scripts/make_jobs.py

"""Build the evaluation job list, ordered so the barrier-crossover region is answered first.

Priority 0  parents (shared across every pair, so they are never recomputed)
Priority 1  naive (alpha=0.5) merges for pythia pairs near the half-decade crossover
Priority 2  near anchors (<0.3 decades) and far anchors (>1.5 decades)
Priority 3  the rest of the pythia pairs
Priority 4  zh-pythia pairs
"""
import json, pandas as pd, numpy as np

SRC = "/root/crossarch-diagnostics/results/checkpoint_pairs.csv"
OUT = "/root/crossarch-accuracy/results/jobs.jsonl"
d = pd.read_csv(SRC)

jobs = []
# parents
for fam, g in d.groupby("family"):
    for s in sorted(set(g.step_a) | set(g.step_b)):
        jobs.append(dict(key=f"{fam}:parent:{s}", prio=0 if fam == "pythia" else 4, fam=fam, kind="parent", step=int(s),
                         tasks="core_zh" if fam == "zhpythia" else "core"))

def prio(fam, r):
    if fam == "zhpythia": return 4
    x = r.log10_step_ratio
    if 0.25 <= x <= 1.10: return 1          # the crossover window
    if x < 0.25 or x >= 1.50: return 2      # anchors on both sides
    return 3

for _, r in d.iterrows():
    fam = r.family
    jobs.append(dict(key=f"{fam}:merge:{int(r.step_a)}:{int(r.step_b)}:0.5", prio=prio(fam, r),
                     fam=fam, kind="merge", step_a=int(r.step_a), step_b=int(r.step_b),
                     alpha=0.5, pair=r.pair, log10_step_ratio=float(r.log10_step_ratio),
                     tasks="core_zh" if fam == "zhpythia" else "core"))

jobs.sort(key=lambda j: (j["prio"], abs(j.get("log10_step_ratio", 0) - 0.6)))
with open(OUT, "w") as f:
    for j in jobs: f.write(json.dumps(j) + "\n")
print("jobs", len(jobs), {p: sum(1 for j in jobs if j["prio"] == p) for p in range(5)})