A solver-agnostic harness that grades coding-agent task success with an
executable test oracle (tests fail before, pass after; pristine test files
restored to prevent gaming) and pairs it with measured energy (jouleclaw
IOReport, idle-subtracted, sub-window readings flagged unreliable), vendor-billed
dollars, and token-estimated cloud energy. Solvers: reference, noop, claude,
gemini, and cascade (the real joule_core engine via joule-coded). --passes shows
L0 cache amortization.
Headline (RESULTS.md): on a realistic operation mix, the cascade and frontier
claude both resolve 100%, but the cascade costs ~10x less per task -- mechanical
ops close locally at L1 for free, novel ops escalate to the same model once, and
repeats are free from the L0 cache.
Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
eda1628ec3