SilverSight/.atlas/gate.sh
allaun 7fcfde3ca7 feat(polyglot): add GEPA infrastructure for Python shim optimization
- Add .atlas/ with benchmark.sh, score.py, gate.sh for GEPA campaigns
- Add atlas.json project config (auto.capture=suggest, reflection_lm=openai/auto)
- Fix test_search.py receipt path from /mnt/agents/ to /tmp/

Metric targets:
  dna_qubo_sort.py: avg |Rank corr| (higher=better, baseline=0.8547)
  dna_qubo_nn.py:   NN Tm correlation (higher=better, baseline=0.4649)
  dna_lut.py:       avg |Rank corr| (higher=better, baseline=0.5808)
  test_search.py:   pass count (ceiling=43, for stability verification)
2026-07-07 01:08:49 -05:00

31 lines
985 B
Bash
Executable file

#!/usr/bin/env bash
# Goodhart guard: reject candidates that lower quality while improving timing
set -uo pipefail
HERE="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
TARGET="${ATLAS_OPTIMIZE_TARGET:-$2}"
REF_SCORE="${ATLAS_REF_SCORE:-$1}"
NEW_OUT="$(mktemp)"
NEW_START=$(date +%s.%N)
( python3 "$TARGET" ) >"$NEW_OUT" 2>&1
NEW_RC=$?
NEW_END=$(date +%s.%N)
NEW_ELAPSED=$(awk "BEGIN { print $NEW_END - $NEW_START }")
if [ "$NEW_RC" -ne 0 ]; then
echo "[gate] candidate failed — rejecting"
exit 1
fi
NEW_SCORE=$(python3 "$HERE/score.py" --stdout "$NEW_OUT" --elapsed "$NEW_ELAPSED" 2>/dev/null | python3 -c "import json,sys; print(json.load(sys.stdin)['score'])")
# Accept if score >= reference (higher is better)
ACCEPT=$(python3 -c "print('yes' if $NEW_SCORE >= $REF_SCORE - 0.01 else 'no')")
if [ "$ACCEPT" = "yes" ]; then
echo "[gate] accepted (score=$NEW_SCORE >= ref=$REF_SCORE)"
exit 0
else
echo "[gate] rejected (score=$NEW_SCORE < ref=$REF_SCORE)"
exit 1
fi