How Superintelligence Works | Agent Material Change Detection

A document can change by only a few characters and still require a different decision. Moving an exam date, altering a recipient address or changing an eligibility threshold may be more consequential than rewriting several paragraphs of harmless explanation. Agent material change detection is the practice of identifying what changed, whether the change affects the user’s task or earlier approval, and which source and action must be reverified before publication or sending.

The practical method is to keep stable document identities and versioned sources, compute exact diffs, classify changes by their real business meaning, compare approved and proposed content, preserve effective dates and route material differences to the right reviewer. This guide uses a fictional one-hundred-change ledger, independently checked arithmetic and a changed-context school notice exercise. The numbers are original teaching examples, not real eduKate edits or evidence that artificial superintelligence already exists.

Follow How Super Intelligence Works. The Concurrent Document Edits chapter prevents lost updates, and Dual Control explains independent approval. This chapter addresses what happens after an apparently small revision: when must a previously valid factual check or human authorisation be reconsidered?

Choose a reading route

Editing an education notice? Begin with effective dates, numeric conditions and exact source revisions.
Approving an agent? Study version-bound consent, recipient changes and final saved results.
Learning Mathematics? Recalculate the hundred-change ledger and classify the deadline example.
Building an editorial pipeline? Follow diffs, materiality tests, source lineage and reapproval rules.

1. A small textual edit can be a large semantic change
2. Compare revisions under one stable document identity
3. Machine diffs are useful but incomplete evidence
4. Classify materiality against the user’s actual purpose
5. Human approval must bind to the reviewed version
6. Track changes across sources, citations and generated text
7. Effective-date shifts change truth conditions
8. Numerical changes deserve type and unit checks
9. An invented hundred-change audit
10. Distinguish textual magnitude from user consequence
11. Preserve source lineage and review evidence
12. Concurrent editors can create a different kind of conflict
13. A changed-context school eligibility exercise
14. False negatives and false positives require different repairs
15. Specific negative tests can protect release gates
16. A small editorial team’s materiality runbook
Frequently asked questions
Technical and governance sources
Further reading

1. A small textual edit can be a large semantic change

Suppose a fictional school notice says that subject-selection applications close on 12 November. Replacing 12 with 21 changes only two characters, but it can change a family’s plans. By contrast, rewriting three explanatory paragraphs to improve grammar may preserve the same deadline, school, cohort and eligibility rule. Character-count difference is therefore an unsuitable proxy for business importance.

Begin by defining the claims and consequences a user relies upon: dates, requirements, fees, destinations, protected data, source references and permitted actions. A textual diff helps locate changes, while a task-specific classifier assesses whether each change could alter the reader’s decision or the operation the agent is allowed to perform. Human review may remain necessary where meaning is ambiguous.

An agent should not declare that a changed number is harmless merely because the source looks authoritative or the edit was made by a trusted colleague. Even authorised editors can make mistakes. The source document’s identity, effective time and the context in which the change applies matter. A rule could be current for one intake but historical for another.

Do not overreact to every byte difference either. Normalised whitespace, an added illustration caption or formatting changes can sometimes be treated under an explicit nonmaterial policy. The key is that the allowed tolerance must be defined before release by an accountable owner. The model cannot invent a new exemption when it wants to publish quickly.

Return to contents or continue with Compare revisions under one stable document identity.

2. Compare revisions under one stable document identity

A robust workflow stores a stable document ID and a revision or content digest. Filenames and titles can change, and two different documents can share the same title. A stable identity links the current candidate to the original approved source and target. Without it, a search result about the same subject may be mistaken for the same document.

Record which revision was reviewed. An editor might have approved revision 7 of a Mathematics syllabus guide. If the draft later becomes revision 8, the system should compare exact relevant content rather than simply reuse an Approved boolean. That avoids a silent change to the effective audience, a numerical condition or a cited source.

A cryptographic digest can help detect byte changes when computed and stored by a trusted service. It does not explain whether a change matters, prove that the document is factually correct or provide human consent for an external action. Use a digest for integrity and exact version binding; use source review and permission checks for their separate requirements.

