Codex 0 · Chapter 06 — The Knowledge Graph (ENG-01)

This chapter specifies the node and edge model, provenance rules, and query obligations of the Knowledge Graph.

06.0 Purpose

The Knowledge Graph is the single linked model connecting capabilities, curriculum, publications, sources, and demonstrated ability. It exists so that "what does this concept connect to, and what is the evidence" has one authoritative answer institution-wide, rather than a different answer per division. This chapter binds ENG-01 (Chapter 05 §05.6.1) to a concrete schema and set of obligations.

06.1 Scope and non-scope

06.1.1 This chapter governs node types, edge types, the capability taxonomy, provenance, versioning, attachment rules for curriculum/publications/assessments, required query patterns, exclusions, and scale behaviour.

06.1.2 This chapter does not govern the Adaptive engine's sequencing algorithm (Chapter 05 §05.6.3), which consumes the graph but is a separate engine, nor the canonical entity definitions the graph references (Chapter 07).

06.1.3 This chapter does not define the public glossary surface (SURF-06) though the taxonomy underlies it; that surface's experience rules are Codex 6's concern.

06.2 Node types

06.2 The node catalogue

06.2.1 Capability — a named, describable human capability at a specific taxonomy level (§06.3). Authored by Research Institute. Canonical entity: ENT-06.

06.2.2 Concept — a unit of knowledge smaller than a capability, used to link curriculum content and publications to the capabilities they support. Concepts do not themselves carry a taxonomy level; they are the connective vocabulary between capabilities and content.

06.2.3 CurriculumUnit — a reference to a school, program, course, module, or lesson (ENT-09). The graph stores a reference node, never a copy of the unit's content.

06.2.4 Publication — a reference to a research output (ENT-13).

06.2.5 Source — a cited reference (paper, dataset, primary document) underlying a publication or a capability definition.

06.2.6 Assessment — a reference to a capability assessment instrument or a specific recorded assessment event (ENT-07), used to link demonstrated ability back to the capability it evidences.

06.2.7 Credential — a reference to an issued credential (ENT-08), linked to the capabilities it certifies.

06.2.8 Person (capability-record projection) — not the Person entity itself, but a minimal projection node representing "this person has demonstrated node," used only to anchor edges from Assessment/Credential nodes. It carries no personal content beyond the person's identifier.

06.2.9 Every node carries: a stable ID, a type, a reference to its canonical source entity (except Concept, which is graph-native), a status (draft, published, deprecated), and a provenance record (§06.4).

06.3 The capability taxonomy and its levels

06.3.1 The taxonomy is a directed acyclic structure of Capability nodes connected by prerequisite-of and part-of edges (§06.4). It is authored and maintained by Research Institute (DIV-03), consistent with Codex 0 §2.

06.3.2 Levels, from broad to specific:

Domain        e.g. "Quantitative Reasoning"
  └─ Strand    e.g. "Statistical Inference"
       └─ Capability   e.g. "Constructs a confidence interval"
            └─ Sub-skill  e.g. "Selects the correct standard error formula"

06.3.3 A Capability node is the smallest unit that a Credential or CapabilityAssessment may reference directly. Sub-skills exist to support curriculum authoring and adaptive sequencing; credentials are never issued at sub-skill granularity alone.

06.3.4 A capability's level is fixed at creation. Reclassifying a capability to a different level is a versioned change (§06.5), not an in-place edit, because existing assessments and credentials reference the prior definition.

06.3.5 Every Capability node states, in plain language, what a person who holds it can do — never only a topic label. "Understands regression" is not a valid capability statement; "constructs and interprets a linear regression model from a given dataset" is.

06.4 Edge types, direction, and cardinality

06.4.1 No edge type may be added informally; a new edge type is proposed with its direction and cardinality stated, following the same decision-record discipline as a schema change (Chapter 07 §07.1).

06.4.2 prerequisite-of edges must not form a cycle. A cycle is a data integrity defect and blocks publication of the offending edge.

06.4.3 An edge from any node to a Person projection node is permitted only for demonstrated-by, and only referencing that person's own record; no other node type may point at a Person projection.

06.5 Provenance and evidence rules

06.5.1 Every node and edge records: who or what produced it (a person, an authoring tool, or a pipeline), when, and under what workflow (curriculum authoring, editorial pipeline, assessment recording).

06.5.2 A Capability node's provenance additionally records the Research Institute review that approved its definition; a capability is never added to the graph without a named reviewer of record.

06.5.3 An evidences or assesses edge without a resolvable Source or Assessment reference is invalid and is rejected at write time, never published in a partial state.

06.5.4 Provenance is retained even after a node is deprecated (§06.6); deprecation removes a node from active query paths, it does not erase the record of how it came to exist.

06.6 Versioning

06.6.1 A published node is never edited in place if the edit changes its meaning (what it asserts a person can do, or what it certifies). A meaning-changing edit creates a new node version connected to the prior one by a supersedes edge.

