GLOSSARY.md

Trio Enterprise — Metrics Glossary

Standard definitions aligned with Frigate NVR / Verkada / industry conventions.

Core Concepts

Detection  = YOLO finds a bounding box in one frame (raw, high-frequency)
Tracking   = ByteTrack links detections across frames into a trajectory (per-object)
Event      = A tracked object's lifecycle (enter → dwell → exit) that meets criteria
Insight    = Multi-signal intelligence from Synthesis layer (the "wow" output)

Detection ≠ Event ≠ Insight. A busy scene has thousands of detections, dozens of events, and maybe 1-2 insights per hour.

Per-Layer Metrics

Layer 1: Perception

Metric	Definition	Unit
frames_processed	Total frames read from stream	count
motion_detected	Frames where MOG2 found significant contours	count
motion_rate	motion_detected / frames_processed	%
hash_skipped	Frames rejected by perceptual hash (identical)	count
stream_status	connected / reconnecting / stalled	enum

Layer 2: Extraction

Metric	Definition	Unit
yolo_triggers	Times YOLO ran (motion confirmed)	count
vlm_calls	Times VLM ran (scene changed)	count
vlm_skip_rate	Frames where YOLO ran but VLM skipped (no change)	%
avg_vlm_latency	Mean VLM inference time	seconds
detections	Total bounding boxes produced by YOLO	count

Layer 3: Memory

Metric	Definition	Unit
entities_learned	Unique entities in semantic memory	count
vehicles_known	Known vehicles (regular + unknown)	count
vehicles_regular	Vehicles seen ≥3 times (trusted)	count
persons_known	Known persons	count
episodes	Events stored in episodic memory	count
memory_actions	ADD/UPDATE/DELETE decisions by memory agent	count

Layer 4: Synthesis

Metric	Definition	Unit
baseline_coverage	% of (hour × dow) slots with ≥10 samples	%
baseline_slots	Number of (camera, hour, dow, metric) baselines	count
insights_generated	Insights produced (anomaly + dwell + pattern)	count
insights_high	High/critical severity insights	count
patterns_discovered	Recurring patterns found by FP-Growth	count
active_dwells	Entities currently being tracked for dwell	count

System

Metric	Definition	Unit
pipelines_running	Active pipeline processes	count
pipelines_total	Registered cameras	count
ram_used_gb	System RAM used	GB
ram_free_gb	System RAM available	GB
cpu_idle_pct	CPU idle percentage	%

Status Indicators

Status	Color	Meaning
active	Green	Pipeline running, producing events
learning	Yellow	Baseline building, <10 samples per slot
ready	Green	Baseline ready, insights being generated
idle	Gray	Pipeline running but no motion detected
offline	Red	Pipeline not running
stalled	Red	Stream connected but no frames received

Event vs Insight

Event (what happened)

{
  "type": "vehicle_arrived",
  "timestamp": "2026-03-26T14:32:00Z",
  "camera_id": "cam_home_front",
  "description": "Black Tesla Model Y SUV parked in driveway",
  "people_count": 0,
  "vehicle_count": 3,
  "severity": "info"
}

Insight (what it means)

{
  "type": "security",
  "severity": "high",
  "confidence": 0.85,
  "narrative": "Dark Gray SUV appeared 4th time during quiet hours (22:00-06:00).
                Average dwell: 2.3h. NOT a regular vehicle. Recommend investigation.",
  "signals": ["frequency: 4x night", "dwell: 2.3h", "regular: false"],
  "actionable": "Investigate"
}