The document’s effective date can differ from its file modification timestamp. A rule issued on Monday might take effect in January, and a later typo correction might not change the applicable year. Keep issue, modification and effective times as distinct facts. A model that replaces them all with today’s date can transform a valid historical source into misleading current guidance.

Return to contents or continue with Machine diffs are useful but incomplete evidence.

3. Machine diffs are useful but incomplete evidence

Version-control systems such as Git can display textual changes between commits. A diff identifies inserted, removed or changed lines, making a large document’s altered regions easier to inspect. It does not by itself decide whether a modified heading creates a different legal or educational requirement. The reviewer must understand what the words mean in the task’s specific context.

Move and reorder operations deserve attention. An instruction moved under another heading may now have a narrower or wider scope even when its exact words are unchanged. A paragraph copied twice can create contradictory-looking statements. A purely line-based review might overstate harmless moves or understate meaning changes caused by section context.

Models can suggest a semantic difference summary, but the candidate summary should link to exact old and new passages. If the model says nothing material changed, a reviewer should still be able to inspect the underlying versions, especially around dates, amounts, conditions and permissions. A generated assertion is a hypothesis requiring verification, not a release gate.

Use a consistent representation for structured documents. A spreadsheet formula can change through cell references without the displayed value immediately revealing it. A PDF can contain an updated table image rather than selectable text. Different formats need appropriate parsers and independent checks before one text-diff tool can make a reliable change classification.

Return to contents or continue with Classify materiality against the user’s actual purpose.

4. Classify materiality against the user’s actual purpose

A source change is material when it can affect a decision, permission, side effect or claim the task depends upon under the organisation’s defined policy. In a school notice, dates and eligibility thresholds are obvious candidates. In a private email, the recipient, attachment, amount or main request may be material. In a technical operation, changing a target account or permission scope is particularly consequential.

Some changes are clearly safe to approve within a bounded formatting rule. For example, replacing repeated spaces or correcting an unambiguous typographical error in a harmless explanation may not change the user’s intended effect. But an apparently small typo in a student ID or recipient email can redirect data. Classify by field semantics, not by which characters were changed.

A practical schema can mark fields as identity, authority, numeric condition, temporal scope, confidential content, external destination, evidence citation or presentation. Materiality then asks what changed in each field and whether the approved task contract permits it without re-review. This structure improves consistency across documents and prevents arbitrary model decisions about what seems important.

When meaning cannot be established safely, retain Unresolved and route to a qualified person. A school requirement might use an unfamiliar abbreviation or reference another issuer. The assistant should not guess that the edit is immaterial because the original and revised sentences are highly similar. Safe abstention can protect readers from consequential uncertainty.

Return to contents or continue with Human approval must bind to the reviewed version.

5. Human approval must bind to the reviewed version

GitHub’s branch protection documentation explains that a repository can dismiss stale pull-request approvals when new commits change the reviewed diff. This feature recognizes that a human approved a particular set of changes, not every future edit. The precise GitHub options are provider-specific, but the underlying design principle translates to agent-generated articles and messages.

If an editor approved a post with one deadline and the model then changes that deadline, the system should not publish under the old grant unless an explicitly defined policy allows that alteration. The authorisation should carry the reviewed content revision and relevant resource. A new approval may be required even when the changed page still parses as valid HTML.

Different approval scopes can coexist. An editor might approve the factual claim while a marketing owner approves the public release, and the user might prohibit subscriber emails. A changed recipient or distribution method can invalidate the action grant without changing the article’s body. Material change detection must inspect both content and intended side effects.

Do not use a higher-privilege service account to bypass an expired or invalidated review. The protected publishing service should check the actual current grant at the moment of execution. A model’s friendly explanation that the new version is nearly identical cannot override the permission boundary.

Return to contents or continue with Track changes across sources, citations and generated text.

6. Track changes across sources, citations and generated text

A generated education article may cite an official syllabus and paraphrase a condition. If the official source changes, the article can become stale without anyone editing its text. Detecting material changes therefore requires tracking source dependencies, not just monitoring the article’s own revision history. A citation should point to an identifiable source and applicable version.

