses “at least” is still wrong; a model that calculates perfectly under a superseded rule is wrong for another reason.
32. A public-facing explanation should be explicit
Instead of “You are not eligible because the current rule requires three workshops,” write “Under the policy applicable to applications submitted on 14 October 2026, two workshops and a score of at least 70 were required. Your toy case meets those conditions.”
That explanation reveals the source, controlling event and scope. It gives the reader a fair way to challenge the result rather than presenting a policy change as an invisible decision.
Frequently asked questions
Is the newest policy always the right one? Not for every historical case. Effective dates and transitional clauses control applicability.
What is an effective date? The date or moment from which a rule governs the cases within its defined scope.
Can two different policies both be valid? Yes, for different periods or populations.
Why must the event date be identified? Submission, enrolment, completion and review dates can differ, and only the authoritative rule determines which controls.
Does eligibility prove that a certificate was issued? No. Issuance needs a separate verified action or record.
What if the source is a draft? It should not automatically be treated as the approved rule.
Can a policy change apply to older applications? Sometimes, if an authoritative provision actually says so; do not invent retroactive effect.
The central lesson
Rules are not timeless just because they are written in confident language. A reliable AI explanation keeps the rule’s authority, version, effective range and controlling event attached to the result. That lets a historical decision remain correct even when today’s rule is different.
When a result looks inconsistent, ask which rule edition was applied and when the relevant action happened. The answer may lie in time and provenance rather than in the arithmetic.
Connected learning and standards
- How Super Intelligence Works — educational curriculum.
- Relative Dates — interpreting calendar references and cutoffs.
- Necessary and Sufficient Conditions — reading qualification rules.
- Boundary Values — exact inclusive and exclusive thresholds.
- W3C PROV-O — preserving document and activity provenance.
- W3C DCAT Version 3 — describing dataset versions and catalogues.
- NIST AI Risk Management Framework — evaluation and oversight of AI applications.