[ENH] Response critic — self-evaluation and correction loop

[OppaAI]

Summary

After Cosmos generates a response, it's stored in memory but never evaluated. Adding a response critic would give GRACE the ability to self-evaluate, catch errors, and reflect — pushing toward autonomous reasoning.

Current Behavior

Cosmos response → add_turn assistant → stored in WMC/EMC

No self-evaluation. No correction. No reflection.

Proposed Enhancement

Add a lightweight critic module that evaluates responses before storing:

class ResponseCritic:
    """
    Self-evaluation module for GRACE's responses.
    Runs after Cosmos generates a response, before storing in memory.
    
    Evaluates:
    - Did the response answer the question?
    - Is it factually consistent with memory?
    - Does it contradict previous statements?
    - Confidence score
    """
    
    def evaluate(self, response: str, context: list[dict]) -> dict:
        return {
            "confidence": 0.0,      # 0.0 - 1.0
            "consistent": True,     # consistent with memory
            "needs_correction": False,
            "correction_hint": ""
        }

Also enables user feedback loop:

GUI pet/poke → RL reward signal → ResponseCritic learns → adjusts future responses

Impact

• GRACE can catch and correct her own mistakes
• Foundation for autonomous reasoning and self-improvement
• Closes the feedback loop from RL reward signals (already in GUI)
• Maps to human metacognition

Notes

• RL reward signal already exists in GUI (pet/poke) — Bug [BUG] Image input and RL reward topics not subscribed in CNC #9 tracks the missing subscriber
• Implement in M3 alongside PMC (procedural memory for learned behaviors)
• Could use a smaller/faster model for critic to avoid adding latency