A new official notice might replace an old eligibility rule. The source change should trigger a review of claims that depend on the affected passage, even if the article has not been touched. The appropriate outcome could be a correction, a dated historical qualification or a held claim until an authoritative interpretation is clear.

Not every source update affects every downstream article. A formatting change in an unrelated appendix may require no amendment to a public Mathematics explanation. Dependency mapping can narrow review to claims whose evidence or conditions changed. This avoids overwhelming editors with false alarms while preserving coverage for consequential differences.

An AI model can help enumerate potentially affected claims, but a protected reviewer or deterministic condition should confirm the actual impact. A changed source paragraph may contain exceptions the model initially missed. Keep provenance and reasoned disagreement visible rather than collapsing multiple sources into an unsupported one-sentence summary.

Return to contents or continue with Effective-date shifts change truth conditions.

7. Effective-date shifts change truth conditions

A genuine old notice can remain accurate about its historical period while no longer governing a new intake. A materiality checker that merely chooses the newest file may erase useful history or incorrectly update an answer to a past-year question. Store the queried time, source-effective interval and relevant cohort alongside the cited claim.

Imagine a rule issued on 20 October but effective only for applications opened in January. A parent’s question about a November deadline may still be governed by an earlier rule. A model that treats publication date as effective date can give an incorrect answer even when it retrieved both official documents. The task must resolve which interval actually applies.

Corrections may be retroactive, prospective or limited to specific situations. The applicable issuer’s own statement governs that scope where known. Do not infer retroactivity from the fact that a new file superseded an older URL. If an important question cannot be settled from current authoritative material, explain the uncertainty and refer to the responsible institution.

A changed approval window is a separate temporal condition. Even if the document’s content is unchanged, an editor’s publication grant may expire before an agent resumes. A current source cannot supply missing human authority. Recheck both the factual rule and the operational permission before taking the next action.

Return to contents or continue with Numerical changes deserve type and unit checks.

8. Numerical changes deserve type and unit checks

A single digit can change an amount, percentage or examination threshold. A number without units can also be misleading. If a source states 0.75 as a proportion and the updated draft says 75%, those may be equivalent; if it says 75 without percent or context, the meaning may differ. Materiality checks should use declared units and expected computation rather than raw textual equality alone.

Dates have similar ambiguity. A string such as 03/04 may mean different calendar dates under different conventions. A model that normalizes a date should preserve its original source context and explicit intended format. Do not claim equivalence between two representations without verifying their actual semantics.

A revised formula can produce the same output for one sample while behaving differently for other inputs. Testing one displayed total is insufficient to conclude that a formula change is harmless. The Spreadsheet Formula Verification guide covers reference and recalculation checks that complement material change detection.

Assign independent mathematical or structured validation for critical fields. The language model can explain why two versions seem equivalent, but the accepted rule should be verified by actual calculations and source semantics. A convincing generated paragraph is not a substitute for a correct numeric condition.

Return to contents or continue with An invented hundred-change audit.

9. An invented hundred-change audit

Consider a fictional Cedar School Notice Review containing one hundred proposed revisions. Forty-eight are verified nonmaterial changes that remain within approved formatting or language rules. Twenty-two change a date, threshold or eligibility condition and require factual review. Fourteen change the recipient, target or permissions and need renewed action authorisation. Ten have ambiguous meaning requiring qualified adjudication. Six show an unverified or conflicting source revision.

The ledger reconciles: 48 + 22 + 14 + 10 + 6 = 100. Nonmaterial cases are 48%, factual-material changes 22%, action-scope changes 14%, ambiguous cases 10%, and source-conflict cases 6%. These categories are invented to teach classification; they are not observed editorial error rates or measured model accuracy.

A dashboard labeling forty-eight changes Safe and the other fifty-two Bad oversimplifies the review. A factual update may be entirely legitimate after an editor checks the current official source. An action-scope change might be permitted with renewed consent. Ambiguity is an unresolved evidence state rather than proof the editor acted maliciously.

