REPROCESS EDGES

Re-evaluate world model edges when the underlying data improves. Deletes edges below a confidence threshold and re-runs their compiled skill patterns for fresh scores. No model calls. No tokens.

The LLM compiled each skill once. REPROCESS EDGES runs the same compiled patterns again — against updated world model data — and produces new edges with recalculated confidence scores.

Syntax#

REPROCESS EDGES WHERE confidence < <threshold>

Example#

REPROCESS EDGES WHERE confidence < 0.7

This:

1. Finds all derived edges with confidence < 0.7
2. Identifies the skill that produced each edge (via skill_id provenance)
3. Deletes the stale edges
4. Re-executes the compiled pattern for those node pairs
5. Materializes new edges with fresh confidence scores from current world model data

When to use it#

The world model is live. As new sensor observations arrive, existing derived edges can become stale — they were scored against older data. REPROCESS EDGES closes that loop:

-- New sensor data arrived with higher reliability readings
-- Refresh any co-occupancy edges that were scored below threshold
 
REPROCESS EDGES WHERE confidence < 0.75
-- => 12 ZONE_PEER edges re-evaluated
-- => avg confidence improved from 0.68 to 0.84

Scoped reprocessing#

Target a specific skill or relationship type:

-- Only refresh co-detection edges
REPROCESS EDGES WHERE confidence < 0.8
  AND type = 'CO_DETECTED'
 
-- Only refresh edges from an older skill version
REPROCESS EDGES WHERE skill_version < 3

How confidence improves without a new model call#

The compiled pattern scores edges from properties already in the graph — sensor reliability, detection confidence on existing relationships, zone hazard levels. When those upstream values improve (a sensor is recalibrated, a detection is corroborated by a second sensor), re-running the same pattern against the updated data produces a higher score.

The LLM authored the scoring logic. The graph engine executes it against current values.

Sensor s1 recalibrated: reliability 0.72 → 0.94
                  ↓
REPROCESS EDGES WHERE confidence < 0.75
                  ↓
  ZonePeerDetector compiled pattern re-runs
  Scores edge (Atlas-01, Rover-02) using s1.reliability = 0.94 now
  New confidence: 0.61 → 0.89

Relationship to skill versions#

REPROCESS EDGES re-runs the current compiled pattern of each skill. If you want the re-evaluation to use a new rule, create a new skill version first:

-- Recompile with a richer rule (one LLM call)
CREATE SKILL ZonePeerDetector_v2
  FROM PROMPT 'Link two Robot nodes as ZONE_PEER if they share a zone
               AND combined detection confidence exceeds 0.80'
  ALLOWED ON [Robot]
  TIER SYMBOLIC
 
-- Refresh all weak edges with the new compiled pattern (zero LLM)
REPROCESS EDGES WHERE confidence < 0.75

Old edges carry skill_version: 1. Refreshed edges carry skill_version: 2. Both are traceable in the provenance chain.

See Also#

• CREATE SKILL — where the LLM compiles your rule (once)
• PROCESS NODE — initial materialization of derived edges
• Confidence & Provenance — how edges carry scores and evidence chains
• Concepts: Skills — the compile-once model explained