SEDIM · Reference architecture

Knowledge has layers.
Make their roles explicit.

SEDIM separates six architectural responsibilities so a current implementation can be evaluated without being confused with a target system or a research claim.

Explore the architectureOpen developer resources
Interactive architecture map

Available today and target architecture, kept visibly separate.

Qwen3-8B · grounding + records
Sources→Grounding→Answer + record
01 · FACIESCurrent

Base-model slot

Provides the language-capable foundation used by the current reference implementation.

Available today
Qwen3-8B supplies the current base-model slot.

Six responsibilities, not six sequential runtime steps.

Responsibility register

Six roles. One inspectable composition.

RoleResponsibilityStatus
01FACIESBase-model slotProvides the language-capable foundation used by the current reference implementation.Current
02VARVEManaged knowledgeCarries versioned knowledge with scope, provenance and lifecycle boundaries.In development
03CLASTLayer initialisationExplores Cross-Layer Asymmetric Spectral Transfer for managed-layer initialisation.Research
04STRIARouting roleRoutes a request across the knowledge responsibilities allowed to participate.Architecture
05STEMMAParticipation recordDeclares the source and layer participation behind a composed result.Current
06CENTOComposed outputPairs the composed answer with the record required to inspect it.Current
Evaluation checklist

Evaluate the path that actually exists.

  1. 01

    State which architecture path is being evaluated: available today or target architecture.

  2. 02

    Inspect the sources, routing signals and participation record—not only the final prose.

  3. 03

    Keep current implementation evidence separate from research and forward-looking claims.

  4. 04

    Test abstention, source removal and re-verification within the declared evaluation scope.

Development boundaries

Important limits remain part of the architecture story.

01

Trained VARVE/CLAST layers and Craton inference are not part of the current production path.

02

A lexical claim-support detector failed its baseline; that separate detector is not live in gamma.

03

Removing a source from serving does not erase every weight artifact or fact already present in a base model.