CatalystRL

Core Concepts

CatalystRL is built on three keystones that work together to create agents that improve themselves over time.

Trust Engine

Progressive capability unlocks based on demonstrated reliability. Agents earn trust through successful operations and lose it through failures.

  • Four trust levels: LOW, MEDIUM, HIGH, AUTONOMOUS
  • Capability-gated operations
  • Trust score persistence across sessions
  • Automatic trust updates based on outcomes

Evolution AI

Controlled experiments on skill configurations with statistical winner selection. Zero-downtime deployment of improvements.

  • A/B testing for skill variants
  • Statistical significance detection
  • Automatic rollout of winners
  • Rollback on performance regression

ABES

Autonomous Bounty Evolution System. Skills detect their own failures and create structured improvement requests that can be resolved without human debugging.

  • 9 bounty principles (P1-P9)
  • Pattern-based error detection
  • Evolver agents for resolution
  • Human-in-the-loop for complex issues

How They Work Together

1

Skill executes an operation

Trust Engine checks if the skill has permission for this operation type.

2

Operation succeeds or fails

Trust score updates based on outcome. Evolution AI records metrics.

3

On failure: ABES creates bounty

Pattern detected, context gathered, bounty created for evolver or human.

4

Bounty resolved, skill improved

Fix applied, skill retested, trust score recovers with successful runs.

Additional Concepts

Gates

Safety checkpoints that validate operations before execution. Four types: Processing Integrity, Escalation, Constraint, and Permission Recovery.

Memory Architecture

Four-layer memory system: Global (Claude), Cross-Session (~/.claude/memory), Repo (CLAUDE.md), and Session (ephemeral).

Skill Families

Related skills grouped under orchestrators. Examples: commit family (scan, stage, create, push), shutdown family (memory, services).

Offline First

All skills work without network. Platform features queue operations for later sync. No cloud dependency for core functionality.