Documentation > About > Bounded Learning

Bounded Learning

Skill refinement has strict guardrails to prevent runaway changes:

Parameter	Value	Purpose
Min corrections	3	Require consistent feedback before suggesting changes
Max change	20%	Prevent drastic alterations in a single refinement
Cooldown	7 days	Allow observation of changes before next refinement
User confirm	Always	Human in the loop for every change
Max cumulative drift	60%	Halt auto-refinements when skill has drifted too far from original (v1.10)
Contradiction detection	Auto	Flag contradictory feedback corrections before applying (v1.10)

Refinement Workflow

# 1. Check eligibility
skill-creator refine my-skill
# Output: "Eligible for refinement (5 corrections)"

# 2. Review suggested changes
# Shows: section, original text, suggested text, reason

# 3. Confirm or cancel
# "Apply these refinements? [y/N]"

# 4. Rollback if needed
skill-creator rollback my-skill

What Gets Refined

• Trigger patterns — Based on when skill activated vs. when it should have
• Content accuracy — Based on corrections you made to skill output
• Missing sections — Based on information you frequently added

Safety Philosophy

Bounded learning ensures the system improves through use while remaining under human control. Every refinement requires explicit user confirmation. The cumulative drift limit (60%) guarantees that no skill can change beyond recognition from its original form through automated refinement alone. If contradictory feedback is detected, the system flags it for human resolution rather than attempting to reconcile conflicting signals automatically.

Guardrail Interaction

The guardrails work together as a layered safety system:

• Min corrections (3) prevents knee-jerk reactions to a single bad interaction
• Max change (20%) ensures each refinement is incremental and reviewable
• Cooldown (7 days) provides observation time to detect regressions
• User confirm keeps a human in the decision loop at all times
• Max cumulative drift (60%) acts as an emergency brake on total accumulated change
• Contradiction detection prevents the system from applying conflicting refinements