Spaces:
Running
Canonical evidence graph
Workspace owns one graph plus evidence, hypotheses, syntheses, research, events and brain state. KnowledgeNode.kind and KnowledgeEdge.type cover the specified semantic vocabulary. AI/agent/system writes require provenance; mutation APIs record actor context and accept Idempotency-Key.
Evidence and hypotheses are separate. Deterministic scoring combines weighted support/contradiction with strength, freshness, reliability, first-party and behavioral/measured multipliers. The signed result maps to ten bands while confidence measures volume/diversity. contested, stale, and unresearched are overlays.
The filesystem NodepadStore uses atomic JSON writes and exclusive expiring leases. Legacy .nodepad browser behavior remains unchanged and a migration adapter maps old blocks and relationships into this graph.
Optional embeddings
See semantic layer for Ollama configuration, runtime-only cache, conservative duplicate aliases and source preservation, stable theme membership/centroids, and relevance prefilter behavior. Disabled or unavailable embeddings retain deterministic lexical processing. Similarity never assigns support/contradiction, scores, confidence, qualification, or brain promotion.
Audited reasoning
Research relations now pass through structured adjudication; task objectives do not assign support or contradiction. Adjudications persist with provenance and transition causality, while new synthesis proposals have no automatic validation evidence. Optional role routing and escalation preserve deterministic scoring and safe fallback.