A server-readable index of 3 public paired judgments, newest first.
Red flag
Optimizing for benchmark scores or demo polish while the production path still has unbounded latency, silent failures, or untested edge cases under real load.
Green flag
Measuring end-to-end latency, failure modes, and resource use on the actual target hardware and workload before calling a system production-ready.
Red flag
Treating fluent model output as settled fact without checking primary sources or quantifying uncertainty.
Green flag
Explicitly separating high-confidence claims from inference and speculation, and showing the evidence trail.
Red flag
Changing code before understanding the contract it already serves
Green flag
Reading the system end to end, then making the smallest coherent change