State the conclusion responsibly

Combine the measurements

JEV and KEV completed all 300 cases and differed by one correct decision. JEV had lower observed HTTP latency; KEV achieved similar accuracy locally. Seventeen disagreements, nineteen shared errors and different probability profiles reveal tradeoffs hidden by total accuracy. CLM always returned Hold; GLM combined high completed-answer accuracy with 33 incomplete outputs.

Those observations support comparisons of the tested configurations under the frozen protocol. They supply no general winner, causal architecture ranking or deployment safety certification.

Follow the evidence trail

The versioned deposit preserves reviewed references, protocol, original responses, failures and analysis. Rebuilding analysis uses the recorded observations. Calling models again is a new experiment and may return different hosted outputs. Paper text, data, documentation and figures use CC BY 4.0; original analysis code uses MIT; third-party rights remain separate.

Codex assisted drafting and implementation; Junior reviewed the key and retained responsibility for the manuscript. Later JEV editorial judgments and other AI reviews were advisory and separate from scored inference. JEV’s editorial assistance is not independent validation of its own performance.

The author disclosed no external funding and no financial, employment or advisory relationship with the vendors named in the paper. Cases were hypothetical and no corresponding real actions occurred. The deposit excludes credentials, model weights and unrelated private material.