intelligence-score

Intelligence Score

ACOS tracks its own effectiveness and learns from every session.

The 4-Component Score (25 points each, 100 total)

Component	Measured by	Scoring
Memory Depth	Trajectory count + memory entries	0-10 = 5pt, 11-50 = 10pt, 51-100 = 15pt, 101-200 = 20pt, 200+ = 25pt
Pattern Quality	patterns.json avg success rate	<50% = 5pt, 50-65% = 10pt, 66-75% = 15pt, 76-85% = 20pt, 86%+ = 25pt
Operational History	Total trajectories + tool diversity	<10 = 5pt, 10-30 = 10pt, 31-60 = 15pt, 61-100 = 20pt, 100+ = 25pt
Learning Velocity	Pattern count growth + success trend	Declining = 5pt, Flat = 10pt, Slight = 15pt, Strong = 20pt, Accelerating = 25pt

Score Display Format

═══════════════════════════════════════════════════
  ACOS Intelligence Report
═══════════════════════════════════════════════════

  Memory Depth        ████████████████████░░░░░  20/25
  Pattern Quality     ██████████████████████░░░  22/25
  Operational History ████████████████████████░  24/25
  Learning Velocity   ████████████████░░░░░░░░░  16/25

  ──────────────────────────────────────────────
  Total               ████████████████████░░░░░  82/100  Grade B+
═══════════════════════════════════════════════════

Grade scale: A+ (95-100), A (90-94), B+ (85-89), B (80-84), C+ (75-79), C (70-74), D (<70)

Data Sources

The intelligence score reads from:

• .claude/trajectories/patterns.json — learned n-gram tool sequences
• .claude/trajectories/ — individual trajectory files (count = operational history)
• .claude-flow/circuit-breaker.json — hook reliability signals
• .claude-flow/metrics/learning-status.json — learning system health

Pattern Learning

ACOS extracts n-gram patterns from successful sessions:

Tool sequence → "Read > Edit > Bash" (3-gram)
Outcome → success/failure
Context → task type, domain

Stored as:
{
  "sequence": "Read > Edit > Bash",
  "occurrences": 10,
  "successRate": 0.89,
  "context": "code_development"
}

Top patterns surface as routing recommendations on session start.

Trajectory System

Each significant session creates a trajectory file in .claude/trajectories/:

• Records the task type, tools used, success outcome
• Feeds pattern extraction on next session start
• Accumulates into institutional memory

Rules:

• Trajectories are append-only (immutable audit trail)
• Pattern extraction runs automatically on session start
• Patterns with <3 occurrences are not surfaced (insufficient evidence)

Continuous Improvement

Intelligence score improves over time through:

1. More trajectories → higher operational history
2. Better patterns → higher pattern quality (success rate improves as bad patterns are deweighted)
3. Voice activation accuracy → skill routing improves as it learns your phrases
4. Hook reliability → zero circuit breaker breaks = perfect hook score

Review weekly with /acos-score. Target: Grade A (90+) within 30 sessions.