scaling-group/eve/implement-evaluation-steps

Implement Evaluation Steps

Overview

Evaluation is a first-class config tree: configs/eve/evaluation/<app>.yaml composed via the run config's defaults: list (e.g. - evaluation: circle_packing). The file declares an ordered evaluation.steps pipeline plus an evaluation.failure_score. Every run config MUST compose an evaluation: group; there is no implicit fallback.

evaluation.steps is an ordered, arbitrarily-long pipeline. Dispatch is by FORM, not a kind field:

• a bare .sh path string -> programmatic (deterministic) step
• a {prompt, immutable?} mapping -> LLM judge step (a first-class worker: its own prompt +

budget, plus an OPTIONAL immutable rubric + score-check subagent). Omit immutable for a lightweight judge that runs from its ENTRYPOINT alone.

Config form

# @package _global_
evaluation:
  steps:
    - configs/eve/evaluation/<app>/evaluation.sh        # programmatic step
    # - { prompt: configs/eve/evaluation/<app>/prompt_assess,
    #     immutable: configs/eve/evaluation/<app>/immutable_assess }   # judge step
  failure_score:
    score: 0.0
    summary: evaluation failed
  seed_solver_score: null
  seed_solver_skip_evaluation: false

Runtime Inputs (shell steps)

Shell evaluators receive these environment variables:

• EVE_WORKSPACE_ROOT: workspace root
• EVE_SOLVER_ROOT: solver candidate repo root, that is EVE_WORKSPACE_ROOT/solver
• EVE_EVAL_LOG_ROOT: formal evaluation log root, that is EVE_WORKSPACE_ROOT/logs/evaluate

Use these paths directly in the shell file instead of reconstructing ad hoc paths.

Requirements

• All shell steps run from EVE_WORKSPACE_ROOT, not EVE_SOLVER_ROOT or the downstream repo.
• Write final evaluation files to EVE_EVAL_LOG_ROOT. This is the only interface that will be extracted.
• EVE_EVAL_LOG_ROOT/score.yaml is required.

score.yaml rules:

• score.yaml must always exist.
• The score payload may be any YAML PyTree.
• If evaluation fails, write a minimal YAML payload such as:

score: 0.0
summary: evaluation failed

You can store files of any format within the EVE_EVAL_LOG_ROOT directory, including multimedia like images and videos. If your logging structure is complex, it is recommended to include a README.md inside EVE_EVAL_LOG_ROOT to document the file organization and contents.

Reference Example

circle_packing ships both forms:

• Default single-step: configs/eve/evaluation/circle_packing.yaml

pointing at configs/eve/evaluation/circle_packing/evaluation.sh.

• Judge pipeline: configs/eve/evaluation/circle_packing.judge.yaml

(performance.sh -> assess -> aggregate), with prompt/immutable dirs under configs/eve/evaluation/circle_packing/.