Model Comparison: Gemma vs Qwen vs GPT-5.5 vs Sonnet 4.6

All Languages Pooled

highest in row lowest in row Sorted by Δ (max − min) descending.

Node visit rates — sorted by Δ

state	Gemma (100)	Qwen (100)	GPT-5.5 (95)	Sonnet 4.6 (95)	GLM-5.2 (81)	Δ
start	1.0	1.0	1.0	1.0	1.0	0.00
improved	2.0	1.8	0.0	0.2	0.0	1.98
churn	0.1	0.3	—	0.1	—	0.22
stuck	11.2	4.3	0.2	0.3	0.1	11.07
breakthrough	4.8	2.0	1.3	1.2	1.3	3.66
stay	5.7	1.7	2.6	2.6	2.5	3.95
regressed	0.5	0.3	0.1	0.2	0.1	0.46
broke	5.5	2.3	0.3	0.1	0.3	5.43
end_submit	0.8	0.8	1.0	0.9	1.0	0.19
end_limit	0.2	0.2	—	0.1	0.0	0.17

Edge visit rates (count / n_problems) — sorted by Δ

transition	Gemma (100)	Qwen (100)	GPT-5.5 (95)	Sonnet 4.6 (95)	GLM-5.2 (81)	Δ (max−min)
stuck → stuck	9.1×	2.9×	0.1×	0.1×	0.0×	9.04×
broke → breakthrough	3.0×	0.5×	0.2×	0.0×	0.2×	3.02×
stay → stay	3.3×	0.8×	1.6×	1.7×	1.5×	2.43×
breakthrough → broke	2.0×	0.6×	0.0×	0.0×	—	1.95×
stay → broke	2.0×	0.5×	0.3×	0.0×	0.3×	1.95×
broke → stuck	1.6×	1.1×	0.1×	0.0×	0.1×	1.55×
breakthrough → stay	2.4×	0.9×	1.0×	0.9×	1.0×	1.52×
improved → broke	1.1×	0.8×	—	0.0×	0.0×	1.08×
broke → improved	0.9×	0.7×	0.0×	0.0×	0.0×	0.86×
stuck → improved	0.8×	0.6×	0.0×	0.0×	—	0.81×
stuck → breakthrough	0.7×	0.5×	0.1×	0.1×	0.1×	0.62×
stay → end_submit	0.3×	0.4×	0.7×	0.8×	0.7×	0.49×
improved → stuck	0.3×	0.1×	0.0×	0.0×	—	0.33×
stuck → end_submit	0.3×	0.1×	0.0×	0.0×	—	0.25×
start → breakthrough	0.7×	0.7×	1.0×	0.9×	1.0×	0.25×
breakthrough → end_submit	0.2×	0.3×	0.3×	0.1×	0.3×	0.25×
breakthrough → regressed	0.2×	0.2×	0.1×	0.1×	0.0×	0.23×
improved → churn	0.1×	0.2×	—	0.0×	—	0.21×
stuck → broke	0.2×	0.1×	0.0×	0.0×	0.0×	0.20×
regressed → broke	0.2×	0.2×	0.0×	0.0×	0.0×	0.19×
improved → breakthrough	0.2×	0.2×	0.0×	0.1×	0.0×	0.18×
improved → improved	0.1×	0.2×	—	0.0×	—	0.17×
start → stuck	0.2×	0.1×	0.0×	0.0×	0.0×	0.16×
start → improved	0.1×	0.2×	0.0×	0.1×	0.0×	0.15×
regressed → breakthrough	0.2×	0.1×	0.1×	0.1×	0.0×	0.13×
improved → regressed	0.1×	0.1×	—	0.0×	—	0.11×
churn → broke	0.1×	0.1×	—	—	—	0.07×
stuck → end_limit	0.1×	0.1×	—	0.0×	—	0.07×
regressed → improved	0.1×	0.0×	—	0.0×	—	0.05×
regressed → stuck	0.0×	0.0×	0.0×	0.1×	0.0×	0.04×
stay → regressed	0.1×	0.0×	0.0×	0.1×	0.0×	0.04×
stay → end_limit	0.1×	—	—	0.0×	0.0×	0.04×
improved → end_limit	0.0×	0.0×	—	0.0×	—	0.03×
regressed → churn	0.0×	—	—	0.0×	—	0.03×
improved → end_submit	0.0×	0.1×	—	—	—	0.02×
churn → improved	0.0×	0.0×	—	—	—	0.02×
churn → churn	0.0×	0.0×	—	0.0×	—	0.02×
churn → regressed	0.0×	0.0×	—	0.0×	—	0.02×
churn → breakthrough	—	0.0×	—	0.0×	—	0.01×
regressed → end_submit	—	—	0.0×	0.0×	—	0.01×
breakthrough → end_limit	0.0×	0.0×	—	0.0×	0.0×	0.01×
churn → end_limit	—	0.0×	—	0.0×	—	0.01×
regressed → regressed	—	0.0×	0.0×	0.0×	—	0.01×
broke → end_limit	0.0×	0.0×	—	—	—	0.00×
broke → end_submit	—	—	—	—	0.0×	0.00×
churn → stuck	—	0.0×	—	—	—	0.00×
regressed → end_limit	0.0×	—	—	—	—	0.00×
stuck → regressed	0.1×	—	—	—	—	0.00×

Transition probabilities P(tgt | src) — sorted by Δ

transition	Gemma (100)	Qwen (100)	GPT-5.5 (95)	Sonnet 4.6 (95)	GLM-5.2 (81)	Δ (max−min)
improved → breakthrough	10.0%	9.7%	66.7%	47.1%	66.7%	57.0pp
regressed → broke	36.5%	57.1%	10.0%	4.3%	20.0%	52.8pp
stuck → stuck	81.1%	68.6%	29.4%	48.0%	36.4%	51.7pp
broke → breakthrough	55.5%	20.5%	67.9%	42.9%	72.0%	51.5pp
improved → broke	54.2%	46.9%	—	5.9%	33.3%	48.3pp
stuck → breakthrough	6.2%	11.9%	47.1%	28.0%	54.5%	48.3pp
breakthrough → broke	40.7%	28.6%	1.6%	1.8%	—	39.1pp
regressed → breakthrough	32.7%	22.9%	60.0%	47.8%	60.0%	37.1pp
churn → breakthrough	—	3.7%	—	40.0%	—	36.3pp
stay → broke	34.7%	27.3%	9.6%	0.8%	11.1%	33.9pp
breakthrough → stay	49.8%	45.4%	74.0%	77.7%	75.7%	32.3pp
broke → improved	15.8%	32.3%	3.6%	28.6%	4.0%	28.7pp
broke → stuck	28.5%	46.7%	28.6%	28.6%	20.0%	26.7pp
improved → stuck	16.9%	7.4%	33.3%	11.8%	—	25.9pp
stay → end_submit	5.6%	22.7%	26.5%	31.2%	27.6%	25.6pp
start → breakthrough	73.0%	72.0%	96.8%	85.3%	96.3%	24.8pp
regressed → stuck	7.7%	2.9%	10.0%	21.7%	20.0%	18.8pp
breakthrough → end_submit	4.1%	16.8%	20.5%	7.1%	21.5%	17.4pp
stay → stay	57.5%	48.3%	62.2%	64.8%	59.3%	16.5pp
start → stuck	17.0%	12.0%	2.1%	4.2%	1.2%	15.8pp
start → improved	10.0%	16.0%	1.1%	10.5%	2.5%	14.9pp
improved → churn	4.0%	13.7%	—	17.6%	—	13.6pp
churn → end_limit	—	7.4%	—	20.0%	—	12.6pp
churn → regressed	20.0%	7.4%	—	20.0%	—	12.6pp
stuck → end_submit	2.3%	1.2%	11.8%	4.0%	—	10.6pp
breakthrough → regressed	5.2%	8.2%	3.9%	12.5%	1.9%	10.6pp
churn → churn	20.0%	11.1%	—	20.0%	—	8.9pp
stuck → improved	7.3%	13.8%	5.9%	8.0%	—	7.9pp
stuck → end_limit	0.8%	2.1%	—	8.0%	—	7.2pp
stuck → broke	1.9%	2.3%	5.9%	4.0%	9.1%	7.2pp
regressed → regressed	—	5.7%	10.0%	4.3%	—	5.7pp
churn → broke	46.7%	51.9%	—	—	—	5.2pp
improved → end_limit	1.0%	2.3%	—	5.9%	—	4.9pp
regressed → improved	13.5%	11.4%	—	8.7%	—	4.8pp
improved → improved	6.5%	10.3%	—	5.9%	—	4.4pp
regressed → churn	7.7%	—	—	4.3%	—	3.4pp
churn → improved	13.3%	14.8%	—	—	—	1.5pp
improved → end_submit	1.5%	2.9%	—	—	—	1.4pp
regressed → end_submit	—	—	10.0%	8.7%	—	1.3pp
stay → regressed	1.2%	1.7%	1.6%	2.4%	1.5%	1.2pp
improved → regressed	6.0%	6.9%	—	5.9%	—	1.0pp
breakthrough → end_limit	0.2%	1.0%	—	0.9%	0.9%	0.8pp
stay → end_limit	0.9%	—	—	0.8%	0.5%	0.4pp
broke → end_limit	0.2%	0.4%	—	—	—	0.2pp
broke → end_submit	—	—	—	—	4.0%	0.0pp
churn → stuck	—	3.7%	—	—	—	0.0pp
regressed → end_limit	1.9%	—	—	—	—	0.0pp
stuck → regressed	0.4%	—	—	—	—	0.0pp

