eralisted

# ERA — empirical-software search via Flat UCB Tree Search Faithful port of Google Research's **ERA** ([Nature 2026](https://www.nature.com/articles/s41586-026-10658-6), [repo](https://github.com/google-research/era)): an AI system that helps write high-quality empirical software. An LLM proposes whole candidate **programs**, a sandbox **scores** each, and a flat **PUCT bandit** (FUTS) decides which existing solution to improve next — keeping a *population* and escaping local optima instead of greedily climbing one trajectory. The candidate **is the source code**; ERA evolves the algorithm/model/pipeline, not a config. **What ERA is for:** optimizing a problem that already has (a) a runnable seed program, (b) a fast automated scalar score, (c) a fixed problem definition. It evolves the code *within* that frame — it does not invent the frame. Skip ERA for problems with no automated score, or a single known change (just make it). ## The engine (`scaffold/futs.py` — FROZEN, vendored verbatim) `futs.py` is vendored **verbatim** from `google-research/era` (`implementation/futs.py`, commit `eb56676ce0a9b22b0c357c218de09ac61e94fc4b`, Apache-2.0). It is the neutral judge — **never edit it during a search** (that invalidates all node scores and cross-run comparison). Primitives: ```python Problem(description) # the problem statement + interface + scoring contract Solution(program) # a complete, self-contained program (source text) Node(index, parent_index, solu