Thanks, Ahmet, that's a good catch. I went back to the per-item results (episodes.csv) and made the paired table for the full agent across the 50 instances: Paraphrase correct Paraphrase wrong Original correct 29 (both) 8 (original only) Original wrong 10 (paraphrase only) 3 (neither) So you were right that the matching totals hide churn: 18 of the 50 instances (36%) succeed in one phrasing and fail in the other. The direction is balanced (8 vs 10, exact McNemar p = 0.81), so neither wording is easier overall. The model just understands a different subset of the instances in each wording. The agent without the repair round shows the same pattern (27 / 8 / 12 / 3). The swapped-block errors are split evenly too: 9 in the original wording and 10 in the paraphrase. The original wording also has 3 missing-goal-fact errors and 1 unsatisfiable goal; the paraphrase has 1 parse failure. Since the 18 flips keep the same initial layout and gold goal and change only the wording, they are already the controlled test you describe, and they say the colour-name mapping is not yet consistent. Your suggested next step makes sense: run several wordings of each fixed instance and measure per-instance consistency before adding more repair logic. As the post shows, the repair round mostly turned invalid plans into valid plans for the wrong goal.