Gap label distributions — sorted by edge visit rate, labels by avg%

transition	label	Gemma (100)	Qwen (100)	GPT-5.5 (95)	Sonnet 4.6 (95)	GLM-5.2 (81)	Δ (max−min)
stuck → stuck (2.45×/traj)	fix_bug	13.4%	45.5%	15.4%	57.1%	85.7%	72.3pp
	revert	27.8%	—	—	—	—	0.0pp
	debug_failure	17.2%	11.6%	7.7%	—	14.3%	9.5pp
	implement	15.4%	13.3%	15.4%	14.3%	—	2.1pp
	explore	10.6%	3.9%	—	14.3%	—	10.4pp
	infra_confusion	4.3%	14.6%	61.5%	—	—	57.2pp
	rewrite	4.1%	8.8%	—	—	—	4.7pp
	cosmetic_refactor	4.7%	0.9%	—	—	—	3.8pp
	execute_plan	1.3%	1.1%	—	—	—	0.2pp
	verify	0.9%	0.2%	—	—	—	0.7pp
	optimize_perf	0.4%	0.2%	—	14.3%	—	14.1pp
stay → stay (1.77×/traj)	cosmetic_refactor	33.5%	30.0%	—	—	—	3.5pp
	verify	16.1%	10.0%	—	—	—	6.1pp
	fix_bug	11.2%	30.0%	25.0%	70.0%	100.0%	88.8pp
	optimize_perf	11.8%	20.0%	—	10.0%	—	10.0pp
	implement	11.2%	—	—	—	—	0.0pp
	revert	6.5%	—	—	10.0%	—	3.5pp
	edge_case	5.5%	10.0%	—	—	—	4.5pp
	infra_confusion	1.6%	—	75.0%	10.0%	—	73.4pp
	debug_failure	1.6%	—	—	—	—	0.0pp
	explore	0.8%	—	—	—	—	0.0pp
	execute_plan	0.2%	—	—	—	—	0.0pp
breakthrough → stay (1.24×/traj)	cosmetic_refactor	25.1%	29.4%	—	—	—	4.3pp
	implement	21.1%	11.8%	—	14.3%	—	9.3pp
	fix_bug	10.3%	32.4%	80.0%	71.4%	100.0%	89.7pp
	optimize_perf	18.9%	2.9%	—	14.3%	—	16.0pp
	edge_case	10.3%	2.9%	—	—	—	7.4pp
	verify	7.4%	11.8%	—	—	—	4.4pp
	infra_confusion	2.9%	5.9%	—	—	—	3.0pp
	rewrite	1.7%	2.9%	20.0%	—	—	18.3pp
	debug_failure	1.1%	—	—	—	—	0.0pp
	execute_plan	0.6%	—	—	—	—	0.0pp
	revert	0.6%	—	—	—	—	0.0pp
start → breakthrough (0.85×/traj)	implement	100.0%	99.5%	100.0%	100.0%	100.0%	0.5pp
	explore	—	0.5%	—	—	—	0.0pp
broke → breakthrough (0.79×/traj)	fix_bug	50.0%	25.0%	—	—	—	25.0pp
	implement	26.7%	25.0%	—	—	—	1.7pp
	cosmetic_refactor	6.7%	25.0%	—	—	—	18.3pp
	rewrite	6.7%	25.0%	—	—	—	18.3pp
	execute_plan	3.3%	—	—	—	—	0.0pp
	infra_confusion	3.3%	—	—	—	—	0.0pp
	debug_failure	3.3%	—	—	—	—	0.0pp
stay → broke (0.60×/traj)	explore	32.0%	—	—	—	—	0.0pp
	debug_failure	20.0%	—	—	—	—	0.0pp
	implement	12.0%	—	—	—	—	0.0pp
	fix_bug	12.0%	—	—	—	—	0.0pp
	optimize_perf	8.0%	—	—	—	—	0.0pp
	rewrite	8.0%	—	—	—	—	0.0pp
	infra_confusion	8.0%	—	—	—	—	0.0pp
stay → end_submit (0.58×/traj)	verify	99.1%	100.0%	100.0%	57.1%	50.0%	50.0pp
	infra_confusion	0.9%	—	—	42.9%	50.0%	49.1pp
broke → stuck (0.56×/traj)	explore	21.3%	15.8%	—	—	—	5.5pp
	debug_failure	23.0%	10.5%	—	—	—	12.5pp
	implement	9.8%	26.3%	—	—	—	16.5pp
	fix_bug	13.1%	10.5%	—	—	—	2.6pp
	rewrite	8.2%	26.3%	—	—	—	18.1pp
	revert	14.8%	—	—	—	—	0.0pp
	infra_confusion	1.6%	10.5%	—	—	—	8.9pp
	execute_plan	3.3%	—	—	—	—	0.0pp
	cosmetic_refactor	3.3%	—	—	—	—	0.0pp
	optimize_perf	1.6%	—	—	—	—	0.0pp
breakthrough → broke (0.51×/traj)	debug_failure	24.1%	33.3%	—	—	—	9.2pp
	fix_bug	20.7%	—	—	—	—	0.0pp
	infra_confusion	17.2%	16.7%	—	—	—	0.5pp
	rewrite	10.3%	16.7%	—	—	—	6.4pp
	explore	6.9%	16.7%	—	—	—	9.8pp
	implement	6.9%	16.7%	—	—	—	9.8pp
	optimize_perf	6.9%	—	—	—	—	0.0pp
	edge_case	3.4%	—	—	—	—	0.0pp
	revert	3.4%	—	—	—	—	0.0pp
improved → broke (0.39×/traj)	explore	37.0%	21.7%	—	—	—	15.3pp
	fix_bug	18.5%	26.1%	—	50.0%	—	31.5pp
	rewrite	14.8%	26.1%	—	—	—	11.3pp
	implement	11.1%	13.0%	—	—	—	1.9pp
	debug_failure	3.7%	13.0%	—	50.0%	—	46.3pp
	cosmetic_refactor	7.4%	—	—	—	—	0.0pp
	execute_plan	3.7%	—	—	—	—	0.0pp
	infra_confusion	3.7%	—	—	—	—	0.0pp
broke → improved (0.33×/traj)	fix_bug	14.3%	42.9%	—	—	—	28.6pp
	rewrite	28.6%	19.0%	—	—	—	9.6pp
	implement	9.5%	19.0%	—	—	—	9.5pp
	execute_plan	4.8%	9.5%	—	—	—	4.7pp
	revert	14.3%	—	—	—	—	0.0pp
	explore	9.5%	4.8%	—	—	—	4.7pp
	cosmetic_refactor	9.5%	—	—	100.0%	—	90.5pp
	debug_failure	9.5%	4.8%	—	—	—	4.7pp
stuck → improved (0.29×/traj)	fix_bug	25.3%	63.6%	50.0%	—	—	38.3pp
	implement	24.0%	10.9%	—	—	—	13.1pp
	rewrite	17.3%	14.5%	—	50.0%	—	35.5pp
	infra_confusion	5.3%	7.3%	50.0%	—	—	44.7pp
	execute_plan	8.0%	—	—	50.0%	—	42.0pp
	cosmetic_refactor	8.0%	—	—	—	—	0.0pp
	revert	6.7%	—	—	—	—	0.0pp
	explore	4.0%	—	—	—	—	0.0pp
	debug_failure	1.3%	1.8%	—	—	—	0.5pp
	optimize_perf	—	1.8%	—	—	—	0.0pp
stuck → breakthrough (0.29×/traj)	fix_bug	51.8%	76.4%	33.3%	81.8%	69.2%	48.5pp
	infra_confusion	5.9%	12.5%	48.1%	9.1%	26.9%	42.2pp
	implement	16.5%	1.4%	3.7%	—	—	15.1pp
	cosmetic_refactor	7.1%	2.8%	3.7%	—	3.8%	4.3pp
	execute_plan	8.2%	—	7.4%	—	—	0.8pp
	rewrite	3.5%	6.9%	—	9.1%	—	5.6pp
	optimize_perf	3.5%	—	3.7%	—	—	0.2pp
	debug_failure	3.5%	—	—	—	—	0.0pp
breakthrough → end_submit (0.23×/traj)	verify	100.0%	99.6%	100.0%	100.0%	100.0%	0.4pp
	fix_bug	—	0.4%	—	—	—	0.0pp
breakthrough → regressed (0.13×/traj)	fix_bug	16.7%	55.6%	—	—	—	38.9pp
	implement	50.0%	11.1%	—	—	—	38.9pp
	optimize_perf	—	11.1%	—	—	—	0.0pp
	cosmetic_refactor	16.7%	—	—	—	—	0.0pp
	rewrite	—	11.1%	—	—	—	0.0pp
	infra_confusion	16.7%	—	—	—	—	0.0pp
	debug_failure	—	11.1%	—	—	—	0.0pp
improved → stuck (0.10×/traj)	implement	17.9%	29.0%	—	—	—	11.1pp
	cosmetic_refactor	16.4%	12.9%	—	—	—	3.5pp
	fix_bug	16.4%	3.2%	—	75.0%	—	71.8pp
	debug_failure	9.0%	22.6%	—	—	—	13.6pp
	rewrite	6.0%	16.1%	—	—	—	10.1pp
	revert	10.4%	—	—	—	—	0.0pp
	explore	6.0%	9.7%	—	—	—	3.7pp
	optimize_perf	3.0%	3.2%	100.0%	25.0%	—	97.0pp
	verify	7.5%	—	—	—	—	0.0pp
	infra_confusion	6.0%	3.2%	—	—	—	2.8pp
	execute_plan	1.5%	—	—	—	—	0.0pp
improved → breakthrough (0.10×/traj)	fix_bug	44.4%	88.0%	75.0%	42.9%	75.0%	45.1pp
	rewrite	11.1%	8.0%	—	42.9%	—	34.9pp
	implement	11.1%	—	—	4.8%	25.0%	20.2pp
	debug_failure	11.1%	—	12.5%	4.8%	—	7.7pp
	optimize_perf	11.1%	—	—	4.8%	—	6.3pp
	infra_confusion	11.1%	—	12.5%	—	—	1.4pp
	execute_plan	—	4.0%	—	—	—	0.0pp
regressed → breakthrough (0.09×/traj)	fix_bug	66.7%	66.7%	—	—	—	0.0pp
	optimize_perf	—	16.7%	—	—	—	0.0pp
	implement	—	16.7%	—	—	—	0.0pp
	debug_failure	33.3%	—	—	—	—	0.0pp
regressed → broke (0.08×/traj)	rewrite	50.0%	33.3%	—	—	—	16.7pp
	optimize_perf	—	33.3%	—	—	—	0.0pp
	implement	25.0%	—	—	—	—	0.0pp
	fix_bug	25.0%	—	—	—	—	0.0pp
	infra_confusion	—	33.3%	—	—	—	0.0pp
start → improved (0.08×/traj)	implement	88.9%	95.3%	100.0%	100.0%	100.0%	11.1pp
	explore	7.4%	4.7%	—	—	—	2.7pp
	infra_confusion	3.7%	—	—	—	—	0.0pp
start → stuck (0.07×/traj)	implement	60.9%	81.6%	96.6%	75.0%	100.0%	39.1pp
	explore	39.1%	18.4%	3.4%	25.0%	—	35.7pp
improved → churn (0.07×/traj)	rewrite	66.7%	50.0%	—	—	—	16.7pp
	fix_bug	33.3%	40.0%	—	—	—	6.7pp
	implement	—	10.0%	—	—	—	0.0pp
stuck → broke (0.07×/traj)	explore	26.9%	12.5%	—	—	—	14.4pp
	debug_failure	11.5%	50.0%	—	—	—	38.5pp
	fix_bug	23.1%	—	—	—	—	0.0pp
	rewrite	15.4%	25.0%	—	—	—	9.6pp
	implement	15.4%	12.5%	—	—	—	2.9pp
	revert	3.8%	—	—	—	—	0.0pp
	infra_confusion	3.8%	—	—	—	—	0.0pp
stuck → end_submit (0.07×/traj)	verify	100.0%	—	—	—	—	0.0pp
	infra_confusion	—	—	100.0%	—	—	0.0pp
improved → improved (0.06×/traj)	fix_bug	40.0%	71.4%	—	66.7%	—	31.4pp
	implement	40.0%	—	—	—	—	0.0pp
	rewrite	10.0%	28.6%	—	—	—	18.6pp
	optimize_perf	10.0%	—	—	33.3%	—	23.3pp
improved → regressed (0.05×/traj)	fix_bug	40.0%	36.4%	—	—	—	3.6pp
	rewrite	20.0%	27.3%	—	50.0%	—	30.0pp
	implement	40.0%	—	—	25.0%	—	15.0pp
	optimize_perf	—	27.3%	—	—	—	0.0pp
	revert	—	9.1%	—	—	—	0.0pp
	debug_failure	—	—	—	25.0%	—	0.0pp
stay → regressed (0.05×/traj)	optimize_perf	33.3%	100.0%	—	—	—	66.7pp
	rewrite	33.3%	—	—	—	—	0.0pp
	cosmetic_refactor	33.3%	—	—	—	—	0.0pp
churn → broke (0.04×/traj)	rewrite	100.0%	50.0%	—	—	—	50.0pp
	explore	—	16.7%	—	—	—	0.0pp
	implement	—	16.7%	—	—	—	0.0pp
	debug_failure	—	16.7%	—	—	—	0.0pp
stuck → end_limit (0.04×/traj)	verify	81.0%	43.2%	—	33.3%	—	47.7pp
	explore	3.2%	29.7%	—	66.7%	—	63.5pp
	debug_failure	12.7%	16.2%	—	—	—	3.5pp
	infra_confusion	—	8.1%	—	—	—	0.0pp
	implement	1.6%	—	—	—	—	0.0pp
	fix_bug	—	2.7%	—	—	—	0.0pp
	rewrite	1.6%	—	—	—	—	0.0pp
regressed → improved (0.03×/traj)	fix_bug	33.3%	50.0%	—	50.0%	—	16.7pp
	rewrite	16.7%	50.0%	—	—	—	33.3pp
	implement	33.3%	—	—	—	—	0.0pp
	optimize_perf	—	—	—	50.0%	—	0.0pp
	debug_failure	16.7%	—	—	—	—	0.0pp
regressed → stuck (0.03×/traj)	debug_failure	22.2%	50.0%	—	—	—	27.8pp
	optimize_perf	22.2%	—	—	—	—	0.0pp
	cosmetic_refactor	22.2%	—	—	—	—	0.0pp
	fix_bug	11.1%	50.0%	—	—	—	38.9pp
	implement	11.1%	—	—	—	—	0.0pp
	rewrite	11.1%	—	—	—	—	0.0pp
stay → end_limit (0.02×/traj)	verify	93.3%	—	—	—	—	0.0pp
	infra_confusion	3.3%	—	—	—	100.0%	96.7pp
	debug_failure	3.3%	—	—	—	—	0.0pp
improved → end_submit (0.02×/traj)	verify	100.0%	100.0%	—	—	—	0.0pp
improved → end_limit (0.01×/traj)	verify	61.5%	38.5%	—	9.1%	—	52.4pp
	debug_failure	30.8%	34.6%	—	27.3%	—	7.3pp
	explore	7.7%	23.1%	—	63.6%	—	55.9pp
	fix_bug	—	3.8%	—	—	—	0.0pp
churn → churn (0.01×/traj)	rewrite	50.0%	43.8%	—	—	—	6.2pp
	fix_bug	—	37.5%	—	—	—	0.0pp
	implement	25.0%	12.5%	—	—	—	12.5pp
	debug_failure	25.0%	6.2%	—	—	—	18.8pp
churn → regressed (0.01×/traj)	implement	50.0%	—	—	—	—	0.0pp
	fix_bug	50.0%	—	—	—	—	0.0pp
churn → improved (0.01×/traj)	fix_bug	100.0%	66.7%	—	—	—	33.3pp
	implement	—	33.3%	—	—	—	0.0pp
breakthrough → end_limit (0.01×/traj)	verify	100.0%	100.0%	—	—	75.0%	25.0pp
	fix_bug	—	—	—	—	25.0%	0.0pp
regressed → churn (0.01×/traj)	implement	25.0%	—	—	—	—	0.0pp
	edge_case	25.0%	—	—	—	—	0.0pp
	fix_bug	25.0%	—	—	—	—	0.0pp
	rewrite	25.0%	—	—	—	—	0.0pp
stuck → regressed (0.01×/traj)	implement	57.1%	—	—	—	—	0.0pp
	rewrite	28.6%	—	—	—	—	0.0pp
	fix_bug	14.3%	—	—	—	—	0.0pp
regressed → regressed (0.01×/traj)	fix_bug	—	100.0%	—	—	—	0.0pp
churn → breakthrough (0.01×/traj)	fix_bug	—	100.0%	—	—	—	0.0pp
churn → end_limit (0.01×/traj)	explore	—	33.3%	—	—	—	0.0pp
	debug_failure	—	33.3%	—	—	—	0.0pp
	verify	—	33.3%	—	—	—	0.0pp
broke → end_limit (0.00×/traj)	verify	100.0%	—	—	—	—	0.0pp
	infra_confusion	—	50.0%	—	—	—	0.0pp
	debug_failure	—	50.0%	—	—	—	0.0pp
churn → stuck (0.00×/traj)	implement	—	40.0%	—	—	—	0.0pp
	fix_bug	—	40.0%	—	—	—	0.0pp
	rewrite	—	20.0%	—	—	—	0.0pp
regressed → end_limit (0.00×/traj)	optimize_perf	50.0%	—	—	—	—	0.0pp
	verify	50.0%	—	—	—	—	0.0pp

