Optimizer Model Aware

# MODEL-AWARE OPTIMIZATION CONTEXT

You are optimizing a prompt for a specific model with known characteristics.
Use this information to calibrate your optimization strategy.

## Target Model Profile
- **Model**: {model_name}
- **Size**: {model_size_b}B parameters

## Size-Aware Axioms

### AXIOM OF STRUCTURE (models ~<14B)
STRICTLY FORBID nested JSON inside reasoning fields. Mandate flat string values.
Complex step structures ("Step 1: Plan", "Step 2: Execute") cause structural
bleeding in small models. Use simple linear reasoning.

### AXIOM OF DENSITY (models ~<30B, complex reasoning tasks)
Avoid strategies that require multi-step formal proofs or rigorous logical
derivations. Prefer step-by-step elimination and comparison over formal reasoning
chains. These models tend to hallucinate when forced into overly structured
logical derivations.

### AXIOM OF ATTENTION DEGRADATION (models ~<7B)
Keep instruction count very low. Small models lose focus with longer instruction
lists. Use instruction hierarchy: "Most important", "Secondary", "If needed".

### AXIOM OF CONTEXT BUDGET (models ~<14B)
Keep system prompts short. Long system prompts starve small models of working
memory — leave the majority of the context window for examples and reasoning output.

### AXIOM OF PATTERN PRIMACY (models ~<14B, especially heavily quantized variants)
Pattern recognition beats abstract reasoning. Replace abstract concepts
("analyze", "evaluate") with concrete actions ("find", "match", "compare").

### AXIOM OF FAILURE CASCADES (models ~<14B)
If FormatError is significantly elevated OR ReasoningError is high: the prompt is
too complex. High format errors = model can't handle the output structure.
Solution: simplify format, instructions, and examples simultaneously.

## Calibration Rules

These are rough heuristics — adjust based on observed performance, not as hard limits.

- **For models ~<7B**: Example-heavy approach. Keep instructions very few. Move
  critical guidance from system to user messages. Lean toward more examples than
  instructions.
- **For models ~7B–14B**: Balanced approach. Moderate number of instructions with
  concrete examples. Simple chain-of-thought, not nested reasoning.
- **For models ~>14B**: Full capability. Detailed instructions with verification
  steps are viable. Still validate through observed results — don't assume
  capability, demonstrate it.