From seven graphs to nine capsule layers: separating records by responsibility

Keep the history without presenting it as the current schema

The original note described Base, Evidence, Analytical, Narrative, Commitment, Case, and Scenario graphs. That was an attempt to separate source material, interpretations, and counterfactual work. It remains useful as design history, but it is not the current capsule layer list. The current design names evidence, claims, situation, temporal, ontology, reasoning, memory, governance, and runtime.

These lists are not equivalent partitions with two names added. A commitment can be represented as a claim with temporal qualifications, situated among actors and constraints. A scenario can involve several layers and must remain distinguishable from an observed record. Migration requires explicit mappings and preserved identifiers, not renaming graph labels until a validator accepts them.

Nine responsibilities in one inspectable object

Names are not enforcement

RDF datasets can group triples in named graphs; SPARQL can select and query those graphs. Those standards provide representational and query mechanisms, not a guarantee that a runtime keeps an unreviewed assertion out of an answer. The application still needs access checks, explicit query scope, and tests for accidental unions across layers. [1][2]

A relational implementation can express the same responsibilities using versioned tables and joins. A document store can preserve a complete package snapshot. A graph projection can make relationship traversal convenient. The choice should follow query patterns, transactional requirements, and operational cost. Nine logical layers are not an argument for nine services or nine independently mutable stores.

Follow a correction across the package

Suppose a source correction changes a meeting date from 8 September to 9 September. Keep the source versions and record the correction. Update the relevant temporal interpretation through review. Then identify derived artifacts: a chronology, a narrative comparison, a memory summary, and any answer that relied on the earlier ordering. Their dependencies matter more than which storage collection happens to contain them.

PROV-O distinguishes entities, activities, derivation, revision, and invalidation. It provides useful terms for this dependency record, but a provenance edge alone does not refresh a summary. A scheduler or reviewer must decide which artifacts to recompute, which to mark stale, and which historical snapshots to preserve. [3]

Select context; do not serialize the entire architecture

An agent comparing two commitments may need the relevant claims, dates, source passages, and a small part of the ontology. Sending all nine layers in full could waste tokens and expose unrelated material. A context assembler should select a task-specific view while retaining identifiers that allow the agent or reviewer to retrieve omitted detail. Context-engineering guidance supports this distinction between a durable record and a working context window. [4]

A compact view is lossy. It must not silently drop a dissenting claim or the condition that makes two undertakings different. Version the selection policy and record which source snapshot it used. Measure context size alongside evidence recall and answer quality; smaller prompts alone are not a successful optimization.

The test is a reconstructable record

Our evaluation should round-trip a capsule, reconstruct an earlier view after a correction, and rebuild derived graph projections from retained records. It should also attempt unauthorized reads and promotion of unreviewed changes. Compare a layered representation with a simpler versioned document or relational design before accepting the extra machinery.

PRAXIS remains the user-facing workspace, with DIALECTICA as the proposed ingestion and review boundary. The nine-layer description is a developing format and architectural direction. Older reference fixtures can contain a different layer set; their manifest and schema version, not this article title, determine what they actually contain.

Read the historical terminology through a dated crosswalk

The seven-graph model describes an earlier research view. The website’s seven communication lenses organize an authored teaching fixture; they are not the same list. Nine architecture responsibilities and four contextual roles describe further concerns. The compiled reference contains eight layers. Keep each artifact’s pinned version and omissions explicit rather than asserting that a new diagram proves a composed runtime. [5]