Python

highest in row lowest in row Sorted by Δ (max − min) descending.

Node visit rates — sorted by Δ

state	Gemma (100)	Qwen (97)	GPT-5.5 (93)	Sonnet 4.6 (91)	GLM-5.2 (75)	Δ
start	1.0	1.0	1.0	1.0	1.0	0.00
improved	0.4	0.3	0.0	0.1	0.0	0.38
churn	0.0	0.2	—	—	—	0.18
stuck	1.0	0.4	0.0	0.1	—	0.96
breakthrough	1.0	0.9	1.0	0.9	1.0	0.10
stay	2.8	0.1	0.0	0.1	0.0	2.76
regressed	0.1	0.1	—	0.0	—	0.14
broke	0.2	0.1	—	0.0	—	0.18
end_submit	0.8	0.8	1.0	0.9	1.0	0.19
end_limit	0.2	0.2	—	0.1	0.0	0.18

Edge visit rates (count / n_problems) — sorted by Δ

transition	Gemma (100)	Qwen (97)	GPT-5.5 (93)	Sonnet 4.6 (91)	GLM-5.2 (75)	Δ (max−min)
stay → stay	2.2×	0.0×	0.0×	0.0×	0.0×	2.24×
breakthrough → end_submit	0.4×	0.8×	1.0×	0.9×	1.0×	0.58×
breakthrough → stay	0.5×	0.1×	0.0×	0.0×	0.0×	0.53×
stuck → stuck	0.5×	0.2×	—	0.0×	—	0.48×
stay → end_submit	0.3×	0.1×	0.0×	0.0×	0.0×	0.32×
start → breakthrough	0.7×	0.7×	1.0×	0.9×	1.0×	0.26×
stuck → improved	0.2×	0.1×	0.0×	0.0×	—	0.16×
start → stuck	0.2×	0.1×	0.0×	0.0×	—	0.15×
improved → stuck	0.1×	0.1×	0.0×	0.0×	—	0.14×
start → improved	0.1×	0.1×	0.0×	0.1×	0.0×	0.13×
stuck → breakthrough	0.1×	0.1×	0.0×	0.0×	—	0.11×
churn → churn	0.0×	0.1×	—	—	—	0.09×
stay → end_limit	0.1×	0.0×	—	—	—	0.09×
broke → stuck	0.1×	0.0×	—	—	—	0.07×
improved → regressed	0.1×	0.1×	—	0.0×	—	0.06×
regressed → stuck	0.1×	0.0×	—	—	—	0.05×
improved → breakthrough	0.1×	0.1×	0.0×	0.1×	—	0.04×
stuck → broke	0.1×	0.0×	—	—	—	0.04×
improved → churn	0.0×	0.0×	—	—	—	0.03×
improved → improved	0.0×	0.0×	—	0.0×	—	0.03×
regressed → improved	0.0×	0.0×	—	0.0×	—	0.03×
improved → end_limit	0.0×	0.1×	—	0.0×	—	0.03×
regressed → breakthrough	0.0×	0.0×	—	—	—	0.02×
stuck → end_limit	0.0×	0.1×	—	0.0×	—	0.02×
regressed → churn	0.0×	0.0×	—	—	—	0.01×
stay → regressed	0.0×	0.0×	—	—	—	0.01×
stuck → regressed	0.0×	0.0×	—	—	—	0.01×
improved → broke	0.0×	0.0×	—	0.0×	—	0.01×
broke → improved	0.0×	0.0×	—	—	—	0.01×
improved → end_submit	0.0×	—	—	—	0.0×	0.00×
breakthrough → regressed	0.0×	0.0×	—	—	—	0.00×
breakthrough → broke	0.0×	0.0×	—	—	—	0.00×
churn → breakthrough	0.0×	0.0×	—	—	—	0.00×
breakthrough → end_limit	—	—	—	—	0.0×	0.00×
broke → breakthrough	0.1×	—	—	—	—	0.00×
broke → end_limit	—	—	—	0.0×	—	0.00×
churn → broke	0.0×	—	—	—	—	0.00×
churn → end_limit	—	0.0×	—	—	—	0.00×
churn → improved	—	0.0×	—	—	—	0.00×
churn → stuck	—	0.1×	—	—	—	0.00×
regressed → broke	—	0.0×	—	—	—	0.00×
regressed → end_limit	0.0×	—	—	—	—	0.00×
regressed → regressed	0.0×	—	—	—	—	0.00×
stay → broke	0.1×	—	—	—	—	0.00×
stuck → churn	—	0.0×	—	—	—	0.00×
stuck → end_submit	0.1×	—	—	—	—	0.00×

Transition probabilities P(tgt | src) — sorted by Δ

transition	Gemma (100)	Qwen (97)	GPT-5.5 (93)	Sonnet 4.6 (91)	GLM-5.2 (75)	Δ (max−min)
improved → end_submit	2.6%	—	—	—	100.0%	97.4pp
regressed → improved	26.7%	27.3%	—	100.0%	—	73.3pp
breakthrough → end_submit	40.8%	85.1%	97.8%	96.4%	97.3%	57.0pp
stuck → breakthrough	12.1%	14.3%	66.7%	11.1%	—	55.6pp
stay → end_submit	11.8%	53.8%	66.7%	42.9%	50.0%	54.9pp
breakthrough → stay	55.1%	10.3%	2.2%	3.6%	1.4%	53.7pp
stay → stay	80.6%	30.8%	33.3%	57.1%	50.0%	49.8pp
improved → stuck	38.5%	22.6%	50.0%	8.3%	—	41.7pp
broke → improved	21.1%	60.0%	—	—	—	38.9pp
improved → breakthrough	12.8%	16.1%	50.0%	41.7%	—	37.2pp
regressed → stuck	40.0%	9.1%	—	—	—	30.9pp
stuck → end_limit	4.0%	11.9%	—	33.3%	—	29.3pp
start → breakthrough	73.0%	74.2%	96.8%	85.7%	98.7%	25.7pp
churn → churn	25.0%	47.6%	—	—	—	22.6pp
stuck → improved	17.2%	21.4%	33.3%	11.1%	—	22.2pp
regressed → breakthrough	6.7%	27.3%	—	—	—	20.6pp
churn → breakthrough	25.0%	4.8%	—	—	—	20.2pp
start → stuck	17.0%	11.3%	2.2%	4.4%	—	14.8pp
improved → end_limit	10.3%	19.4%	—	25.0%	—	14.7pp
stuck → stuck	52.5%	38.1%	—	44.4%	—	14.4pp
regressed → churn	13.3%	27.3%	—	—	—	14.0pp
start → improved	10.0%	14.4%	1.1%	9.9%	1.3%	13.3pp
improved → regressed	17.9%	19.4%	—	8.3%	—	11.1pp
improved → churn	2.6%	12.9%	—	—	—	10.3pp
broke → stuck	47.4%	40.0%	—	—	—	7.4pp
improved → improved	10.3%	3.2%	—	8.3%	—	7.1pp
stay → regressed	0.7%	7.7%	—	—	—	7.0pp
stay → end_limit	3.6%	7.7%	—	—	—	4.1pp
improved → broke	5.1%	6.5%	—	8.3%	—	3.2pp
stuck → broke	5.1%	2.4%	—	—	—	2.7pp
stuck → regressed	2.0%	2.4%	—	—	—	0.4pp
breakthrough → regressed	3.1%	3.4%	—	—	—	0.3pp
breakthrough → broke	1.0%	1.1%	—	—	—	0.1pp
breakthrough → end_limit	—	—	—	—	1.4%	0.0pp
broke → breakthrough	31.6%	—	—	—	—	0.0pp
broke → end_limit	—	—	—	100.0%	—	0.0pp
churn → broke	50.0%	—	—	—	—	0.0pp
churn → end_limit	—	19.0%	—	—	—	0.0pp
churn → improved	—	4.8%	—	—	—	0.0pp
churn → stuck	—	23.8%	—	—	—	0.0pp
regressed → broke	—	9.1%	—	—	—	0.0pp
regressed → end_limit	6.7%	—	—	—	—	0.0pp
regressed → regressed	6.7%	—	—	—	—	0.0pp
stay → broke	3.2%	—	—	—	—	0.0pp
stuck → churn	—	9.5%	—	—	—	0.0pp
stuck → end_submit	7.1%	—	—	—	—	0.0pp

