scaling-group/eve/run-math-proof-quickstart

Run this skill from the repository root. A Math Proof quickstart is a real local attempt to solve the user's problem, not a smoke test.

Do not run Git commands as part of this skill. Do not git add, git commit, git push, git checkout, or git stash. Quickstart application configs use application.path so the runner can read local task files directly.

Create The Task Pipeline

1. Choose a stable <task_slug> using lowercase letters, digits, and

underscores. Derive <task_slug_hyphen> by replacing underscores with hyphens. For example, infinitely_many_primes becomes infinitely-many-primes. Quickstart copies are intentionally local scratch files, so place the task under examples/tmp/.

1. Refuse to overwrite an existing task unless the user explicitly asks. Check:

test ! -e examples/tmp/math_proof_<task_slug>

1. Copy the template task:

cp -R examples/math_proof/_quickstart examples/tmp/math_proof_<task_slug>

1. Replace examples/tmp/math_proof_<task_slug>/problem/problem.md with the

user's problem. Keep proof/.gitkeep.

1. Check that the copied task files no longer contain template references:

rg "math_proof_quickstart|math-proof-quickstart|examples/math_proof/_quickstart|<task_slug>|<title>" \
  examples/tmp/math_proof_<task_slug>

Any remaining hit must be intentional. Fix stale template references before launch. This generated quickstart task path is ignored by git.

Launch

Run the quickstart config with task-specific overrides:

uv run python -m scaling_evolve.algorithms.eve.runner \
  --config-name=math_proof_quickstart \
  application.name=math-proof-<task_slug_hyphen> \
  application.path=examples/tmp/math_proof_<task_slug> \
  label=math-proof-<task_slug_hyphen>

Supervise And Report

After launch, supervise the live run with supervise-run skill. Use inspect-population skill to read the completed population, evolved proof files, and score artifacts.

For the final quickstart report, include:

• config name and exact launch command;
• run root;
• completed iteration count;
• whether the run completed, stalled, failed, or was interrupted;
• best candidate proof paths under solver_workspaces/*/solver/proof/;
• raw evaluation dimensions from score artifacts;
• evaluator, runner, boundary, or repair failure summaries;
• any useful next action, such as resume the same run or inspect a specific

candidate proof.

Do not invent a scalar score for Math Proof. Preserve the vector dimensions and state any presentation ordering as an operator view, not a stored score.