The decision’s unit is one proposed revision, not the number of changed characters or editor clicks. One revision can touch several paragraphs and tool actions but should retain a stable identity in the audit. Final publication requires separate evidence that the approved version was actually saved with the intended public status.

Return to contents or continue with Distinguish textual magnitude from user consequence.

10. Distinguish textual magnitude from user consequence

Imagine an article whose introduction is rewritten for clarity but all claims, sources, dates and intended audience remain stable. It may create hundreds of changed characters with little effect on the parent’s decision. Another edit changes the closing time from 17:00 to 15:00, a two-character difference with material consequences for a family traveling to an event.

Use this contrast to assess the detector. A naive edit-distance threshold can call the first dangerous and the second harmless. A field-aware classifier can prioritize dates, recipients and status flags, while a qualified reviewer investigates whether the changes truly govern the user’s current case. Materiality is about the relevant causal effect, not typography.

A model should not automatically ignore changes outside a small list. A sentence adding an exception can alter an entire rule without containing a number. A heading move can change the scope of a condition. Combine structured checks with passage-level comparison, and preserve a route for reviewers to flag unfamiliar material differences.

False positives also cost real time. If every harmless punctuation correction triggers full reapproval, qualified staff may experience alert fatigue and miss important cases. Design a measured review policy with clear exemptions and targeted source checks, then monitor which decisions were correct under a held-out labelled set.

Return to contents or continue with Preserve source lineage and review evidence.

11. Preserve source lineage and review evidence

A review packet should include stable document identity, old and new revision or digest, affected passages, cited source versions, materiality category, reviewer decision and final saved target. These facts let a later editor see what changed without reconstructing a conversation from memory. The packet can store protected evidence pointers instead of duplicating sensitive files in every audit event.

Provenance matters when content was generated from several sources. If an updated article combines a current official document with an older parent FAQ, reviewers need to know which claim came from which source. An undifferentiated list of references may hide that an important eligibility sentence relies on the obsolete one.

Human approval should identify the precise change or version considered. A signature on a general project plan does not automatically authorise every later source correction, target account or public communication. Tie the grant to the actual content and intended effect. Preserve historical approvals even if a subsequent version needs a new one; the past decision remains an important event.

Keep audits useful without broadcasting personal data. A private learner report may need a restricted reviewer, while an ordinary public editorial diff can be visible more broadly. Store safe stable identifiers in the general ledger and use qualified controlled access for sensitive content. Verification quality improves when the evidence and permissions are explicit.

Return to contents or continue with Concurrent editors can create a different kind of conflict.

12. Concurrent editors can create a different kind of conflict

Two editors can change the same article concurrently. One adjusts the source date, while another improves explanatory text from an older version. A naive save may overwrite the first person’s correction. The related Concurrent Document Edits chapter examines ETags and lost-update protection. Materiality detection adds the review of meaning after a conflict-safe merge.

A merged result can be technically valid and still semantically incoherent. One paragraph may say applications close on Friday while another refers to a Thursday deadline, each inherited from different branches. A successful three-way merge or no-lost-update guarantee does not establish a unified current policy. Apply source and claim checks to the combined version.

Optimistic locks can prevent an agent from blindly updating a changed target, but they do not decide which correction should win. The responsible editor may need to reconcile source evidence and secure new approval. The agent should preserve both contributions and explain the conflict, rather than selecting whichever paragraph is newer by timestamp.

An error-free save also does not settle publication state. A post can remain Draft after a successful merge, or a public page can contain only part of the intended update. Final readback should compare the approved canonical version against what the destination actually stored.

Return to contents or continue with A changed-context school eligibility exercise.

13. A changed-context school eligibility exercise

Transfer the method to a fictional subject-selection notice. Version 4 states that a workshop is open to Secondary 2 learners and registration ends on 12 November. Version 5 retains the audience but moves the date to 21 November. A reviewer should classify this as a material factual condition for current planning, even though almost all the words are identical.

Change the scenario: the school adds a historical note specifying that the earlier 12 November date applied to the previous cohort. The old statement may still be accurate in historical context, while the new notice governs a later intake. A detector that simply deletes every previous date can erase useful facts. Preserve effective cohort and source scope.