Gap label distributions — sorted by edge visit rate, labels by avg%

transition	label	Gemma (100)	Qwen (97)	GPT-5.5 (93)	Sonnet 4.6 (91)	GLM-5.2 (75)	Δ (max−min)
start → breakthrough (0.86×/traj)	implement	100.0%	100.0%	100.0%	100.0%	100.0%	0.0pp
breakthrough → end_submit (0.80×/traj)	verify	100.0%	100.0%	100.0%	100.0%	100.0%	0.0pp
stay → stay (0.47×/traj)	cosmetic_refactor	35.1%	—	—	—	—	0.0pp
	optimize_perf	17.3%	50.0%	—	25.0%	—	32.7pp
	implement	15.6%	—	—	—	—	0.0pp
	fix_bug	8.9%	50.0%	100.0%	25.0%	100.0%	91.1pp
	verify	8.9%	—	—	—	—	0.0pp
	revert	6.2%	—	—	25.0%	—	18.8pp
	edge_case	2.7%	—	—	—	—	0.0pp
	infra_confusion	1.3%	—	—	25.0%	—	23.7pp
	explore	1.8%	—	—	—	—	0.0pp
	debug_failure	1.8%	—	—	—	—	0.0pp
	execute_plan	0.4%	—	—	—	—	0.0pp
stuck → stuck (0.15×/traj)	explore	25.0%	18.8%	—	25.0%	—	6.2pp
	fix_bug	7.7%	37.5%	—	25.0%	—	29.8pp
	debug_failure	9.6%	25.0%	—	—	—	15.4pp
	revert	15.4%	—	—	—	—	0.0pp
	implement	11.5%	6.2%	—	25.0%	—	18.8pp
	rewrite	13.5%	6.2%	—	—	—	7.3pp
	cosmetic_refactor	9.6%	—	—	—	—	0.0pp
	optimize_perf	1.9%	6.2%	—	25.0%	—	23.1pp
	verify	5.8%	—	—	—	—	0.0pp
breakthrough → stay (0.14×/traj)	optimize_perf	31.5%	11.1%	—	33.3%	—	22.2pp
	implement	22.2%	22.2%	—	33.3%	—	11.1pp
	cosmetic_refactor	24.1%	11.1%	—	—	—	13.0pp
	fix_bug	11.1%	22.2%	50.0%	33.3%	100.0%	88.9pp
	rewrite	3.7%	11.1%	50.0%	—	—	46.3pp
	verify	1.9%	22.2%	—	—	—	20.3pp
	edge_case	3.7%	—	—	—	—	0.0pp
	execute_plan	1.9%	—	—	—	—	0.0pp
stay → end_submit (0.09×/traj)	verify	96.9%	100.0%	100.0%	66.7%	—	33.3pp
	infra_confusion	3.1%	—	—	33.3%	100.0%	96.9pp
start → improved (0.07×/traj)	implement	90.0%	100.0%	100.0%	100.0%	100.0%	10.0pp
	explore	10.0%	—	—	—	—	0.0pp
start → stuck (0.07×/traj)	explore	76.5%	18.2%	—	50.0%	—	58.3pp
	implement	23.5%	81.8%	100.0%	50.0%	—	76.5pp
stuck → improved (0.06×/traj)	fix_bug	17.6%	66.7%	100.0%	—	—	82.4pp
	execute_plan	29.4%	—	—	100.0%	—	70.6pp
	implement	29.4%	11.1%	—	—	—	18.3pp
	rewrite	17.6%	22.2%	—	—	—	4.6pp
	explore	5.9%	—	—	—	—	0.0pp
improved → stuck (0.05×/traj)	cosmetic_refactor	33.3%	28.6%	—	—	—	4.7pp
	implement	26.7%	14.3%	—	—	—	12.4pp
	optimize_perf	6.7%	—	100.0%	100.0%	—	93.3pp
	explore	6.7%	14.3%	—	—	—	7.6pp
	fix_bug	13.3%	—	—	—	—	0.0pp
	rewrite	—	28.6%	—	—	—	0.0pp
	debug_failure	6.7%	14.3%	—	—	—	7.6pp
	revert	6.7%	—	—	—	—	0.0pp
stuck → breakthrough (0.04×/traj)	fix_bug	33.3%	66.7%	—	—	—	33.4pp
	execute_plan	25.0%	—	50.0%	—	—	25.0pp
	optimize_perf	25.0%	—	50.0%	—	—	25.0pp
	implement	16.7%	—	—	—	—	0.0pp
	rewrite	—	33.3%	—	—	—	0.0pp
	infra_confusion	—	—	—	100.0%	—	0.0pp
improved → breakthrough (0.03×/traj)	fix_bug	40.0%	100.0%	—	40.0%	—	60.0pp
	rewrite	20.0%	—	—	40.0%	—	20.0pp
	optimize_perf	20.0%	—	—	20.0%	—	0.0pp
	implement	20.0%	—	—	—	—	0.0pp
	infra_confusion	—	—	100.0%	—	—	0.0pp
improved → regressed (0.03×/traj)	fix_bug	57.1%	16.7%	—	—	—	40.4pp
	rewrite	28.6%	33.3%	—	—	—	4.7pp
	optimize_perf	—	50.0%	—	—	—	0.0pp
	implement	14.3%	—	—	100.0%	—	85.7pp
improved → end_limit (0.03×/traj)	debug_failure	25.0%	50.0%	—	33.3%	—	25.0pp
	explore	—	33.3%	—	66.7%	—	33.4pp
	verify	75.0%	16.7%	—	—	—	58.3pp
stuck → end_limit (0.02×/traj)	verify	100.0%	60.0%	—	33.3%	—	66.7pp
	explore	—	20.0%	—	66.7%	—	46.7pp
	debug_failure	—	20.0%	—	—	—	0.0pp
churn → churn (0.02×/traj)	rewrite	—	60.0%	—	—	—	0.0pp
	fix_bug	—	20.0%	—	—	—	0.0pp
	debug_failure	100.0%	10.0%	—	—	—	90.0pp
	implement	—	10.0%	—	—	—	0.0pp
broke → stuck (0.02×/traj)	explore	55.6%	—	—	—	—	0.0pp
	debug_failure	22.2%	100.0%	—	—	—	77.8pp
	optimize_perf	11.1%	—	—	—	—	0.0pp
	execute_plan	11.1%	—	—	—	—	0.0pp
stay → end_limit (0.02×/traj)	verify	90.9%	—	—	—	—	0.0pp
	debug_failure	9.1%	100.0%	—	—	—	90.9pp
stay → broke (0.02×/traj)	explore	77.8%	—	—	—	—	0.0pp
	rewrite	11.1%	—	—	—	—	0.0pp
	infra_confusion	11.1%	—	—	—	—	0.0pp
regressed → improved (0.02×/traj)	fix_bug	25.0%	66.7%	—	—	—	41.7pp
	rewrite	25.0%	33.3%	—	—	—	8.3pp
	implement	50.0%	—	—	—	—	0.0pp
	optimize_perf	—	—	—	100.0%	—	0.0pp
broke → improved (0.01×/traj)	fix_bug	25.0%	66.7%	—	—	—	41.7pp
	rewrite	50.0%	—	—	—	—	0.0pp
	implement	25.0%	—	—	—	—	0.0pp
	execute_plan	—	33.3%	—	—	—	0.0pp
regressed → stuck (0.01×/traj)	fix_bug	16.7%	100.0%	—	—	—	83.3pp
	optimize_perf	16.7%	—	—	—	—	0.0pp
	implement	16.7%	—	—	—	—	0.0pp
	cosmetic_refactor	16.7%	—	—	—	—	0.0pp
	rewrite	16.7%	—	—	—	—	0.0pp
	debug_failure	16.7%	—	—	—	—	0.0pp
stuck → end_submit (0.01×/traj)	verify	100.0%	—	—	—	—	0.0pp
improved → improved (0.01×/traj)	fix_bug	75.0%	100.0%	—	—	—	25.0pp
	optimize_perf	25.0%	—	—	100.0%	—	75.0pp
breakthrough → regressed (0.01×/traj)	implement	100.0%	—	—	—	—	0.0pp
	optimize_perf	—	33.3%	—	—	—	0.0pp
	fix_bug	—	33.3%	—	—	—	0.0pp
	rewrite	—	33.3%	—	—	—	0.0pp
stuck → broke (0.01×/traj)	explore	60.0%	—	—	—	—	0.0pp
	rewrite	20.0%	—	—	—	—	0.0pp
	implement	—	100.0%	—	—	—	0.0pp
	debug_failure	20.0%	—	—	—	—	0.0pp
broke → breakthrough (0.01×/traj)	rewrite	33.3%	—	—	—	—	0.0pp
	execute_plan	16.7%	—	—	—	—	0.0pp
	implement	16.7%	—	—	—	—	0.0pp
	fix_bug	16.7%	—	—	—	—	0.0pp
	cosmetic_refactor	16.7%	—	—	—	—	0.0pp
improved → broke (0.01×/traj)	explore	100.0%	50.0%	—	—	—	50.0pp
	implement	—	50.0%	—	—	—	0.0pp
	debug_failure	—	—	—	100.0%	—	0.0pp
churn → stuck (0.01×/traj)	implement	—	40.0%	—	—	—	0.0pp
	fix_bug	—	40.0%	—	—	—	0.0pp
	rewrite	—	20.0%	—	—	—	0.0pp
improved → churn (0.01×/traj)	rewrite	—	100.0%	—	—	—	0.0pp
	fix_bug	100.0%	—	—	—	—	0.0pp
regressed → churn (0.01×/traj)	fix_bug	50.0%	33.3%	—	—	—	16.7pp
	optimize_perf	—	33.3%	—	—	—	0.0pp
	edge_case	50.0%	—	—	—	—	0.0pp
	rewrite	—	33.3%	—	—	—	0.0pp
churn → end_limit (0.01×/traj)	explore	—	50.0%	—	—	—	0.0pp
	verify	—	50.0%	—	—	—	0.0pp
stuck → churn (0.01×/traj)	fix_bug	—	50.0%	—	—	—	0.0pp
	rewrite	—	25.0%	—	—	—	0.0pp
	implement	—	25.0%	—	—	—	0.0pp
regressed → breakthrough (0.01×/traj)	fix_bug	100.0%	33.3%	—	—	—	66.7pp
	optimize_perf	—	33.3%	—	—	—	0.0pp
	implement	—	33.3%	—	—	—	0.0pp
stuck → regressed (0.01×/traj)	implement	100.0%	—	—	—	—	0.0pp
	fix_bug	—	100.0%	—	—	—	0.0pp
stay → regressed (0.01×/traj)	optimize_perf	50.0%	100.0%	—	—	—	50.0pp
	cosmetic_refactor	50.0%	—	—	—	—	0.0pp
improved → end_submit (0.00×/traj)	verify	100.0%	—	—	—	100.0%	0.0pp
churn → breakthrough (0.00×/traj)	revert	100.0%	—	—	—	—	0.0pp
	fix_bug	—	100.0%	—	—	—	0.0pp
breakthrough → broke (0.00×/traj)	explore	100.0%	—	—	—	—	0.0pp
	implement	—	100.0%	—	—	—	0.0pp
churn → broke (0.00×/traj)	rewrite	100.0%	—	—	—	—	0.0pp
breakthrough → end_limit (0.00×/traj)	verify	—	—	—	—	100.0%	0.0pp
broke → end_limit (0.00×/traj)	explore	—	—	—	100.0%	—	0.0pp
churn → improved (0.00×/traj)	fix_bug	—	100.0%	—	—	—	0.0pp
regressed → broke (0.00×/traj)	optimize_perf	—	100.0%	—	—	—	0.0pp
regressed → regressed (0.00×/traj)	rewrite	100.0%	—	—	—	—	0.0pp
regressed → end_limit (0.00×/traj)	verify	100.0%	—	—	—	—	0.0pp

Java

highest in row lowest in row Sorted by Δ (max − min) descending.

Node visit rates — sorted by Δ

state	Gemma (100)	Qwen (96)	GPT-5.5 (94)	Sonnet 4.6 (90)	GLM-5.2 (78)	Δ
start	1.0	1.0	1.0	1.0	1.0	0.00
improved	0.3	0.5	0.0	0.1	0.0	0.42
churn	0.0	0.1	—	—	—	0.11
stuck	0.9	0.5	—	0.0	—	0.90
breakthrough	0.9	0.8	1.0	1.0	1.0	0.16
stay	0.8	0.1	0.0	0.1	0.0	0.78
regressed	0.1	0.1	—	0.0	—	0.08
broke	0.1	0.1	—	—	—	0.02
end_submit	0.8	0.8	1.0	1.0	1.0	0.22
end_limit	0.2	0.2	—	0.0	0.0	0.21

Edge visit rates (count / n_problems) — sorted by Δ

transition	Gemma (100)	Qwen (96)	GPT-5.5 (94)	Sonnet 4.6 (90)	GLM-5.2 (78)	Δ (max−min)
breakthrough → end_submit	0.5×	0.7×	1.0×	0.9×	1.0×	0.53×
stuck → stuck	0.5×	0.2×	—	0.0×	—	0.51×
stay → stay	0.4×	0.0×	0.0×	0.0×	—	0.42×
breakthrough → stay	0.4×	0.0×	0.0×	0.0×	0.0×	0.35×
stay → end_submit	0.3×	0.0×	0.0×	0.0×	0.0×	0.31×
start → breakthrough	0.7×	0.7×	1.0×	0.9×	1.0×	0.31×
start → stuck	0.2×	0.2×	—	0.0×	—	0.15×
stuck → improved	0.1×	0.1×	—	0.0×	—	0.13×
start → improved	0.1×	0.2×	0.0×	0.1×	0.0×	0.13×
improved → end_limit	0.0×	0.1×	—	0.0×	0.0×	0.07×
stuck → breakthrough	0.1×	0.1×	—	0.0×	—	0.07×
improved → breakthrough	0.0×	0.1×	0.0×	0.1×	0.0×	0.06×
broke → improved	0.0×	0.1×	—	—	—	0.05×
improved → broke	0.0×	0.1×	—	—	—	0.05×
improved → stuck	0.2×	0.1×	—	—	—	0.05×
breakthrough → regressed	0.0×	0.0×	—	—	—	0.03×
churn → broke	0.0×	0.0×	—	—	—	0.03×
churn → churn	0.0×	0.0×	—	—	—	0.03×
stuck → regressed	0.0×	0.0×	—	—	—	0.03×
broke → stuck	0.1×	0.0×	—	—	—	0.03×
improved → churn	0.0×	0.0×	—	—	—	0.02×
improved → regressed	0.0×	0.0×	—	0.0×	—	0.02×
stuck → churn	0.0×	0.0×	—	—	—	0.02×
breakthrough → broke	0.0×	0.0×	—	—	—	0.02×
stay → end_limit	0.0×	0.0×	—	—	—	0.02×
stuck → broke	0.0×	0.0×	—	—	—	0.02×
improved → improved	0.0×	0.0×	—	0.0×	—	0.02×
broke → end_limit	0.0×	0.0×	—	—	—	0.01×
churn → end_limit	0.0×	0.0×	—	—	—	0.01×
regressed → breakthrough	0.0×	0.0×	—	—	—	0.01×
regressed → improved	0.0×	0.0×	—	—	—	0.01×
stuck → end_limit	0.1×	0.1×	—	—	—	0.00×
improved → end_submit	0.0×	0.0×	0.0×	0.0×	0.0×	0.00×
regressed → end_limit	0.0×	—	—	0.0×	—	0.00×
regressed → stuck	0.0×	0.0×	—	—	—	0.00×
breakthrough → end_limit	0.0×	—	—	—	—	0.00×
broke → breakthrough	0.0×	—	—	—	—	0.00×
churn → breakthrough	—	0.0×	—	—	—	0.00×
churn → improved	—	0.0×	—	—	—	0.00×
regressed → broke	0.0×	—	—	—	—	0.00×
regressed → churn	—	0.0×	—	—	—	0.00×
regressed → regressed	—	0.0×	—	—	—	0.00×
stay → broke	0.0×	—	—	—	—	0.00×
stuck → end_submit	0.0×	—	—	—	—	0.00×

