mirror of
https://github.com/allaunthefox/Research-Stack.git
synced 2026-08-19 00:10:35 +00:00
180bf43cce
Some checks failed
Wolfram Alpha Verification Check / wolfram-verification (push) Has been cancelled
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)
30 lines
1.3 KiB
Bash
Executable file
30 lines
1.3 KiB
Bash
Executable file
#!/usr/bin/env bash
|
|
# Generated by `atlas autoresearch` — this file is YOURS, edit it freely.
|
|
#
|
|
# Contract: write {"score": <float>, "examples": [...]} to $ATLAS_OPTIMIZE_RESULT.
|
|
# The candidate artifact has already been written into this worktree at
|
|
# ${ATLAS_OPTIMIZE_TARGET} (the repo-relative path: 4-Infrastructure/shim/braid_search.py)
|
|
# so the only thing a candidate can change is that file — the optimizer runs
|
|
# every candidate in its own throwaway git worktree, which keeps the evaluator
|
|
# and data pinned. score.py turns the run's output into the score; .atlas/gate.sh
|
|
# is the Goodhart guard.
|
|
set -uo pipefail
|
|
HERE="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
|
|
OUT="$(mktemp)"
|
|
|
|
# --- the repo's run command (auto-detected) --------------------------------
|
|
# Edit this line if the campaign should run something different.
|
|
TARGET="${ATLAS_OPTIMIZE_TARGET:-$1}"
|
|
( python3 "$TARGET" ) >"$OUT" 2>&1
|
|
RC=$?
|
|
# ---------------------------------------------------------------------------
|
|
|
|
if [ "$RC" -ne 0 ]; then
|
|
echo "[benchmark] run command exited $RC — candidate failed" >&2
|
|
tail -n 40 "$OUT" >&2
|
|
exit "$RC"
|
|
fi
|
|
|
|
# Parse the optimized metric out of the run output (edit .atlas/score.py to
|
|
# change which number is read, or to emit richer per-example feedback).
|
|
python3 "$HERE/score.py" --stdout "$OUT"
|