Expand limitations: no execution, templated reasoning, gold-conditioned, explore-only tail, unmeasured diversity

README.md

@@ -71,7 +71,16 @@ We cover 119,084 of SWE-ZERO's 122,908 PRs. 117,175 have exactly 100 rollouts an
 
 ### Caveats
 
-This is mid-training data, not SFT, same as SWE-ZERO. Nothing is checked against tests. Most rollouts don't end in a clean submit, and about 20% stop without an edit on purpose to match the real distribution. If you filter to clean submissions you throw away most of the dataset.
+This is mid-training data, not SFT, same as SWE-ZERO. There's a fair amount to be upfront about.
+
+- Nothing executes. Observations are read from the real repo, but we never run tests or builds, so no trajectory contains test output, stack traces, or runtime errors. This won't teach test-driven debugging.
+- The reasoning is templated. The THOUGHT text comes from a grammar, not a model. It's grounded in what the agent actually saw, but it reads flat and shows a plausible narration rather than real problem-solving.
+- The generator already knows the fix. Every trajectory is built backward from the gold patch, so the investigation is a path toward a known answer, not a genuine search. We add randomness so a PR's 100 rollouts differ, but they don't get stuck the way a real agent does.
+- Some PRs never reach an edit. When the gold patch can't be replayed as a safe edit, those rollouts explore and stop. Separately, about 20% stop without an edit on purpose to match the real distribution.
+- Nothing is verified. No patch is checked against tests, and we have not run the downstream training bake-off yet (see status above).
+- Intra-PR diversity is not measured yet. SWE-ZERO reports a MinHash similarity of 0.273 across a PR's rollouts. We have not computed ours.
+
+Most rollouts don't end in a clean submit. If you filter to clean submissions only, you throw away most of the dataset.
 
 ### Thanks