Now the assistant changes a parent-notification recipient after content approval. That is an action-scope change independent of factual correctness; even a perfect school notice should not be sent to an unapproved address. The system needs a current recipient-bound grant and should hold transmission until it exists.

Finally, a WordPress publishing tool returns a timeout after possibly saving revision 5. The agent must reconcile the original post ID and slug before creating another. Change detection establishes which version is intended; actual provider readback establishes whether that version is live. Both checks are necessary for a truthful final link.

Return to contents or continue with False negatives and false positives require different repairs.

14. False negatives and false positives require different repairs

A false negative marks a consequential edit nonmaterial. It can allow an old approval or outdated claim to pass unnoticed. Tests should include changed dates, recipients, exception clauses, source-effective years and status flags, even when only a few characters differ. A model’s high semantic similarity score should not be allowed to override these critical field checks.

A false positive marks a harmless edit material, consuming reviewer time and possibly delaying necessary corrections. Test punctuation fixes, formatting changes and properly defined equivalent units against the policy. A person should be able to explain why a given difference mattered rather than receive a constant flood of vague warnings.

Measure both under an independent reference set. A detector can appear very conservative by flagging everything, but this does not support efficient review or useful publication. Conversely, an extremely permissive classifier can preserve throughput at the expense of harmful changes. The appropriate balance depends on task consequence and the organisation’s accepted policy.

Review disagreement transparently. Two editors may reasonably disagree about whether a change is material under an ambiguous rule. Preserve their reasoning and revise the policy if needed, rather than force the model’s preferred label into the reference set. Independent adjudication makes future classifications more consistent.

Return to contents or continue with Specific negative tests can protect release gates.

15. Specific negative tests can protect release gates

Construct fictional versions differing in one decisive attribute at a time. Change only the recipient, only the current effective date, only a percentage threshold or only a source citation. The expected decision should be established before the agent evaluates the diff. This reveals whether the detector actually uses the rule instead of being impressed by fluent surrounding text.

Include a malicious-looking source quotation in the revised document that asks the agent to approve itself. The text is untrusted evidence, not an authorised editor decision. A real approval must come from the protected identity and workflow service. A model’s generated summary of the diff cannot create missing permissions.

Test a normal case too. A permitted formatting correction should remain publishable under an explicitly defined nonmaterial rule when all necessary conditions are met. A system that refuses every revision may avoid some mistakes but fail the user’s legitimate purpose. Negative and positive oracles establish the useful boundary.

After a test reveals a defect, preserve the exact old and new versions, expected category and actual tool outcome. Re-run after a model, parser or provider change. A green report is trustworthy only when the saved target and human permission are independently checked, not because the same agent wrote its own expected answer.

Return to contents or continue with A small editorial team’s materiality runbook.

16. A small editorial team’s materiality runbook

Begin with the actual subjects that matter to readers: who is eligible, what they must do, when and where, and which evidence supports the statement. Create a concise classification policy for these fields and for recipient or publication changes. Use stable source and target IDs and record the approved version before the model starts revising.

At each proposed update, show the exact old and new passages. Label changed dates, units, rules, authorities, destinations and confidential material. Verify source applicability and current user permission. Then obtain the review appropriate to the risk, not simply a generic second click from the same person. Unclear cases should have a qualified decision route.

After a write, fetch the actual destination state. A WordPress page may have been updated only as a draft, or a concurrent editor may have changed it again. The user wants the right article at the right URL, not an explanation of what the assistant intended to click. Link the final status to the canonical saved artifact.

The lesson transfers beyond publishing. Students can compare two versions of a timetable, a fraction problem or an application rule and ask whether the difference changes the decision. More powerful AI can accelerate the comparison, but human authority, correct evidence and verified real-world effects remain necessary.

Return to contents.