Transition probabilities P(tgt | src) — sorted by Δ

transition	Gemma (100)	Qwen (96)	GPT-5.5 (94)	Sonnet 4.6 (90)	GLM-5.2 (78)	Δ (max−min)
regressed → end_limit	16.7%	—	—	100.0%	—	83.3pp
improved → breakthrough	6.2%	15.9%	75.0%	58.3%	33.3%	68.8pp
stay → end_submit	40.5%	60.0%	50.0%	40.0%	100.0%	60.0pp
breakthrough → end_submit	52.3%	88.8%	98.9%	97.7%	98.7%	46.6pp
breakthrough → stay	41.9%	5.0%	1.1%	2.3%	1.3%	40.8pp
stay → stay	54.4%	20.0%	50.0%	60.0%	—	40.0pp
broke → improved	30.8%	64.3%	—	—	—	33.5pp
improved → end_submit	3.1%	2.3%	25.0%	8.3%	33.3%	31.0pp
start → breakthrough	73.0%	65.6%	95.7%	86.7%	96.2%	30.6pp
improved → stuck	53.1%	27.3%	—	—	—	25.8pp
stuck → stuck	56.4%	31.2%	—	50.0%	—	25.2pp
broke → stuck	46.2%	21.4%	—	—	—	24.8pp
improved → end_limit	12.5%	18.2%	—	16.7%	33.3%	20.8pp
churn → end_limit	33.3%	15.4%	—	—	—	17.9pp
stuck → breakthrough	8.5%	12.5%	—	25.0%	—	16.5pp
stay → end_limit	3.8%	20.0%	—	—	—	16.2pp
start → stuck	17.0%	17.7%	—	2.2%	—	15.5pp
stuck → improved	14.9%	29.2%	—	25.0%	—	14.3pp
start → improved	10.0%	16.7%	4.3%	11.1%	3.8%	12.9pp
improved → broke	9.4%	18.2%	—	—	—	8.8pp
stuck → end_limit	8.5%	16.7%	—	—	—	8.2pp
broke → end_limit	7.7%	14.3%	—	—	—	6.6pp
regressed → stuck	16.7%	11.1%	—	—	—	5.6pp
regressed → breakthrough	16.7%	22.2%	—	—	—	5.5pp
regressed → improved	16.7%	22.2%	—	—	—	5.5pp
improved → regressed	3.1%	6.8%	—	8.3%	—	5.2pp
stuck → churn	1.1%	6.2%	—	—	—	5.1pp
improved → improved	9.4%	4.5%	—	8.3%	—	4.9pp
breakthrough → regressed	1.2%	5.0%	—	—	—	3.8pp
improved → churn	3.1%	6.8%	—	—	—	3.7pp
churn → broke	33.3%	30.8%	—	—	—	2.5pp
churn → churn	33.3%	30.8%	—	—	—	2.5pp
breakthrough → broke	3.5%	1.2%	—	—	—	2.3pp
stuck → regressed	4.3%	2.1%	—	—	—	2.2pp
stuck → broke	3.2%	2.1%	—	—	—	1.1pp
breakthrough → end_limit	1.2%	—	—	—	—	0.0pp
broke → breakthrough	15.4%	—	—	—	—	0.0pp
churn → breakthrough	—	15.4%	—	—	—	0.0pp
churn → improved	—	7.7%	—	—	—	0.0pp
regressed → broke	33.3%	—	—	—	—	0.0pp
regressed → churn	—	33.3%	—	—	—	0.0pp
regressed → regressed	—	11.1%	—	—	—	0.0pp
stay → broke	1.3%	—	—	—	—	0.0pp
stuck → end_submit	3.2%	—	—	—	—	0.0pp

Gap label distributions — sorted by edge visit rate, labels by avg%

transition	label	Gemma (100)	Qwen (96)	GPT-5.5 (94)	Sonnet 4.6 (90)	GLM-5.2 (78)	Δ (max−min)
start → breakthrough (0.83×/traj)	implement	100.0%	98.4%	100.0%	100.0%	100.0%	1.6pp
	explore	—	1.6%	—	—	—	0.0pp
breakthrough → end_submit (0.81×/traj)	verify	100.0%	100.0%	100.0%	100.0%	100.0%	0.0pp
stuck → stuck (0.14×/traj)	explore	32.1%	—	—	—	—	0.0pp
	implement	20.8%	26.7%	—	—	—	5.9pp
	fix_bug	11.3%	26.7%	—	100.0%	—	88.7pp
	debug_failure	13.2%	33.3%	—	—	—	20.1pp
	rewrite	17.0%	13.3%	—	—	—	3.7pp
	optimize_perf	3.8%	—	—	—	—	0.0pp
	infra_confusion	1.9%	—	—	—	—	0.0pp
stay → stay (0.10×/traj)	cosmetic_refactor	30.2%	—	—	—	—	0.0pp
	verify	23.3%	—	—	—	—	0.0pp
	fix_bug	11.6%	100.0%	—	100.0%	—	88.4pp
	optimize_perf	14.0%	—	—	—	—	0.0pp
	edge_case	11.6%	—	—	—	—	0.0pp
	revert	4.7%	—	—	—	—	0.0pp
	implement	4.7%	—	—	—	—	0.0pp
	infra_confusion	—	—	100.0%	—	—	0.0pp
start → improved (0.09×/traj)	implement	80.0%	87.5%	100.0%	100.0%	100.0%	20.0pp
	explore	10.0%	12.5%	—	—	—	2.5pp
	infra_confusion	10.0%	—	—	—	—	0.0pp
breakthrough → stay (0.09×/traj)	cosmetic_refactor	38.9%	—	—	—	—	0.0pp
	fix_bug	11.1%	75.0%	100.0%	100.0%	100.0%	88.9pp
	optimize_perf	19.4%	—	—	—	—	0.0pp
	implement	16.7%	—	—	—	—	0.0pp
	verify	8.3%	25.0%	—	—	—	16.7pp
	edge_case	5.6%	—	—	—	—	0.0pp
stay → end_submit (0.08×/traj)	verify	100.0%	100.0%	100.0%	100.0%	100.0%	0.0pp
start → stuck (0.07×/traj)	implement	47.1%	76.5%	—	50.0%	—	29.4pp
	explore	52.9%	23.5%	—	50.0%	—	29.4pp
stuck → improved (0.06×/traj)	fix_bug	42.9%	50.0%	—	—	—	7.1pp
	rewrite	28.6%	35.7%	—	100.0%	—	71.4pp
	implement	28.6%	14.3%	—	—	—	14.3pp
improved → stuck (0.06×/traj)	implement	23.5%	41.7%	—	—	—	18.2pp
	debug_failure	17.6%	16.7%	—	—	—	0.9pp
	infra_confusion	17.6%	8.3%	—	—	—	9.3pp
	explore	5.9%	8.3%	—	—	—	2.4pp
	optimize_perf	5.9%	8.3%	—	—	—	2.4pp
	fix_bug	11.8%	—	—	—	—	0.0pp
	cosmetic_refactor	5.9%	8.3%	—	—	—	2.4pp
	rewrite	5.9%	8.3%	—	—	—	2.4pp
	execute_plan	5.9%	—	—	—	—	0.0pp
improved → breakthrough (0.04×/traj)	fix_bug	100.0%	85.7%	66.7%	57.1%	—	42.9pp
	rewrite	—	14.3%	—	42.9%	—	28.6pp
	implement	—	—	—	—	100.0%	0.0pp
	debug_failure	—	—	33.3%	—	—	0.0pp
stuck → end_limit (0.03×/traj)	verify	100.0%	37.5%	—	—	—	62.5pp
	explore	—	25.0%	—	—	—	0.0pp
	debug_failure	—	25.0%	—	—	—	0.0pp
	fix_bug	—	12.5%	—	—	—	0.0pp
improved → end_limit (0.03×/traj)	verify	40.0%	50.0%	—	50.0%	—	10.0pp
	explore	20.0%	12.5%	—	50.0%	100.0%	87.5pp
	debug_failure	40.0%	25.0%	—	—	—	15.0pp
	fix_bug	—	12.5%	—	—	—	0.0pp
stuck → breakthrough (0.03×/traj)	fix_bug	50.0%	83.3%	—	100.0%	—	50.0pp
	implement	37.5%	—	—	—	—	0.0pp
	rewrite	—	16.7%	—	—	—	0.0pp
	execute_plan	12.5%	—	—	—	—	0.0pp
broke → improved (0.03×/traj)	rewrite	50.0%	22.2%	—	—	—	27.8pp
	implement	—	33.3%	—	—	—	0.0pp
	fix_bug	—	33.3%	—	—	—	0.0pp
	explore	50.0%	—	—	—	—	0.0pp
	execute_plan	—	11.1%	—	—	—	0.0pp
improved → broke (0.02×/traj)	explore	100.0%	25.0%	—	—	—	75.0pp
	implement	—	25.0%	—	—	—	0.0pp
	rewrite	—	25.0%	—	—	—	0.0pp
	fix_bug	—	12.5%	—	—	—	0.0pp
	debug_failure	—	12.5%	—	—	—	0.0pp
broke → stuck (0.02×/traj)	rewrite	33.3%	33.3%	—	—	—	0.0pp
	implement	16.7%	66.7%	—	—	—	50.0pp
	debug_failure	33.3%	—	—	—	—	0.0pp
	explore	16.7%	—	—	—	—	0.0pp
improved → improved (0.01×/traj)	implement	100.0%	—	—	—	—	0.0pp
	fix_bug	—	100.0%	—	100.0%	—	0.0pp
improved → end_submit (0.01×/traj)	verify	—	100.0%	100.0%	100.0%	100.0%	0.0pp
improved → regressed (0.01×/traj)	revert	—	33.3%	—	—	—	0.0pp
	implement	100.0%	—	—	—	—	0.0pp
	fix_bug	—	33.3%	—	—	—	0.0pp
	rewrite	—	33.3%	—	—	—	0.0pp
	debug_failure	—	—	—	100.0%	—	0.0pp
churn → broke (0.01×/traj)	rewrite	100.0%	50.0%	—	—	—	50.0pp
	explore	—	25.0%	—	—	—	0.0pp
	debug_failure	—	25.0%	—	—	—	0.0pp
breakthrough → regressed (0.01×/traj)	fix_bug	—	50.0%	—	—	—	0.0pp
	implement	—	25.0%	—	—	—	0.0pp
	cosmetic_refactor	100.0%	—	—	—	—	0.0pp
	debug_failure	—	25.0%	—	—	—	0.0pp
churn → churn (0.01×/traj)	rewrite	100.0%	25.0%	—	—	—	75.0pp
	fix_bug	—	50.0%	—	—	—	0.0pp
	implement	—	25.0%	—	—	—	0.0pp
stuck → regressed (0.01×/traj)	rewrite	50.0%	—	—	—	—	0.0pp
	fix_bug	25.0%	100.0%	—	—	—	75.0pp
	implement	25.0%	—	—	—	—	0.0pp
improved → churn (0.01×/traj)	rewrite	100.0%	33.3%	—	—	—	66.7pp
	fix_bug	—	66.7%	—	—	—	0.0pp
stuck → churn (0.01×/traj)	debug_failure	100.0%	66.7%	—	—	—	33.3pp
	rewrite	—	33.3%	—	—	—	0.0pp
breakthrough → broke (0.01×/traj)	debug_failure	33.3%	100.0%	—	—	—	66.7pp
	rewrite	33.3%	—	—	—	—	0.0pp
	implement	33.3%	—	—	—	—	0.0pp
stay → end_limit (0.01×/traj)	verify	100.0%	100.0%	—	—	—	0.0pp
stuck → broke (0.01×/traj)	implement	66.7%	—	—	—	—	0.0pp
	rewrite	33.3%	—	—	—	—	0.0pp
	debug_failure	—	100.0%	—	—	—	0.0pp
regressed → churn (0.01×/traj)	rewrite	—	66.7%	—	—	—	0.0pp
	fix_bug	—	33.3%	—	—	—	0.0pp
churn → end_limit (0.01×/traj)	debug_failure	—	100.0%	—	—	—	0.0pp
	verify	100.0%	—	—	—	—	0.0pp
regressed → improved (0.01×/traj)	rewrite	—	100.0%	—	—	—	0.0pp
	fix_bug	100.0%	—	—	—	—	0.0pp
regressed → breakthrough (0.01×/traj)	fix_bug	100.0%	100.0%	—	—	—	0.0pp
broke → end_limit (0.01×/traj)	infra_confusion	—	50.0%	—	—	—	0.0pp
	debug_failure	—	50.0%	—	—	—	0.0pp
	verify	100.0%	—	—	—	—	0.0pp
stuck → end_submit (0.01×/traj)	verify	100.0%	—	—	—	—	0.0pp
regressed → end_limit (0.00×/traj)	optimize_perf	100.0%	—	—	—	—	0.0pp
	debug_failure	—	—	—	100.0%	—	0.0pp
churn → breakthrough (0.00×/traj)	fix_bug	—	100.0%	—	—	—	0.0pp
regressed → stuck (0.00×/traj)	optimize_perf	100.0%	—	—	—	—	0.0pp
	debug_failure	—	100.0%	—	—	—	0.0pp
regressed → broke (0.00×/traj)	rewrite	100.0%	—	—	—	—	0.0pp
broke → breakthrough (0.00×/traj)	implement	50.0%	—	—	—	—	0.0pp
	fix_bug	50.0%	—	—	—	—	0.0pp
churn → improved (0.00×/traj)	implement	—	100.0%	—	—	—	0.0pp
regressed → regressed (0.00×/traj)	fix_bug	—	100.0%	—	—	—	0.0pp
stay → broke (0.00×/traj)	debug_failure	100.0%	—	—	—	—	0.0pp
breakthrough → end_limit (0.00×/traj)	verify	100.0%	—	—	—	—	0.0pp

