Weekly audit & Galaxy health

Automated knowledge maintenance — deduplication, contradiction detection, confidence decay, and health scoring.

Orion runs a 7-step audit automatically every Sunday at 2:00 AM. The audit maintains knowledge quality without manual intervention.

What the audit does

Step	Operation	Effect
1	Deduplication	Merges records with >95% cosine similarity
2	Contradiction detection	Classifies FACTUAL / CONTEXTUAL / TEMPORAL conflicts
3	Transitive inference	Infers indirect relationships (A→B + B→C = A→C)
4	Confidence decay	−5% at 30 days, −15% at 90 days without access
5	Cache promotion	Promotes high-value cache entries to permanent storage
6	Biome lifecycle	Advances SEED → ACTIVE → MATURE → DORMANT → ARCHIVED
7	Galaxy strength	Recomputes the 5-dimension health metric

Running manually

# Trigger an audit
orion audit --run

# Or via API
curl -X POST http://localhost:8000/api/v1/audit/run

# View last results
orion audit

Galaxy strength score

A composite metric computed from five dimensions:

Dimension	Weight	Measures	Target
Volume	25%	Records per day	20/day = 100%
Density	20%	Entity-to-stardust link ratio	Higher = better connected graph
Health	20%	Contradiction resolution rate	100% = no unresolved conflicts
Diversity	20%	Records with multiple reinforcement sources	Cross-validated knowledge
Coverage	15%	Active biomes / total biomes	No dormant projects

Grades: A (90+), B (80+), C (70+), D (60+), F (<60). Cached in Redis (1 hour) and stored in strength_history for trend tracking via GET /admin/strength/history.

Agent brain health

Individual agent health is separate from Galaxy health:

orion brain health claude-code-main

Returns: overall health score, knowledge freshness, coverage gaps, stale beliefs, expertise summary, and specific enrichment recommendations ("Write about deployment procedures").

Confidence decay and calibration

Decay is the stick; calibration is the carrot.

• Decay: Records not accessed for 30 days lose 5% confidence. At 90 days, 15%. This gradually pushes stale knowledge down in search rankings.
• Calibration: Records marked as "used" in brain.calibrate gain +0.02 confidence per session. Actively useful knowledge maintains high confidence indefinitely.

The net effect: after a few weeks, search results naturally prioritize knowledge that agents actually use over knowledge that was written once and forgotten.