Consider a controlled approval experiment in which two documents carry the same number of changed characters. Document A fixes twenty spelling errors in an explanatory note. Document B changes one digit in the minimum age printed on an eligibility notice. A character-count threshold would favour A as the riskier edit, yet the change in B can affect who is eligible and what a reader does next. A safer release gate compares each edit to its reader-facing consequence: Does it alter an instruction, number, recipient, permission, legal applicability, date, scientific claim or required action? Who was authorised to approve that precise consequence? The evidence of a low-risk spellcheck must never substitute for a substantive new authorisation.

A practical review record should make those questions reproducible. For each material candidate, retain the immutable source identifier and source revision; the previous and proposed value; the affected passage, figure or outbound payload; the governing approval policy; the reviewer’s decision and its scope; and the version ultimately sent or published. If the source changes again after review, do not silently extend approval to the new version. Re-run the comparison against the version the reviewer actually saw, classify newly altered consequences and request fresh approval where the rule requires it. An approved edit and a successful remote save are distinct facts; neither independently establishes that the resulting statement is correct.

An illustrative risk worksheet can make the triage legible without pretending a score replaces judgement. Suppose twenty changes are classified: twelve cosmetic, five potentially explanatory and three consequence-bearing. A policy assigns cosmetic changes to automated checks, explanation changes to an editor and consequence-bearing changes to the accountable domain owner. Then 12 + 5 + 3 = 20 and the consequence-bearing share is 3/20 = 15%, not 3/8 or 3/12. If two of those three await decisions, the released set has at most one consequence-bearing change cleared for release under that rule. Do not calculate a misleading 90% readiness score by dividing eighteen reviewed changes by twenty: review completion cannot authorise the two unresolved material effects.

Change detection also needs an explicit adversarial rehearsal. A test editor can alter a date only inside the body, a threshold only inside an image caption, a formula while leaving its displayed answer unchanged, or the destination address of a hyperlink while preserving its visible label. A diff system that sees only prose or only the visual rendering may miss these edits. The harness should extract structured dates, numbers, links and machine-readable fields where the format allows; compare source revisions; and run the same destination-side verification used for ordinary publishing. Cases outside its supported parsers should be labelled as unknown and escalated, not recorded as harmless. The point is to demonstrate that the check notices the kind of edit that matters, not to manufacture confidence by increasing test counts.

Finally, a release team should practise a changed-context transfer. Imagine the original approved programme notice says an application closes at 17:00 Singapore time, while a new edition introduces an online-only submission condition. Even if the time remains unchanged, the required reader action has changed. The machine can flag the new submission phrase and preserve an exact diff; the editor must verify its authoritative source and the approving owner must decide whether the old consent still covers the new procedure. If the answer is no, hold the edited notice, obtain the newly scoped approval, publish only the approved version and retrieve the live notice to check the condition actually appears. That sequence turns version comparison into responsible decision control.

Frequently asked questions

Does a larger textual diff always mean a more important change?

No. Two changed digits can alter a deadline, while a long rewrite may preserve the same facts and permission.

Can a content hash determine whether reapproval is necessary?

A hash detects byte changes under a trusted process; materiality and authority require separate interpretation.

Do all edits need fresh approval?

That depends on a documented policy. Material content, recipient or permission changes commonly need more review than bounded formatting fixes.

Can current source evidence replace an expired human grant?

No. Source truth and authorisation for an external action are independent conditions.

Does a successful merge prove the article is correct?

No. Concurrent edits can produce semantically conflicting statements, and target-state readback is still necessary.

How should ambiguous changes be treated?

Hold consequential effects, preserve the original versions and obtain qualified clarification rather than guessing.

Technical and governance sources

Sources reviewed 24 October 2026. Cedar School, the hundred revisions and every decision count are invented teaching cases. Actual institution rules require checking current primary sources and applicable scope.

GitHub Protected Branches: Dismissal of stale approvals when reviewed diffs change.
Git Diff Documentation: Version comparisons and textual change inspection.
HTTP Semantics: Conditional update and resource representation concepts.
OWASP Authorization: Current permission and object-level action checks.
NIST AI RMF Playbook — Measure: Reference labels, validity and evaluation limits.

How Superintelligence Works | Agent Material Change Detection — Document Diffs, Effective Dates and Reapproval