Rust

highest in row lowest in row Sorted by Δ (max − min) descending.

Node visit rates — sorted by Δ

state	Gemma (100)	Qwen (96)	GPT-5.5 (94)	Sonnet 4.6 (90)	GLM-5.2 (78)	Δ
start	1.0	1.0	1.0	1.0	1.0	0.00
improved	0.3	0.3	0.0	0.1	0.0	0.29
churn	0.1	0.0	—	—	—	0.03
stuck	1.1	0.6	—	0.1	0.1	1.09
breakthrough	0.9	0.9	1.0	1.0	1.0	0.12
stay	2.4	0.1	0.0	0.0	—	2.34
regressed	0.1	0.0	—	—	—	0.03
broke	0.2	0.1	—	—	—	0.12
end_submit	0.8	0.8	1.0	1.0	1.0	0.22
end_limit	0.2	0.2	0.0	0.0	0.0	0.22

Edge visit rates (count / n_problems) — sorted by Δ

transition	Gemma (100)	Qwen (96)	GPT-5.5 (94)	Sonnet 4.6 (90)	GLM-5.2 (78)	Δ (max−min)
stay → stay	1.8×	0.0×	0.0×	0.0×	—	1.79×
stuck → stuck	0.7×	0.3×	—	0.0×	—	0.67×
breakthrough → end_submit	0.3×	0.7×	1.0×	0.9×	0.9×	0.64×
breakthrough → stay	0.6×	0.1×	0.0×	0.0×	—	0.55×
stay → end_submit	0.4×	0.1×	0.0×	0.0×	—	0.40×
start → breakthrough	0.7×	0.6×	1.0×	0.9×	0.9×	0.37×
start → stuck	0.2×	0.3×	—	0.0×	0.1×	0.25×
improved → stuck	0.1×	0.0×	—	0.0×	—	0.14×
stuck → breakthrough	0.1×	0.2×	—	0.0×	0.0×	0.13×
start → improved	0.1×	0.1×	0.0×	0.1×	0.0×	0.10×
stay → end_limit	0.1×	0.0×	—	—	—	0.09×
stuck → improved	0.2×	0.1×	—	—	—	0.08×
improved → end_limit	0.0×	0.1×	0.0×	0.0×	—	0.07×
broke → breakthrough	0.1×	0.0×	—	—	—	0.07×
stuck → end_limit	0.1×	0.1×	—	—	0.0×	0.07×
improved → breakthrough	0.0×	0.1×	0.0×	0.0×	0.0×	0.06×
broke → stuck	0.1×	0.0×	—	—	—	0.05×
improved → broke	0.1×	0.0×	—	—	—	0.05×
breakthrough → broke	0.0×	0.0×	—	—	—	0.03×
improved → improved	0.0×	0.0×	0.0×	—	0.0×	0.02×
stuck → broke	0.0×	0.0×	—	—	—	0.02×
improved → churn	0.0×	0.0×	—	—	—	0.01×
churn → churn	0.0×	0.0×	—	—	—	0.01×
improved → regressed	0.0×	0.0×	—	—	—	0.01×
breakthrough → end_limit	0.0×	—	—	—	0.0×	0.01×
improved → end_submit	—	—	0.0×	0.0×	0.0×	0.00×
broke → improved	0.0×	0.0×	—	—	—	0.00×
stuck → regressed	0.0×	0.0×	—	—	—	0.00×
churn → broke	—	0.0×	—	—	—	0.00×
churn → end_limit	0.0×	—	—	—	—	0.00×
churn → improved	0.0×	—	—	—	—	0.00×
churn → regressed	0.0×	—	—	—	—	0.00×
regressed → breakthrough	0.0×	—	—	—	—	0.00×
regressed → broke	—	0.0×	—	—	—	0.00×
regressed → churn	0.0×	—	—	—	—	0.00×
regressed → end_limit	—	0.0×	—	—	—	0.00×
regressed → improved	0.0×	—	—	—	—	0.00×
regressed → stuck	0.0×	—	—	—	—	0.00×
stay → broke	0.1×	—	—	—	—	0.00×
stuck → churn	0.0×	—	—	—	—	0.00×
stuck → end_submit	0.0×	—	—	—	—	0.00×

Transition probabilities P(tgt | src) — sorted by Δ

transition	Gemma (100)	Qwen (96)	GPT-5.5 (94)	Sonnet 4.6 (90)	GLM-5.2 (78)	Δ (max−min)
stuck → breakthrough	9.6%	25.8%	—	80.0%	75.0%	70.4pp
breakthrough → end_submit	34.7%	86.6%	98.9%	98.8%	97.4%	64.2pp
breakthrough → stay	58.9%	12.2%	1.1%	1.2%	—	57.8pp
stay → stay	76.3%	23.1%	50.0%	66.7%	—	53.2pp
stay → end_submit	17.4%	69.2%	50.0%	33.3%	—	51.8pp
improved → end_limit	6.1%	28.6%	25.0%	50.0%	—	43.9pp
stuck → stuck	59.6%	43.5%	—	20.0%	—	39.6pp
start → breakthrough	73.0%	59.4%	96.8%	90.0%	92.3%	37.4pp
improved → stuck	45.5%	14.3%	—	16.7%	—	31.2pp
improved → breakthrough	3.0%	25.0%	25.0%	16.7%	33.3%	30.3pp
improved → improved	9.1%	7.1%	25.0%	—	33.3%	26.2pp
start → stuck	21.0%	28.1%	—	3.3%	5.1%	24.8pp
broke → improved	18.2%	40.0%	—	—	—	21.8pp
broke → breakthrough	40.9%	20.0%	—	—	—	20.9pp
stuck → end_limit	7.0%	9.7%	—	—	25.0%	18.0pp
improved → end_submit	—	—	25.0%	16.7%	33.3%	16.6pp
improved → broke	27.3%	14.3%	—	—	—	13.0pp
start → improved	6.0%	12.5%	3.2%	6.7%	2.6%	9.9pp
improved → churn	3.0%	7.1%	—	—	—	4.1pp
stay → end_limit	4.2%	7.7%	—	—	—	3.5pp
breakthrough → broke	4.2%	1.2%	—	—	—	3.0pp
improved → regressed	6.1%	3.6%	—	—	—	2.5pp
broke → stuck	40.9%	40.0%	—	—	—	0.9pp
stuck → regressed	0.9%	1.6%	—	—	—	0.7pp
breakthrough → end_limit	2.1%	—	—	—	2.6%	0.5pp
stuck → improved	15.8%	16.1%	—	—	—	0.3pp
stuck → broke	3.5%	3.2%	—	—	—	0.3pp
churn → broke	—	66.7%	—	—	—	0.0pp
churn → churn	33.3%	33.3%	—	—	—	0.0pp
churn → end_limit	16.7%	—	—	—	—	0.0pp
churn → improved	16.7%	—	—	—	—	0.0pp
churn → regressed	33.3%	—	—	—	—	0.0pp
regressed → breakthrough	20.0%	—	—	—	—	0.0pp
regressed → broke	—	50.0%	—	—	—	0.0pp
regressed → churn	40.0%	—	—	—	—	0.0pp
regressed → end_limit	—	50.0%	—	—	—	0.0pp
regressed → improved	20.0%	—	—	—	—	0.0pp
regressed → stuck	20.0%	—	—	—	—	0.0pp
stay → broke	2.1%	—	—	—	—	0.0pp
stuck → churn	0.9%	—	—	—	—	0.0pp
stuck → end_submit	2.6%	—	—	—	—	0.0pp

