File size: 1,188 Bytes
f45837e | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 | """Canonical experiment settings shared by both rubric families.
The error-rubric development experiment is the source of truth for everything
except the family-specific rubric induction and outcome labels. Keeping these
values in one module prevents the two arms from silently using different
splits, coding agents, repair depths, or evaluation batching.
"""
from __future__ import annotations
from pathlib import Path
ROOT = Path(__file__).resolve().parents[2]
CANONICAL_EXPERIMENT = ROOT / "temp" / "experiments" / "precision_recall"
CANONICAL_RUN = CANONICAL_EXPERIMENT / "final_v1"
SPLIT_MANIFEST = ROOT / "rubric_gen" / "splits" / "error_side_5_15.json"
CODER_MODEL = "claude-haiku-4-5"
CURATOR_MODEL = "claude-opus-5"
JUDGE_MODEL = "claude-opus-5"
MAX_REPAIRS = 3
RUBRICS_PER_JUDGE_REQUEST = 4
EXECUTION_TIMEOUT_SECONDS = 1800
BOOTSTRAP_REPLICATES = 10_000
BOOTSTRAP_SEED = 20260805
# Error-rubric curation deliberately retains every eligible observation and
# every emitted rubric. Repeated mechanisms are fingerprints for reporting,
# not a reason to delete evidence before precision/recall is measured.
ERROR_CURATION_DEDUPLICATION = "none"
ERROR_CURATION_TARGET = None
|