Skip to content

Latest commit

 

History

History
31 lines (23 loc) · 1.85 KB

File metadata and controls

31 lines (23 loc) · 1.85 KB

Grounded Evolution — Roadmap

Status: Alpha — single-node evolution loop operational.


Near-Term (v0.3 – v0.5)

  • Eval harness rewrite — parameterized benchmarks so the score is meaningful across runs
  • Population persistence v2 — SQLite backend instead of JSON for concurrent access
  • Execution cost tracking — log LLM tokens, wall time, and per-cycle cost to enable regression-aware pruning
  • Prompt diversity enforcement — deduplicate near-identical prompts before insertion
  • Multi-provider LLM support — graceful fallback if one provider is rate-limited
  • Explicit mutation hyper-parameters — expose mutation rate, crossover rate, tournament size as CLI/env overrides
  • Regressions tests in CI — run evaluator/ suite on every PR

Medium-Term (v0.6 – v1.0)

  • Parallel generation — generate + evaluate multiple child prompts per cycle
  • Archive of generated projects — retain every project artifact, not just the latest
  • Prompt migration — keep a rolling window of high-scoring prompts; archive the rest
  • Human-in-the-loop — optional approval gate before mutation is applied
  • Dashboard — simple web UI for monitoring cycles, viewing generated projects, comparing scores
  • Git-based experiment tracking — tag each commit with cycle metadata for reproducibility

Long-Term (v1.0+)

  • Distributed evolution — multiple agents running on different machines, sharing a common population DB
  • Cross-domain benchmarks — evolve prompts for code, documentation YAML, SQL, shell scripts
  • Automated meta-parameter tuning — treat mutation rate / temperature / selection pressure as evolvable parameters
  • LLM fine-tuning loop — fine-tune a small adapter using high-scoring prompts as training data