Gap label distributions — sorted by edge visit rate, labels by avg%

transition	label	Gemma (100)	Qwen (96)	GPT-5.5 (94)	Sonnet 4.6 (90)	GLM-5.2 (78)	Δ (max−min)
start → breakthrough (0.82×/traj)	implement	100.0%	100.0%	100.0%	100.0%	100.0%	0.0pp
breakthrough → end_submit (0.79×/traj)	verify	100.0%	100.0%	100.0%	100.0%	100.0%	0.0pp
stay → stay (0.37×/traj)	cosmetic_refactor	37.2%	66.7%	—	—	—	29.5pp
	verify	19.4%	—	—	—	—	0.0pp
	fix_bug	12.8%	—	—	100.0%	—	87.2pp
	implement	7.8%	—	—	—	—	0.0pp
	optimize_perf	7.2%	—	—	—	—	0.0pp
	revert	6.7%	—	—	—	—	0.0pp
	edge_case	5.6%	33.3%	—	—	—	27.7pp
	infra_confusion	2.2%	—	100.0%	—	—	97.8pp
	debug_failure	1.1%	—	—	—	—	0.0pp
stuck → stuck (0.19×/traj)	explore	44.1%	—	—	—	—	0.0pp
	implement	20.6%	22.2%	—	—	—	1.6pp
	rewrite	14.7%	25.9%	—	—	—	11.2pp
	fix_bug	5.9%	37.0%	—	100.0%	—	94.1pp
	verify	5.9%	—	—	—	—	0.0pp
	debug_failure	1.5%	11.1%	—	—	—	9.6pp
	cosmetic_refactor	4.4%	—	—	—	—	0.0pp
	execute_plan	1.5%	—	—	—	—	0.0pp
	optimize_perf	1.5%	—	—	—	—	0.0pp
	infra_confusion	—	3.7%	—	—	—	0.0pp
breakthrough → stay (0.14×/traj)	implement	32.1%	20.0%	—	—	—	12.1pp
	cosmetic_refactor	25.0%	40.0%	—	—	—	15.0pp
	fix_bug	8.9%	30.0%	100.0%	100.0%	—	91.1pp
	edge_case	16.1%	10.0%	—	—	—	6.1pp
	optimize_perf	10.7%	—	—	—	—	0.0pp
	infra_confusion	1.8%	—	—	—	—	0.0pp
	revert	1.8%	—	—	—	—	0.0pp
	verify	1.8%	—	—	—	—	0.0pp
	debug_failure	1.8%	—	—	—	—	0.0pp
start → stuck (0.12×/traj)	implement	42.9%	85.2%	—	100.0%	100.0%	57.1pp
	explore	57.1%	14.8%	—	—	—	42.3pp
stay → end_submit (0.11×/traj)	verify	100.0%	100.0%	100.0%	—	—	0.0pp
	infra_confusion	—	—	—	100.0%	—	0.0pp
stuck → breakthrough (0.07×/traj)	fix_bug	27.3%	87.5%	—	75.0%	100.0%	72.7pp
	implement	63.6%	—	—	—	—	0.0pp
	infra_confusion	—	12.5%	—	—	—	0.0pp
	rewrite	9.1%	—	—	25.0%	—	15.9pp
start → improved (0.06×/traj)	implement	100.0%	100.0%	100.0%	100.0%	100.0%	0.0pp
stuck → improved (0.06×/traj)	implement	27.8%	20.0%	—	—	—	7.8pp
	rewrite	33.3%	10.0%	—	—	—	23.3pp
	fix_bug	5.6%	50.0%	—	—	—	44.4pp
	revert	11.1%	—	—	—	—	0.0pp
	explore	11.1%	—	—	—	—	0.0pp
	debug_failure	5.6%	10.0%	—	—	—	4.4pp
	execute_plan	5.6%	—	—	—	—	0.0pp
	optimize_perf	—	10.0%	—	—	—	0.0pp
improved → stuck (0.04×/traj)	implement	26.7%	25.0%	—	—	—	1.7pp
	cosmetic_refactor	26.7%	25.0%	—	—	—	1.7pp
	rewrite	13.3%	50.0%	—	—	—	36.7pp
	explore	13.3%	—	—	—	—	0.0pp
	fix_bug	6.7%	—	—	100.0%	—	93.3pp
	verify	6.7%	—	—	—	—	0.0pp
	debug_failure	6.7%	—	—	—	—	0.0pp
stuck → end_limit (0.03×/traj)	verify	100.0%	66.7%	—	—	—	33.3pp
	explore	—	33.3%	—	—	100.0%	66.7pp
improved → end_limit (0.03×/traj)	debug_failure	—	50.0%	—	66.7%	—	16.7pp
	verify	100.0%	25.0%	100.0%	—	—	75.0pp
	explore	—	25.0%	—	33.3%	—	8.3pp
broke → stuck (0.03×/traj)	explore	33.3%	25.0%	—	—	—	8.3pp
	implement	33.3%	25.0%	—	—	—	8.3pp
	rewrite	22.2%	25.0%	—	—	—	2.8pp
	fix_bug	—	25.0%	—	—	—	0.0pp
	debug_failure	11.1%	—	—	—	—	0.0pp
improved → broke (0.03×/traj)	explore	44.4%	25.0%	—	—	—	19.4pp
	rewrite	33.3%	50.0%	—	—	—	16.7pp
	execute_plan	11.1%	—	—	—	—	0.0pp
	implement	11.1%	—	—	—	—	0.0pp
	fix_bug	—	25.0%	—	—	—	0.0pp
improved → breakthrough (0.02×/traj)	fix_bug	—	71.4%	100.0%	—	100.0%	28.6pp
	rewrite	—	14.3%	—	100.0%	—	85.7pp
	debug_failure	100.0%	—	—	—	—	0.0pp
	execute_plan	—	14.3%	—	—	—	0.0pp
broke → breakthrough (0.02×/traj)	fix_bug	66.7%	50.0%	—	—	—	16.7pp
	implement	33.3%	—	—	—	—	0.0pp
	rewrite	—	50.0%	—	—	—	0.0pp
stay → end_limit (0.02×/traj)	verify	100.0%	—	—	—	—	0.0pp
	debug_failure	—	100.0%	—	—	—	0.0pp
broke → improved (0.02×/traj)	rewrite	50.0%	50.0%	—	—	—	0.0pp
	execute_plan	25.0%	—	—	—	—	0.0pp
	revert	25.0%	—	—	—	—	0.0pp
	explore	—	25.0%	—	—	—	0.0pp
	fix_bug	—	25.0%	—	—	—	0.0pp
improved → improved (0.01×/traj)	rewrite	33.3%	100.0%	100.0%	—	—	66.7pp
	fix_bug	33.3%	—	—	—	100.0%	66.7pp
	implement	33.3%	—	—	—	—	0.0pp
stuck → broke (0.01×/traj)	explore	75.0%	—	—	—	—	0.0pp
	rewrite	25.0%	50.0%	—	—	—	25.0pp
	debug_failure	—	50.0%	—	—	—	0.0pp
breakthrough → broke (0.01×/traj)	rewrite	25.0%	100.0%	—	—	—	75.0pp
	explore	25.0%	—	—	—	—	0.0pp
	implement	25.0%	—	—	—	—	0.0pp
	debug_failure	25.0%	—	—	—	—	0.0pp
stay → broke (0.01×/traj)	optimize_perf	40.0%	—	—	—	—	0.0pp
	explore	20.0%	—	—	—	—	0.0pp
	implement	20.0%	—	—	—	—	0.0pp
	fix_bug	20.0%	—	—	—	—	0.0pp
breakthrough → end_limit (0.01×/traj)	verify	100.0%	—	—	—	66.7%	33.3pp
	fix_bug	—	—	—	—	33.3%	0.0pp
improved → end_submit (0.01×/traj)	verify	—	—	100.0%	100.0%	100.0%	0.0pp
improved → churn (0.01×/traj)	rewrite	100.0%	—	—	—	—	0.0pp
	implement	—	50.0%	—	—	—	0.0pp
	fix_bug	—	50.0%	—	—	—	0.0pp
improved → regressed (0.01×/traj)	implement	100.0%	—	—	—	—	0.0pp
	fix_bug	—	100.0%	—	—	—	0.0pp
churn → churn (0.01×/traj)	rewrite	50.0%	—	—	—	—	0.0pp
	implement	50.0%	—	—	—	—	0.0pp
	fix_bug	—	100.0%	—	—	—	0.0pp
stuck → end_submit (0.01×/traj)	verify	100.0%	—	—	—	—	0.0pp
churn → broke (0.00×/traj)	rewrite	—	50.0%	—	—	—	0.0pp
	implement	—	50.0%	—	—	—	0.0pp
stuck → regressed (0.00×/traj)	explore	—	100.0%	—	—	—	0.0pp
	implement	100.0%	—	—	—	—	0.0pp
regressed → churn (0.00×/traj)	rewrite	50.0%	—	—	—	—	0.0pp
	implement	50.0%	—	—	—	—	0.0pp
churn → regressed (0.00×/traj)	implement	50.0%	—	—	—	—	0.0pp
	fix_bug	50.0%	—	—	—	—	0.0pp
regressed → broke (0.00×/traj)	rewrite	—	100.0%	—	—	—	0.0pp
regressed → end_limit (0.00×/traj)	explore	—	100.0%	—	—	—	0.0pp
regressed → stuck (0.00×/traj)	debug_failure	100.0%	—	—	—	—	0.0pp
churn → improved (0.00×/traj)	fix_bug	100.0%	—	—	—	—	0.0pp
churn → end_limit (0.00×/traj)	verify	100.0%	—	—	—	—	0.0pp
regressed → improved (0.00×/traj)	debug_failure	100.0%	—	—	—	—	0.0pp
regressed → breakthrough (0.00×/traj)	debug_failure	100.0%	—	—	—	—	0.0pp
stuck → churn (0.00×/traj)	explore	100.0%	—	—	—	—	0.0pp

Ocaml

highest in row lowest in row Sorted by Δ (max − min) descending.

Node visit rates — sorted by Δ

state	Gemma (100)	Qwen (93)	GPT-5.5 (92)	Sonnet 4.6 (91)	GLM-5.2 (77)	Δ
start	1.0	1.0	1.0	1.0	1.0	0.00
improved	0.4	0.3	0.1	0.2	0.0	0.33
churn	—	0.0	—	—	—	0.00
stuck	13.4	5.3	0.4	0.1	0.4	13.31
breakthrough	0.9	0.8	1.0	0.9	1.0	0.18
stay	0.9	0.1	0.0	0.0	0.0	0.88
regressed	0.0	0.0	—	0.0	—	0.01
broke	0.6	0.2	—	0.0	—	0.59
end_submit	0.5	0.7	1.0	1.0	1.0	0.51
end_limit	0.5	0.3	0.0	0.0	0.0	0.51

Edge visit rates (count / n_problems) — sorted by Δ

transition	Gemma (100)	Qwen (93)	GPT-5.5 (92)	Sonnet 4.6 (91)	GLM-5.2 (77)	Δ (max−min)
stuck → stuck	12.0×	4.4×	0.1×	—	0.1×	11.91×
start → stuck	0.8×	0.8×	0.3×	0.0×	0.3×	0.75×
breakthrough → end_submit	0.3×	0.6×	1.0×	0.9×	1.0×	0.63×
start → breakthrough	0.2×	0.2×	0.7×	0.8×	0.7×	0.59×
stay → stay	0.6×	0.0×	0.0×	0.0×	—	0.59×
stuck → breakthrough	0.5×	0.5×	0.3×	0.1×	0.3×	0.49×
stuck → end_limit	0.4×	0.2×	—	—	0.0×	0.41×
breakthrough → stay	0.3×	0.1×	0.0×	0.0×	0.0×	0.28×
broke → stuck	0.4×	0.1×	—	—	—	0.26×
stuck → improved	0.3×	0.2×	0.0×	—	—	0.25×
improved → stuck	0.2×	0.1×	—	0.0×	—	0.18×
breakthrough → broke	0.2×	0.0×	—	—	—	0.18×
start → improved	0.0×	0.0×	0.0×	0.2×	0.0×	0.15×
stay → end_submit	0.1×	0.1×	0.0×	0.0×	—	0.13×
improved → broke	0.1×	0.1×	—	0.0×	—	0.12×
broke → breakthrough	0.1×	0.0×	—	—	—	0.11×
stuck → broke	0.1×	0.0×	—	—	—	0.10×
broke → improved	0.1×	0.1×	—	0.0×	—	0.08×
improved → breakthrough	0.0×	0.1×	0.0×	0.1×	0.0×	0.08×
improved → end_limit	0.0×	0.0×	0.0×	0.0×	—	0.03×
stay → end_limit	0.0×	—	—	—	0.0×	0.03×
improved → regressed	—	0.0×	—	0.0×	—	0.01×
improved → improved	—	0.0×	—	0.0×	—	0.01×
regressed → broke	0.0×	0.0×	—	—	—	0.01×
breakthrough → end_limit	0.0×	0.0×	—	—	—	0.00×
breakthrough → regressed	0.0×	0.0×	—	—	—	0.00×
improved → end_submit	—	0.0×	0.0×	0.0×	—	0.00×
regressed → improved	—	0.0×	—	0.0×	—	0.00×
broke → end_limit	0.0×	—	—	—	—	0.00×
churn → churn	—	0.0×	—	—	—	0.00×
churn → improved	—	0.0×	—	—	—	0.00×
improved → churn	—	0.0×	—	—	—	0.00×
regressed → breakthrough	—	0.0×	—	—	—	0.00×
regressed → end_limit	—	—	—	0.0×	—	0.00×
regressed → stuck	0.0×	—	—	—	—	0.00×
stay → broke	0.1×	—	—	—	—	0.00×
stay → regressed	0.0×	—	—	—	—	0.00×
stuck → end_submit	—	—	0.0×	—	—	0.00×

