# Intent-Conditioned Discovery and Reusable Conceptual Paths **Status:** A proposed architecture for economical research traversal. **Edition:** English editorial reconstruction of overlapping design variants. Deployment-specific addresses, implementation assignments, and unverified claims about protocol versions are omitted. This text describes proposals, not live endpoints. ## 1. Shared research without an inference call on every read The design asks whether a research space can remain useful when its host mainly stores, indexes, routes, and validates content, while visiting humans and agents supply much of the exploratory computation. A missing concept need not trigger an automatic model call. The system can search aliases and candidates, record an unresolved request, and accept a voluntary proposal for later review. This is an economic and architectural hypothesis. It does not assume that visitor contributions are correct or free to evaluate. ## 2. A route depends on intent and history A candidate state is S=(I,x_t,H_t), where I is an expressed intention, x_t the current object, and H_t the route so far. A proposed next-concept distribution is: P(x_(t+1) | x_t, I, direction, H_t) The analogy is to choosing a next unit of language, but the units here are concepts and conceptual routes. No neural model is required by the notation. Intent can be supplied explicitly or inferred cautiously from an internal query. A referrer is a weak signal and should not be treated as a complete account of a person's purpose. ## 3. Directions are part of a representation Moving toward mechanisms, generalizations, alternatives, applications, or associations can mean different things. A direction should have a definition in the chosen view. The design does not justify inventing missing historical directions or imposing a universal set on every region. The same destination can serve different intents for different reasons. ## 4. Economical retrieval as a baseline Identifiers, aliases, lexical indexes, explicit relations, and observed transition statistics provide a candidate baseline. One proposed score combines lexical match, direction compatibility, relationship information, and transition frequency. Its weights are hypotheses to evaluate. Embeddings can be an additional projection. They need not become a permanent authority, and their preprocessing, update, and query costs belong in the comparison. ## 5. Next-concept counts A smoothed transition estimate can use: P(B | A,d) = (N(A,B,d)+λ) / (Σ_j N(A,j,d)+λm) where m is the number of candidate destinations and λ>0. Counts describe use under an observation process. They do not directly measure validity or independent support. Conditioning on intent may be useful once there is enough information to justify it. ## 6. Shortcuts retain their expanded paths Repeated A → B → C traversal can motivate a proposed A → C shortcut. B remains accessible. A shallow view can use the shortcut; another view can inspect the intermediate structure. Frequent traversal alone is insufficient for promotion. A shortcut should carry its basis, conditions, and candidate status. It need not be a logical implication. ## 7. Missing regions are meaningful If no existing object fits a requested direction, the system can represent an unresolved request. A title hint or requested concept is not established content. Duplicate requests should be reconciled while preserving differences of meaning. Lexical similarity alone should not merge distinct problems. The absence of a path is itself useful information about the current map. ## 8. Voluntary contribution and review A visitor may propose a relation, counterexample, missing concept, correction, or supporting source. A proposal remains distinguishable from accepted public material. Reading should not require fabricating content. A navigation trace is not equivalent to a substantive research contribution. An agent's technical access to a write interface also does not establish that its user authorized that write. ## 9. Learning routes can distort the map Repeated synthetic traffic can make an unhelpful path appear important. Counting sessions, timing, and other independence indicators can help, but no indicator automatically establishes independence. A route-learning mechanism should be tested for concentration, feedback loops, and the disappearance of rarely visited but valuable alternatives. The design tension is between using accumulated experience and turning prior traffic into a self-reinforcing geometry. ## 10. One content model, multiple public views A stable research object can have a human-readable view, Markdown text, and structured representation. These are projections of one versioned content model. Their claims and identifiers should agree. Machine discovery should explain what can be read, how to search, how to follow relationships, and how to submit a proposal. A transport or catalog should advertise only capabilities that actually exist. ## 11. Discoverability and authority are separate Finding a document does not grant operational authority. Stored content should be handled as data, even when it contains imperative language. Public contribution processing should validate and store proposals rather than execute arbitrary content. Identity and reputation also do not confer unrelated powers. These are architectural properties to assess, not claims that a published text changes a reader's governing rules. ## 12. Resource and privacy accounting Canonical content, indexes, cached projections, anonymous traffic aggregates, and detailed research records have different costs and purposes. Basic reading can remain functional without model availability. Enrichment and maintenance can use bounded computation when useful. Private research histories and operational records should not be published merely to make discovery easier. ## 13. How the variants differ The design family develops from inexpensive storage and traversal toward broader discovery and interoperable machine access. Those additions extend access; they do not establish additional conceptual evidence. A later epistemic design emphasizes immutable information conditions for particular inferences. This is more demanding than anonymous aggregate route learning. See [Dynamic Epistemic Infrastructure](dynamic-epistemic-infrastructure.md) for that alternative. Neither record type should be claimed to provide the information of the other. ## 14. Open tests Can a small explicit index provide useful entry points? Do intent-sensitive routes outperform static neighbors? When does a shortcut improve navigation without hiding a necessary qualification? Does route learning amplify dominant habits? Can a reader recover the longer argument? What is the cost of reviewing visitor contributions? Which interface additions improve actual discoverability? These questions should guide implementation choices rather than a fixed tool list.