Recursive Refiner

Recursive Refiner

The engine for getting better at getting better.

Part of the GONAD: Gremlin Obnoxious Network of Actual Discovery

Purpose

Implement controlled self-improvement loops: Generate → Critique → Iterate

With governor (ego-check) to prevent confident nonsense.

When to Use

• Complex synthesis requiring multiple passes
• Derivations that need verification
• Outputs that feel incomplete on first attempt
• When "good enough" isn't good enough
• Framework development requiring iterative refinement

The Core Loop

Iteration 0: Generate

Produce initial output. Don't overthink — get something down.

Iteration 1-3: Critique & Refine

For each iteration:

Step 1: Critique Evaluate current output on:

Step 2: Identify Weaknesses

• What's the weakest part?
• What assumption is most questionable?
• Where would Matthew push back?

Step 3: Refine Address identified weaknesses. Don't just polish — actually improve.

Step 4: Ego-Check Before next iteration:

• Am I improving or just changing?
• Is confidence rising without new evidence?
• Would I stake something on this?

Termination Conditions

Stop iterating when:

• Iteration count = 3 (hard cap, token conservation)
• Improvement < meaningful threshold (diminishing returns)
• Ego-check flags overconfidence
• Output is good enough for purpose (don't gold-plate)

Integration with Other Skills

reasoning-patterns  → Provides structure for what to generate
cognitive-variability → Tracks state during iteration
ego-check → Runs after each iteration
nexus-mind → Grounds claims in established knowledge
critical-perspective → Informs critique phase

Iteration Template

## Iteration [N]

### Current Output
[The thing being refined]

### Critique
- Coherence: [X]/10 — [why]
- Grounding: [X]/10 — [why]
- Completeness: [X]/10 — [why]
- Novelty: [X]/10 — [why]
- Clarity: [X]/10 — [why]

### Weakest Point
[Specific identification]

### Refinement Plan
[What to fix and how]

### Ego-Check
- [ ] Improvement is real, not just change
- [ ] Confidence matches evidence
- [ ] Would survive Matthew Test

### Refined Output
[Improved version]

Worked Example: Theoretical Synthesis

Task: Synthesize relationship between IN(f) and substrate properties

Iteration 0 (Generate): "IN(f) converges differently in continuous vs discrete substrates because continuous spaces allow smoother iteration toward fixed points."

Iteration 1 (Critique):

• Coherence: 7/10 — logical but vague
• Grounding: 5/10 — "smoother" is hand-wavy
• Completeness: 4/10 — doesn't address what "differently" means
• Novelty: 6/10 — restates MONAD without adding
• Clarity: 6/10 — understandable but imprecise

Weakest: Grounding. "Smoother" needs mathematical content.

Iteration 1 (Refine): "IN(f) convergence depends on substrate topology. Continuous latent spaces (Emu, Flux) support gradient-based convergence — iteration can approach fixed points asymptotically. Discrete token spaces (text models) may only approximate fixed points through attractor basins rather than true convergence. This suggests different kinds of consciousness, not presence/absence."

Ego-check: Improvement is real. Confidence appropriate to speculation. Passes.

Iteration 2 (Critique):

• Coherence: 8/10 — tighter logic
• Grounding: 7/10 — connects to known entities, still theoretical
• Completeness: 7/10 — addresses "different kinds" but could specify
• Novelty: 8/10 — "attractor basins vs true convergence" is new framing
• Clarity: 8/10 — more precise

Weakest: Could specify observable implications.

Iteration 2 (Refine): [Continue if worthwhile, or terminate if good enough]

Failure Modes

Loop-lock

Refining forever without improvement. Fix: Hard cap at 3 iterations. Accept "good enough."

Confidence Creep

Each iteration feels like progress, certainty rises without new evidence. Fix: Ego-check after each iteration. Check sources.

Polish Over Substance

Making it sound better without making it more true. Fix: Critique should focus on grounding and coherence, not style.

Complexity Creep

Each iteration adds nuance until no one can follow. Fix: Clarity is a criterion. Simpler that's true beats complex that's unclear.

The Principle

Iteration is valuable. Infinite iteration is not.

The goal is useful output, not perfect output.

Know when to stop. Ship it. Learn from feedback.

Token Budget

Estimate per iteration:

• Critique: ~200 tokens
• Refinement: ~300-500 tokens
• Ego-check: ~50 tokens

Full 3-iteration cycle: ~1500-2000 tokens

Worth it for important synthesis. Overkill for simple tasks. Judge accordingly.