06.6.2 Non-meaning-changing edits (correcting a typo in a label, fixing a broken source link) may be made in place and are logged as an Event (ENT-18).

06.6.3 An Assessment or Credential that references a capability continues to resolve to the exact version of that capability current at the time the assessment or credential was created, even after the capability is superseded. This is what makes a credential's meaning stable over decades (Chapter 07 §07.4, Codex 5 "Contract rules").

06.6.4 A deprecated node remains resolvable by ID indefinitely; it is excluded from default query results (adaptive sequencing, discovery) but never deleted while any Assessment, Credential, or Publication still references it.

06.7 Attachment of curriculum, publications, and assessments

06.7.1 A curriculum unit is attached to the graph via one or more teaches edges, authored at the point the unit is created or revised in curriculum authoring (TOOL-01). A unit with no teaches edge is a defect: Adaptive cannot reason about it and Learning should flag it.

06.7.2 A publication is attached via discusses and cites edges, authored during editorial review (ENG-09) before publication, never retrofitted casually after the fact — retrofitting requires the same review discipline as original attachment.

06.7.3 An assessment instrument is attached via assesses edges at the time the instrument is authored; a specific assessment event is attached via demonstrated-by at the time it is recorded, never edited retroactively except by the correction-as-new-version rule (§06.6.1).

06.8 Query patterns the graph must serve

06.8.1 Adaptive sequencing — given a person's demonstrated capabilities (via demonstrated-by and assesses edges) and a target capability, return the shortest valid path of prerequisite-of edges not yet demonstrated, with the teaches units that build each step.

06.8.2 Tutor grounding — given a concept or capability a scholar is asking about, return its definition, its part-of parent, its evidencing sources, and the curriculum units that teach it, so the Tutor Agent (AGT-02) can ground its explanation in the same material a human teacher would cite.

06.8.3 Research linkage — given a capability, return every publication that discusses it and every source that evidences it, supporting Research Institute's ability to show a capability's definition is not invented but grounded (Codex 1 Article VI.8).

06.8.4 Credential meaning — given a credential, return the exact version of every capability it certifies, resolvable permanently, satisfying the twenty-year verification requirement (Codex 5, Chapter 07 §07.9).

06.8.5 Each of these query patterns is a named, tested interface of ENG-01, not an ad hoc traversal left to each consumer to reinvent.

06.9 What may never become a node

06.9.1 A member's personal content — journal entries, private goals, private financial data, private engagement material — is never represented as a graph node, even anonymized, because the graph is institution-wide and queryable by design; personal content belongs in its owning engine's private state (Chapter 05 §05.2.2).

06.9.2 A specific person's identity is never a general-purpose node; the only permitted person-referencing node is the minimal Person projection used solely to anchor demonstrated-by edges (§06.2.8), and it is never traversable by anyone other than that person and explicitly permitted roles.

06.9.3 Advisory engagement detail, venture financial detail, and other engagement-scoped material (ENT-12, ENT-15, ENT-16) are never graph nodes; only anonymized, reviewed derivations (per Chapter 07 §07.11) may ever surface as Concept or Source material, and only through the same editorial review as any other source.

06.9.4 Draft, unreviewed capability definitions are never published nodes; they remain in Research's editorial pipeline state (ENG-09) until reviewed.

06.10 Scale behaviour

06.10.1 The graph is designed to answer the four query patterns of §06.8 in bounded time regardless of institution size; a query pattern that degrades unpredictably as nodes grow is a defect, not an acceptable tradeoff, given Codex 1 Article VII's hundred-year horizon.

06.10.2 Node and edge writes are expected to be orders of magnitude less frequent than reads; the graph is optimized for traversal, and write paths (curriculum authoring, editorial review, assessment recording) are the only permitted mutation sources (§06.7).

06.10.3 As the taxonomy grows across domains and languages, part-of and prerequisite-of structure must remain navigable without a full-graph scan; query patterns are scoped by domain or capability neighborhood rather than requiring global traversal.

06.10.4 Multi-language capability statements (Codex 0 §15.3 localization) are modeled as localized labels on the same node, never as separate nodes per language, so a credential's meaning does not fragment by locale.

06.11 Interfaces

06.11.1 Published as API-09 (Knowledge Graph), serving ENG-03, ENG-09, AGT-02, AGT-03, per the registry.

06.11.2 Write access is restricted to the producing systems named in §06.7: curriculum authoring, the editorial pipeline, and the assessment/credential recording path. No general-purpose write endpoint exists.

06.11.3 Read access follows the same server-side authorization as any other query: a query pattern is available to a role if the underlying nodes it would return are available to that role (§06.9.2's Person-projection restriction applies regardless of the calling engine).

