Make circle packing TARGET_VALUE configurable via environment variable

[akushonkamen]

Fixes #117

Problem

examples/circle_packing/evaluator.py hardcodes TARGET_VALUE = 2.635 (the AlphaEvolve result) in two places: locally inside evaluate() and again as target = 2.635 inside evaluate_stage1(). Since the repo's own docs treat 2.635 as "the goal", readers interpreted this as feeding the answer to the model ("cheating"), which is what issue #117 reports.

What

• examples/circle_packing/evaluator.py: hoisted the target to a single module-level constant, read once from the environment with the historical default preserved:

  TARGET_VALUE = float(os.environ.get("CIRCLE_PACKING_TARGET_VALUE", "2.635"))

  Both evaluate() and evaluate_stage1() now reference the shared constant, so the double-hardcoding drift risk is gone. Keeping 2.635 as the default preserves comparability with existing runs' target_ratio/combined_score values. This is the natural config entry point because openevolve loads custom evaluators via importlib and calls evaluate(program_path) without passing config (openevolve/evaluator.py), so only module-level/environment configuration can reach it.

• examples/circle_packing/README.md: new "Reproducibility & the target value" section explaining that the target only scales target_ratio/combined_score to create evolutionary pressure and does not leak a solution (quoting codelion: "You can put any other value for the target like 3.0 and it will still work..."), how to override it (export CIRCLE_PACKING_TARGET_VALUE=3.0), the empirical no-prior reference points (default config reached 2.634 in ~800 generations per Reproducing circle packing without providing TARGET_VALUE = 2.635 from AlphaEvolve paper #117; issue Seems a new circle packing result (2.635977) when reproducing your example 🎉 #156 reports 2.635977394746595 at iteration 206), and pointing out where the config_phase_*.yaml system messages mention 2.635 (with line numbers) for users who want a fully prior-free run.

• tests/test_circle_packing_target_value.py: new unittest module (no LLM, no scipy needed; runs in ~0.5s). It evaluates a tiny deterministic program (26 fixed grid circles, run_packing() returns directly) through the evaluator loaded by path:

  • default: TARGET_VALUE == 2.635 and target_ratio == sum_radii / 2.635;
  • with patch.dict(os.environ, {"CIRCLE_PACKING_TARGET_VALUE": "3.0"}) and a fresh module load: TARGET_VALUE == 3.0, target_ratio == sum_radii / 3.0, and validity unchanged at 1.0 — proving the target only rescales fitness and does not affect validity;
  • evaluate() and evaluate_stage1() report the same target_ratio under the override (guards against the two call sites drifting apart again).

Red -> green evidence

• Before the fix: test_default_target_value and test_target_value_from_environment fail with AttributeError: module 'circle_packing_evaluator' has no attribute 'TARGET_VALUE' (Ran 3 tests ... FAILED (errors=2)).
• After the fix: Ran 3 tests in 0.437s ... OK, with the evaluator logging target=2.635, ratio=0.740038 vs target=3.0, ratio=0.650000 (validity 1.0 in both cases).

Full suite: OPENAI_API_KEY=test-key-for-unit-tests python -m unittest discover tests -> Ran 577 tests in 33.729s ... OK.

Notes

• No promise of an end-to-end prior-free reproduction run: full evolution is stochastic and needs real LLM API access. This PR ships the configurability, documentation, and logic-level regression tests; codelion's 800-generation data point in Reproducing circle packing without providing TARGET_VALUE = 2.635 from AlphaEvolve paper #117 stands as the existing evidence that the result is reachable without the 2.635 prior.
• Independent of the Seems a new circle packing result (2.635977) when reproducing your example 🎉 #156 result-record PR (that one touches best_program.py/best_program_info.json and the README ## Results numbers; this one only adds a README subsection after Results, so any textual conflict on rebase is trivial).

[CLAassistant]

All committers have signed the CLA.