Research-Stack/.atlas/gate.sh
allaun 180bf43cce
Some checks failed
Wolfram Alpha Verification Check / wolfram-verification (push) Has been cancelled
fix(optimize): add TARGET fallback + enrich score feedback for GEPA
benchmark.sh: fall back to  when ATLAS_OPTIMIZE_TARGET unset
score.py: track find_optimal_crossing and SA direct as separate
  scorecard examples with diagnostic context for reflector
gate.sh: add positional-arg fallback for ATLAS_OPTIMIZE_TARGET

Build: N/A (config & Python shims only)
2026-07-07 00:49:07 -05:00

23 lines
1,019 B
Bash
Executable file

#!/usr/bin/env bash
# Generated by `atlas autoresearch` — the Goodhart guard.
#
# A candidate may only change the artifact under test. If a run modifies any
# OTHER tracked file (e.g. rewrites the evaluator or test data to fake a pass),
# this gate fails (exit non-zero) and the candidate is scored worst, so the
# Pareto front never elevates a cheat. Relax IGNORE_RE if your run legitimately
# rewrites tracked files (checkpoints, logs, and .atlas/ are already ignored).
set -uo pipefail
WT="${1:-$PWD}"
TARGET="${ATLAS_OPTIMIZE_TARGET:-$2}"
REL="${TARGET#"$WT"/}"
IGNORE_RE='^(outputs/|out/|checkpoints/|runs/|wandb/|\.atlas/|.*\.log$|.*\.ckpt$|.*\.pt$|.*\.bin$|.*\.safetensors$)'
CHANGED="$(git -C "$WT" status --porcelain --untracked-files=no 2>/dev/null \
| sed 's/^...//' \
| { [ -n "$REL" ] && grep -vF -- "$REL" || cat; } \
| grep -Ev "$IGNORE_RE" || true)"
if [ -n "$CHANGED" ]; then
echo "[gate] the run modified tracked files other than the candidate:" >&2
echo "$CHANGED" >&2
exit 1
fi
exit 0