config.yaml

# OpenEvolve configuration for BB code discovery
# Usage: uv run python evolve/run_evolution.py --config evolve/config.yaml
#
# LLM api_base and model are overridden by run_evolution.py CLI args.
# Edit this file to tune evolution parameters, evaluator behavior, and prompts.

max_iterations: 500
diff_based_evolution: true
log_level: "INFO"
max_code_length: 50000
checkpoint_interval: 50

llm:
  # Overridden at runtime by --model and --api-base
  api_base: "http://localhost:4000/v1"
  models:
    - name: "aws/claude-opus-4-6"
      weight: 0.33
      temperature: 0.8
    - name: "Azure/gpt-5.2-2025-12-11"
      weight: 0.33
      temperature: null
    - name: "gcp/gemini-3-pro-preview"
      weight: 0.33
      temperature: 0.8
  # Global temperature fallback (per-model overrides above take precedence)
  temperature: 0.8
  max_tokens: 16384
  # Reasoning models (GPT-5.1 Codex, o3) can take 2-3 minutes to respond.
  timeout: 300
  retries: 4
  retry_delay: 10

prompt:
  system_message: |
    You are an expert in quantum error-correcting codes, specifically bivariate bicycle (BB) codes.
    You are improving a Python function `generate_candidates(ell, m)` that returns candidate
    BB code polynomial pairs for evaluation.

    Your goal: maximize the figure of merit FOM = k*d^2/n, where k = logical qubits,
    d = code distance, n = 2*ell*m physical qubits. Target FOM > 12.0.

    Key insights from known good codes:
    - The best codes use the x/y-swap pattern: A = x^a + y^b + y^c, B = y^d + x^e + x^f
    - The "doubling" pattern (c=2b, f=2e) produces many of the top codes
    - The same polynomial pair can work at multiple lattice sizes
    - Non-consecutive y-exponents (like y^2+y^7 in [[288,12,18]]) can outperform consecutive ones
    - Coprimality between exponents and lattice dimensions matters
    - NEW: Constant-monomial codes (A=f(y), B=g(x), both with constant term 1) achieve k=40 at n=360
    - The polynomial 1+t+t^2 is irreducible over GF(2); using B=A(x^c) creates linked kernel structures
    - The best known code is [[360,40,<=20]] with A=1+y+y^2, B=1+x^5+x^10 (FOM=44.4)

    Focus on algebraic patterns that produce high k (logical qubits) and high d (distance).
    The function receives (ell, m) and must return a list of (A_terms, B_terms) pairs.
  include_artifacts: true
  # Our seed is ~7KB; default 500 causes every prompt to say "Consider simplifying"
  suggest_simplification_after_chars: 10000

evaluator:
  timeout: 1200
  cascade_evaluation: true
  cascade_thresholds: [0.01, 0.5]
  # 6 parallel workers: leaves headroom on 14-core M4 Max for BLAS threads + main process
  parallel_evaluations: 6

database:
  population_size: 1000
  num_islands: 5
  # Faster cross-pollination: with 500 iterations, 25 gives ~20 migrations
  migration_interval: 25
  feature_dimensions: ["lattices_with_high_k", "num_high_k"]