06.12 Invariants

06.12.1 No prerequisite-of cycle.

06.12.2 No edge points at a Person projection except demonstrated-by, and never across people.

06.12.3 Every published Capability node has a named review and a resolvable evidence trail.

06.12.4 A superseded node remains resolvable for as long as any Assessment, Credential, or Publication references it.

06.12.5 No node duplicates a canonical entity's fields; it references the entity's ID (Chapter 07 §07.1).

06.12.6 An embedding index contains no content that is not already present in a canonical, published node (§06.15.2, §06.15.3).

06.12.7 Every published Capability node carries a review date; overdue review is visible, never hidden (§06.16.1).

06.13 Prohibitions

06.13.1 No personal, private, or engagement-scoped content becomes a node (§06.9).

06.13.2 No in-place meaning-changing edit to a published node (§06.6.1).

06.13.3 No write path outside the named producing systems (§06.11.2).

06.13.4 No cross-locale node fragmentation for the same capability (§06.10.4).

06.13.5 No unreviewed capability definition is published.

06.13.6 No embedding generated from personal, private, or engagement-scoped content (§06.15.2).

06.13.7 No silent embedding-model swap that could change retrieved results without a logged reindex event (§06.15.5).

06.13.8 No deletion of a decayed or flagged node/edge as a substitute for editorial review (§06.16.5).

06.14 Open questions

06.14.1 Whether Concept nodes require their own review workflow distinct from Capability review, or may be authored more loosely by curriculum authors — reserved pending ENG-09 build-out.

06.14.2 How the graph represents capability equivalence across partner institutions' own frameworks (Codex 0 §15.2) — reserved for a decision record when the first partner integration is proposed.

06.14.3 Whether sub-skill nodes should be queryable directly by the Tutor Agent or remain an internal Adaptive/Learning concern — reserved.

06.14.4 The exact review cadence for Capability node revalidation referenced in §06.16.1 is not yet fixed by Research Institute.

06.14.5 Which embedding model and hosting arrangement is used, and whether it is self-hosted to avoid sending institutional content to a third party, is undecided (§06.15).

06.15 Retrieval and embedding policy

06.15.1 The graph's structural traversal (§06.8) is the authoritative answer to "what is connected to what." Vector embeddings are a supplementary retrieval index over node text (labels, definitions, publication abstracts) used only to shortlist candidates for a structural query or a natural-language question put to an agent; an embedding similarity score is never itself treated as evidence that a capability is a prerequisite, part-of, or evidenced by another node.

06.15.2 Embeddings are generated only from published, institution-owned node content — never from a member's personal content, submitted artifact, or private engagement material (§06.9). Regenerating embeddings from a deprecated or superseded node is permitted only to keep historical search results resolvable; it does not restore the node to active query paths (§06.6.4).

06.15.3 An embedding index is a derived cache, not a canonical store. It is rebuildable in full from the node/edge store at any time and carries no data that does not already exist in a canonical node; loss of the embedding index degrades search quality, never correctness of a structural answer.

06.15.4 Where an agent (AGT-02, AGT-03) uses embedding retrieval to ground a response, the response cites the specific node IDs retrieved, satisfying the same explainability requirement as a structural query (§06.8.2, Codex 1 Article VI.4). Retrieval that cannot be traced to specific nodes is not used to ground a claim presented to a person.

06.15.5 Embedding models and their versions are recorded against the index they produced. A change of embedding model is a reindex, logged as an Event (ENT-18), never a silent swap that could shift search results without explanation.

06.16 Quality and decay

06.16.1 A published Capability node is reviewed on a standing cadence set by Research Institute (not less than once per taxonomy revision cycle, §06.14 pending decision) to confirm its statement, level, and prerequisite edges still reflect current disciplinary understanding. A node overdue for review is flagged, never silently trusted indefinitely.

06.16.2 A teaches edge whose curriculum unit has been substantially revised without a corresponding review is treated as stale; Learning (ENG-02) surfaces staleness to curriculum authors rather than continuing to feed it to Adaptive sequencing unexamined.

06.16.3 Evidence decay: an evidences edge to a Source that has been retracted, corrected, or superseded in its own field is flagged for editorial re-review; the edge is not silently removed, because Assessments and Credentials may already resolve through the capability it supported (§06.6.3).

06.16.4 Graph quality metrics — orphaned nodes (no incoming or outgoing edge other than provenance), nodes overdue for review, and unresolved evidences/assesses references — are reported to Research Institute and Technology on a standing basis; an ever-growing count of any of these is a defect, not a acceptable steady state.

06.16.5 Decay handling never deletes a node or edge to make a metric look better (§06.6.4); it surfaces the condition for a named reviewer to act on.

06.17 Governing Codices

Codex 3 (Knowledge Graph named as engine); Codex 5 (normalization and contract rules); Codex 1 Article VI.8 (evidentiary claims); Chapter 05 (ENG-01 treatment); Chapter 07 (canonical entities referenced by nodes); Chapter 08 (access to graph queries).