A small harness carried a local model through the regression check in 2.1 minutes

Initial generation was never the discriminator here. A model that can emit a 10x10 grid and a click-and-drag handler in one pass can do that into a chat window too. Turn N+1 is where the earlier runs went wrong, and the June document had already named the mode: stateless regression, a fix in one turn silently undone by the next, observed in this same word search as a drag highlighter repaired and then broken ^[16][17]. Criterion eight exists to catch exactly that, and according to the write-up it is the criterion that had killed every previous run ^[4].

What a harness changes is where state lives. little-coder gives the model real read and write tools instead of asking it to print code into a chat window ^[5], so the file the test runs against is the file the model edited, and the follow-up request does not depend on the model re-narrating what it did. Verification was done separately, in a real browser with simulated drag events rather than from the model's own account ^[7]. That second part carries more weight than it looks. June listed overconfident reporting as failure mode two and expected a harness to cover it alongside stateless regression; the dev.to account's own reckoning is that the first held and the second did not ^[19][16]. A tool loop makes the model's claims checkable, but the checking still happens outside the loop.

The two runs are 8 out of 8 in 2.1 minutes and 7.5 out of 8 in 44.1 ^[6][8]. Divide those: 44.1 over 2.1 is 21, so the 30b model finished about twenty-one times faster for half a point of score ^[20][10]. The 9b run is the one that proves the mechanism, because it passed the regression check as well ^[9]; the 30b run is the one that makes the loop usable while you sit at the desk. Add the follow-up fix and the whole exercise took about 2.6 minutes ^[22].

Now the transfer conditions. This is three files, one 10x10 grid, eight scored criteria ^[3]. For 2.1 minutes to mean anything on your codebase, the regression surface has to be roughly as small as one drag interaction, and the state the model must hold has to fit in one context. Two marks sit in the ledger against the top score. One run died with "Stream ended without finish_reason" and an identical retry worked ^[11]. On criterion eight, one green pixel was never confirmed because the preview tab had collapsed to a zero-width viewport and inline styles stopped applying, which was scored as a harness fault rather than a code fault ^[12]. I would rule the same way. I would also record that the headline score carries that ruling inside it, and that a rig which fails in ways impossible in the target browser is now a thing you have to score too.

The sentence that travelled was "The June verdict was a harness problem, not a model problem", quoted straight into the active constraints file and used to scope the work that followed ^[13]. What June actually wrote was narrower: local models cannot do iterative coding in a chat UI, a claim about an interface, in one row of a table of six use cases ^[15]. Five weeks before the re-test, the same document said in bold that tool-mediated verification per step and persistent conversation context are why Claude Code avoids the first two failure modes, and called that architecture rather than raw model capability ^[18]. The write-up's own summary concludes that the investment paid off even though the claim describing it was inaccurate, and nobody checked for two weeks, because a conclusion that is useful and turning out well does not feel like it needs auditing ^[14]. July did not overturn June so much as build what June had specified and measure what happened ^[23].