Task 15.4 reuses the admitted Task 13.4 supervised load receipt. It reports five operational families and 900 verified outcomes. The billed model family contains 100 exact-one-tool successes with TTFT, total latency, input/output tokens, iterations, provider errors, billed USD per verified outcome, and CPU/RSS resource proxies. The four local families retain latency, errors, integrity floors, and resource proxies; they are unpriced, not “free.”
Cache reads/writes and calibrated energy were unavailable and remain null with explicit limitations. No optimization is admitted: this is a baseline, not a paired challenger, so outcome, safety, privacy, accessibility, and reliability floors cannot yet be compared.