glpi-neuroforge-mega/docs/UNIFIED-GRAPH.md

Unified Graph Explorer (v1.4.0)

The Control Center remains read-only and now normalizes operational, evidence, learning, research and engineering relationships into one graph contract (nodes[], edges[], bounded metadata).

Views

• Runtime & Trust — services, external systems, scoped credentials and authority boundaries.
• Ticket Evidence — ticket, run, KB candidates, validated outcomes, policy checks, model attempts, proposed reply and human decision.
• Learning Lineage — accepted/corrected outcomes, NeuroForge memories and immutable supersedes chains.
• Research Provenance — goal -> query -> source -> claim/evidence -> memory without exposing full source bodies or prompts.
• NeuroForge Brain — bounded/redacted memory/synapse/consolidation view; vectors and full memory exports are not returned.
• Engineering Graph — reproducible Go AST + root Compose snapshot with components, packages, files, functions, HTTP routes and service dependencies.
• Change Impact — bounded bidirectional dependency traversal for a file/symbol/route query with a conservative static risk hint.

Visualization

The browser uses a dependency-free canvas renderer. 2D is the operational default. 3D is an optional pseudo-perspective explorer for bounded subgraphs. Node budgets and server-side filtering prevent accidental full-graph rendering.

The graph is an explanation/inspection surface, not a decision authority. A high change-impact hint does not replace tests, code review or runtime evidence.

Trust boundaries

The Control Center never receives Agent admin/basic-auth credentials. Agent graph reads require CONTROL_READ_TOKEN; NeuroForge graph reads use the existing scoped app key. Graph endpoints are GET-only and return redacted/bounded representations.

Reproducibility

Regenerate the engineering snapshot after structural code changes:

make engineering-graph
make engineering-graph-check