The router was not the point↗ (external link)

The pitch is cost-aware routing: instead of guessing which frontier model should take a repository issue from the issue text alone, send a cheap 7B "searcher" into the repo first, let it write up what it found, then route on that.

The result is real. On the full Python slice of SWE-bench Pro — 266 tasks — SuperScout resolves 159 of 266 (59.77%) at $0.230 per solve, against Claude Opus 4.6 running solo at 158 of 266 (59.40%) for $1.274 per solve. Same solve rate, roughly a fifth of the cost. The searcher's own GPU time adds less than half a cent per task.

Then the ablation, which the authors report plainly and which quietly dismantles the framing: a variant with no router at all ties the routed system. The routing is not what is doing the work. The handoff document is.

So look at the handoff instead. It contains implicated files with line regions ranked by confidence, a reproduction attempt naming a file, a command and the observed output, the dead ends the searcher already tried, and free-form notes. And before any fixer sees it, "a sandbox replays the claimed reproduction command against the unpatched repository. A claim that does not genuinely fail is stripped: both the test file and the claim itself are removed from the handoff."

That sentence is the paper. A small model produces a plausible account of the bug; a sandbox checks which parts of that account are true; the unverified parts are deleted rather than passed along with a hedge. What reaches the expensive model is context that has been executed, not context that has been generated.

The authors are careful about what this buys. The handoff "redistributes rather than adds solving ability" — cheap fixers get lifted, the strongest one gets very slightly worse. Their calibration study is N=99, so per-fixer effects are "directional only." Nobody has shown that scouting makes a model smarter. They have shown that most of what separates a $1.27 solve from a $0.23 solve is knowing where to look, and that knowing where to look is cheap if you are willing to verify it.

The fixer pool, for the record, is Claude Opus 4.6, GPT-5.2, Gemini 3 Flash and Kimi K2.5, and adding a new one requires no retraining. That is the genuinely practical bit of engineering here, and it is not in the title either.