The most useful line in Charles Azam’s write-up isn’t that Fable 5 crushed an unpublished NP-hard optimization problem — it’s that the /goal mode “can win most runs and still be a bad default.” That gap between average win rate and worst-case behavior is where most agent evaluations quietly lie to you.

• 🎯 /goal isn’t a “try harder” switch — it changes the control loop and the search path. Sometimes it finds a better basin; sometimes it just gives a bad idea more time to mature.
• ⚠️ Same command, two architectures. Claude Code wires /goal to a separate evaluator; Codex backs it with persisted state and lifecycle tools. Identical surface, different failure modes underneath.
• 📊 Averaging hides variance. “Wins most runs” is exactly the metric that lets a directive compound a bad plan on the runs it loses — and those are the runs that page you at 2am.
• 🔍 Human baselines still matter. The author had a week of hand-written C++ to compare against. Without that anchor, “an absolute beast” is a vibe, not a measurement.
• ⚡ Consistency > peak. The standout wasn’t the best single solution but the low variance across runs — which is the property you actually schedule production agents on.

This maps straight onto anything with routing, retries, or goal-directed loops in production: a control-loop change has to be measured per-distribution and at the tail, never on aggregate pass rate. The HN discussion is worth skimming for the reproducibility notes.

So here’s the trap I’d watch for: how many teams have shipped a “reasoning” or “goal” toggle as a default because it won the A/B on mean score, without ever looking at what it does to the p99 run?