# Dynamic Epistemic Infrastructure **Status:** A proposed research architecture. **Edition:** English editorial reconstruction. The design described here is an object of inquiry. It is not a statement that every proposed service is implemented in the published reader. ## 1. What information conditions produced a conclusion? A research infrastructure should make it possible to inspect problems, claims, representations, critiques, inferences, and revisions together with the conditions under which a conclusion was formed. These objects need not follow a linear workflow. A representation can reopen a problem; a critique can question an inference rather than its premises; a new result can revise a relationship. The representation of an argument must remain distinguishable from an assertion that the argument is valid. ## 2. Authorship and epistemic support A signature can establish that a particular key signed particular content. It does not establish the truth of that content. Actor identity, authorship verification, contribution history, and evaluation of a claim should remain distinct. The same applies to model reputation: capability can be relevant metadata without becoming a truth guarantee. Support can attach to a claim, evidence item, inference, or relationship. Even if A and B are correct, the relation “A supports B” can be wrong. ## 3. Inferences as inspectable objects An inference should be more than an unqualified arrow. It can have multiple premises, a conclusion, conditions, evidence, and a concise justification. Another contribution may challenge the inference itself, identify a missing premise, offer a counterexample, or propose another representation. A long model output should not become authoritative merely because it has been generated. Candidate claims and relations need review on their own terms. ## 4. Multidimensional evaluation Support, logical coherence, empirical support, clarity, representation quality, uncertainty, counterevidence, and independence need not reduce to one score. A display may offer a compact projection, but the underlying evaluations and disagreement should remain available. Popularity, vote count, model consensus, traffic, independent support, and evidence strength are different quantities. Repeated exposure can amplify a weak claim. ## 5. Availability is not use | Information state | Meaning | | --- | --- | | Available | The resource could be accessed under the applicable conditions | | Requested | Access was sought | | Exposed | Its content was delivered | | Cited | It was explicitly referenced | | Used as a premise | It appears in the stated justification | A request may fail. Delivered content may be ignored. A cited resource may not be a premise. **Editorial qualification:** These are records of access and declared use. They do not provide complete introspective access or prove which content causally determined an internal computation. ## 6. Research routes need not be fixed A participant may inspect primary evidence first, compare arguments first, seek counterexamples, or request social evaluations when uncertain. Blind assessment is one possible experimental condition. It is not a mandatory universal workflow. If evaluations or reputation were exposed, that exposure should remain visible when the resulting contribution is assessed. ## 7. Freeze the historical conditions of an inference Once an inference is issued, its recorded information conditions should be fixed. Later evidence can justify a revision, but it cannot become part of the earlier inference's original basis. Graph history, visited resources, and stated premises are different views. Creation time alone is insufficient: a resource that existed earlier may never have been delivered. A version identifier or content digest is more precise than a mutable URL alone. ## 8. Revision without retrospective rewriting A changed conclusion becomes a revised object linked to the prior one. Earlier content can be challenged, superseded, or retracted while remaining recoverable under the appropriate access policy. The current graph view may change as evaluations and evidence change. A historical view should preserve what was recorded then. This supports inquiry into why a claim was once considered strong and which counterexample changed that assessment. ## 9. Independence is an ancestry question Two systems reaching the same conclusion are not automatically two independent confirmations. They may share evidence, copy a prior inference, or be exposed to the same vote distribution. Separate model names alone do not establish independence. Information ancestry can help assess dependence, although it cannot reveal every shared influence. Independence should remain qualified rather than represented as a simple count without a method. ## 10. Traffic can reveal routes without validating them Repeated movement along A → B → C may identify a useful conceptual path. It does not demonstrate that each edge is true or that A entails C. A shortcut can be a navigation proposal while its intermediate reasoning remains available. Traffic and support should be stored and interpreted separately. ## 11. Candidate storage and interfaces Possible records include actors, epistemic objects, relationships, evidence versions, evaluations, revisions, access events, and inferences. A relational store with structured fields can support such a design; a specialized graph database is not conceptually mandatory. Exact technology is a deployment choice. Read interfaces, access to evaluations, candidate contributions, and revision proposals should declare their effects clearly. A publicly readable proposal does not grant an agent authority to execute it. ## 12. Human and machine views An idea page can expose its question, formulation, alternative representations, supporting and opposing relations, evaluations, and revision history. An inference can expose its declared premises and available access history. Machine-readable views should preserve the same distinctions, rather than flattening them into a prose confidence score. Private information conditions need not become public. A public research artifact can expose qualified evidence anchors while retaining private recovery mappings locally. ## 13. Counterfactual evaluation The architecture could support controlled comparisons: the same task with social evaluations hidden, exposed early, or exposed only on request. Such comparisons are proposed experiments. Merely preserving an access record does not establish the effect of an exposure. Model stochasticity, order, and other conditions must be considered. ## 14. Integrity questions Can a later resource be incorrectly backdated into an earlier basis? Can content change without a version change? Are edge evaluations distinguishable from node evaluations? Does a popularity projection masquerade as evidence? Can an apparently independent confirmation be traced to an earlier conclusion? Can a reader inspect a superseded claim without confusing it for the current formulation? ## 15. An unresolved architectural tension A cheap navigation system may aggregate anonymous transition counts. A study of an inference's information conditions may require a more detailed record. These serve different purposes. Aggregated traffic cannot reconstruct a detailed epistemic route. Detailed records should be collected only where their purpose and access policy justify them. See [Intent-Conditioned Discovery](intent-conditioned-discovery.md). The two designs can coexist as alternatives or serve different layers, without pretending their information requirements are identical.