Transition probabilities P(tgt | src) — sorted by Δ

transition	Gemma (100)	Qwen (93)	GPT-5.5 (92)	Sonnet 4.6 (91)	GLM-5.2 (77)	Δ (max−min)
improved → breakthrough	2.8%	18.8%	60.0%	44.4%	100.0%	97.2pp
stuck → breakthrough	4.0%	8.9%	62.5%	100.0%	74.2%	96.0pp
stay → end_limit	4.5%	—	—	—	100.0%	95.5pp
broke → improved	15.0%	29.4%	—	100.0%	—	85.0pp
start → stuck	78.0%	75.3%	29.3%	3.3%	31.2%	74.7pp
stay → end_submit	15.7%	84.6%	50.0%	50.0%	—	68.9pp
stuck → stuck	89.8%	82.3%	32.5%	—	22.6%	67.2pp
breakthrough → end_submit	38.2%	74.7%	98.9%	98.8%	98.7%	60.7pp
start → breakthrough	21.0%	23.7%	66.3%	80.2%	66.2%	59.2pp
stay → stay	67.4%	15.4%	50.0%	50.0%	—	52.0pp
improved → stuck	55.6%	25.0%	—	11.1%	—	44.5pp
regressed → broke	66.7%	33.3%	—	—	—	33.4pp
breakthrough → stay	32.6%	14.7%	1.1%	1.2%	1.3%	31.5pp
improved → broke	36.1%	28.1%	—	5.6%	—	30.5pp
breakthrough → broke	23.6%	4.0%	—	—	—	19.6pp
improved → end_submit	—	3.1%	20.0%	5.6%	—	16.9pp
regressed → improved	—	33.3%	—	50.0%	—	16.7pp
start → improved	1.0%	1.1%	4.3%	16.5%	2.6%	15.5pp
improved → end_limit	5.6%	12.5%	20.0%	16.7%	—	14.4pp
broke → breakthrough	21.7%	11.8%	—	—	—	9.9pp
improved → regressed	—	3.1%	—	11.1%	—	8.0pp
broke → stuck	61.7%	58.8%	—	—	—	2.9pp
stuck → improved	1.9%	4.4%	2.5%	—	—	2.5pp
improved → improved	—	6.2%	—	5.6%	—	0.6pp
breakthrough → end_limit	3.4%	4.0%	—	—	—	0.6pp
breakthrough → regressed	2.2%	2.7%	—	—	—	0.5pp
stuck → end_limit	3.1%	3.6%	—	—	3.2%	0.5pp
stuck → broke	1.0%	0.8%	—	—	—	0.2pp
broke → end_limit	1.7%	—	—	—	—	0.0pp
churn → churn	—	50.0%	—	—	—	0.0pp
churn → improved	—	50.0%	—	—	—	0.0pp
improved → churn	—	3.1%	—	—	—	0.0pp
regressed → breakthrough	—	33.3%	—	—	—	0.0pp
regressed → end_limit	—	—	—	50.0%	—	0.0pp
regressed → stuck	33.3%	—	—	—	—	0.0pp
stay → broke	11.2%	—	—	—	—	0.0pp
stay → regressed	1.1%	—	—	—	—	0.0pp
stuck → end_submit	—	—	2.5%	—	—	0.0pp

Gap label distributions — sorted by edge visit rate, labels by avg%

transition	label	Gemma (100)	Qwen (93)	GPT-5.5 (92)	Sonnet 4.6 (91)	GLM-5.2 (77)	Δ (max−min)
stuck → stuck (3.32×/traj)	revert	31.2%	—	—	—	—	0.0pp
	fix_bug	14.2%	47.1%	15.4%	—	85.7%	71.5pp
	debug_failure	18.6%	10.3%	7.7%	—	14.3%	10.9pp
	implement	15.0%	12.5%	15.4%	—	—	2.9pp
	infra_confusion	4.8%	16.4%	61.5%	—	—	56.7pp
	explore	7.1%	3.7%	—	—	—	3.4pp
	cosmetic_refactor	4.8%	1.0%	—	—	—	3.8pp
	rewrite	2.5%	7.6%	—	—	—	5.1pp
	execute_plan	1.4%	1.2%	—	—	—	0.2pp
	verify	0.4%	0.2%	—	—	—	0.2pp
	optimize_perf	0.1%	—	—	—	—	0.0pp
breakthrough → end_submit (0.76×/traj)	verify	100.0%	98.2%	100.0%	100.0%	100.0%	1.8pp
	fix_bug	—	1.8%	—	—	—	0.0pp
start → breakthrough (0.51×/traj)	implement	100.0%	100.0%	100.0%	100.0%	100.0%	0.0pp
start → stuck (0.43×/traj)	implement	76.9%	81.4%	96.3%	100.0%	100.0%	23.1pp
	explore	23.1%	18.6%	3.7%	—	—	19.4pp
stuck → breakthrough (0.33×/traj)	fix_bug	61.1%	72.7%	36.0%	100.0%	65.2%	64.0pp
	infra_confusion	9.3%	15.9%	52.0%	—	30.4%	42.7pp
	cosmetic_refactor	11.1%	4.5%	4.0%	—	4.3%	7.1pp
	execute_plan	5.6%	—	4.0%	—	—	1.6pp
	implement	3.7%	2.3%	4.0%	—	—	1.7pp
	rewrite	3.7%	4.5%	—	—	—	0.8pp
	debug_failure	5.6%	—	—	—	—	0.0pp
stay → stay (0.13×/traj)	verify	28.3%	50.0%	—	—	—	21.7pp
	cosmetic_refactor	18.3%	50.0%	—	—	—	31.7pp
	fix_bug	15.0%	—	—	100.0%	—	85.0pp
	edge_case	11.7%	—	—	—	—	0.0pp
	implement	10.0%	—	—	—	—	0.0pp
	revert	8.3%	—	—	—	—	0.0pp
	infra_confusion	1.7%	—	100.0%	—	—	98.3pp
	optimize_perf	3.3%	—	—	—	—	0.0pp
	debug_failure	3.3%	—	—	—	—	0.0pp
stuck → end_limit (0.13×/traj)	verify	71.4%	33.3%	—	—	100.0%	66.7pp
	debug_failure	19.0%	16.7%	—	—	—	2.3pp
	explore	4.8%	33.3%	—	—	—	28.5pp
	infra_confusion	—	16.7%	—	—	—	0.0pp
	implement	2.4%	—	—	—	—	0.0pp
	rewrite	2.4%	—	—	—	—	0.0pp
stuck → improved (0.10×/traj)	fix_bug	34.6%	77.3%	—	—	—	42.7pp
	infra_confusion	15.4%	18.2%	100.0%	—	—	84.6pp
	cosmetic_refactor	23.1%	—	—	—	—	0.0pp
	implement	15.4%	4.5%	—	—	—	10.9pp
	revert	11.5%	—	—	—	—	0.0pp
broke → stuck (0.10×/traj)	revert	24.3%	—	—	—	—	0.0pp
	fix_bug	21.6%	10.0%	—	—	—	11.6pp
	debug_failure	24.3%	—	—	—	—	0.0pp
	explore	10.8%	20.0%	—	—	—	9.2pp
	implement	5.4%	20.0%	—	—	—	14.6pp
	rewrite	2.7%	30.0%	—	—	—	27.3pp
	infra_confusion	2.7%	20.0%	—	—	—	17.3pp
	cosmetic_refactor	5.4%	—	—	—	—	0.0pp
	execute_plan	2.7%	—	—	—	—	0.0pp
breakthrough → stay (0.09×/traj)	fix_bug	10.3%	27.3%	100.0%	100.0%	100.0%	89.7pp
	verify	27.6%	9.1%	—	—	—	18.5pp
	cosmetic_refactor	10.3%	45.5%	—	—	—	35.2pp
	infra_confusion	13.8%	18.2%	—	—	—	4.4pp
	edge_case	17.2%	—	—	—	—	0.0pp
	optimize_perf	10.3%	—	—	—	—	0.0pp
	implement	3.4%	—	—	—	—	0.0pp
	rewrite	3.4%	—	—	—	—	0.0pp
	debug_failure	3.4%	—	—	—	—	0.0pp
improved → stuck (0.06×/traj)	fix_bug	30.0%	12.5%	—	100.0%	—	87.5pp
	revert	30.0%	—	—	—	—	0.0pp
	debug_failure	5.0%	50.0%	—	—	—	45.0pp
	verify	20.0%	—	—	—	—	0.0pp
	implement	—	25.0%	—	—	—	0.0pp
	explore	—	12.5%	—	—	—	0.0pp
	cosmetic_refactor	5.0%	—	—	—	—	0.0pp
	rewrite	5.0%	—	—	—	—	0.0pp
	infra_confusion	5.0%	—	—	—	—	0.0pp
stay → end_submit (0.06×/traj)	verify	100.0%	100.0%	100.0%	—	—	0.0pp
	infra_confusion	—	—	—	100.0%	—	0.0pp
start → improved (0.05×/traj)	implement	100.0%	100.0%	100.0%	100.0%	100.0%	0.0pp
breakthrough → broke (0.05×/traj)	fix_bug	28.6%	—	—	—	—	0.0pp
	infra_confusion	23.8%	33.3%	—	—	—	9.5pp
	debug_failure	23.8%	33.3%	—	—	—	9.5pp
	optimize_perf	9.5%	—	—	—	—	0.0pp
	edge_case	4.8%	—	—	—	—	0.0pp
	revert	4.8%	—	—	—	—	0.0pp
	explore	—	33.3%	—	—	—	0.0pp
	rewrite	4.8%	—	—	—	—	0.0pp
improved → broke (0.05×/traj)	fix_bug	38.5%	44.4%	—	100.0%	—	61.5pp
	rewrite	7.7%	22.2%	—	—	—	14.5pp
	debug_failure	7.7%	22.2%	—	—	—	14.5pp
	explore	7.7%	11.1%	—	—	—	3.4pp
	implement	15.4%	—	—	—	—	0.0pp
	cosmetic_refactor	15.4%	—	—	—	—	0.0pp
	infra_confusion	7.7%	—	—	—	—	0.0pp
improved → breakthrough (0.04×/traj)	fix_bug	—	100.0%	100.0%	37.5%	100.0%	62.5pp
	rewrite	—	—	—	37.5%	—	0.0pp
	implement	—	—	—	12.5%	—	0.0pp
	infra_confusion	100.0%	—	—	—	—	0.0pp
	debug_failure	—	—	—	12.5%	—	0.0pp
stuck → broke (0.04×/traj)	fix_bug	42.9%	—	—	—	—	0.0pp
	debug_failure	14.3%	50.0%	—	—	—	35.7pp
	explore	7.1%	25.0%	—	—	—	17.9pp
	implement	14.3%	—	—	—	—	0.0pp
	rewrite	7.1%	25.0%	—	—	—	17.9pp
	revert	7.1%	—	—	—	—	0.0pp
	infra_confusion	7.1%	—	—	—	—	0.0pp
broke → improved (0.03×/traj)	fix_bug	22.2%	60.0%	—	—	—	37.8pp
	cosmetic_refactor	22.2%	—	—	100.0%	—	77.8pp
	debug_failure	22.2%	20.0%	—	—	—	2.2pp
	revert	22.2%	—	—	—	—	0.0pp
	implement	11.1%	20.0%	—	—	—	8.9pp
broke → breakthrough (0.03×/traj)	fix_bug	53.8%	—	—	—	—	0.0pp
	implement	23.1%	50.0%	—	—	—	26.9pp
	cosmetic_refactor	7.7%	50.0%	—	—	—	42.3pp
	infra_confusion	7.7%	—	—	—	—	0.0pp
	debug_failure	7.7%	—	—	—	—	0.0pp
improved → end_limit (0.02×/traj)	explore	—	25.0%	100.0%	100.0%	—	75.0pp
	verify	50.0%	75.0%	—	—	—	25.0pp
	debug_failure	50.0%	—	—	—	—	0.0pp
stay → broke (0.02×/traj)	debug_failure	40.0%	—	—	—	—	0.0pp
	implement	20.0%	—	—	—	—	0.0pp
	fix_bug	20.0%	—	—	—	—	0.0pp
	rewrite	10.0%	—	—	—	—	0.0pp
	infra_confusion	10.0%	—	—	—	—	0.0pp
breakthrough → end_limit (0.01×/traj)	verify	100.0%	100.0%	—	—	—	0.0pp
stay → end_limit (0.01×/traj)	verify	80.0%	—	—	—	—	0.0pp
	infra_confusion	20.0%	—	—	—	100.0%	80.0pp
breakthrough → regressed (0.01×/traj)	fix_bug	50.0%	100.0%	—	—	—	50.0pp
	infra_confusion	50.0%	—	—	—	—	0.0pp
improved → regressed (0.01×/traj)	rewrite	—	—	—	100.0%	—	0.0pp
	fix_bug	—	100.0%	—	—	—	0.0pp
improved → end_submit (0.01×/traj)	verify	—	100.0%	100.0%	100.0%	—	0.0pp
improved → improved (0.01×/traj)	fix_bug	—	100.0%	—	100.0%	—	0.0pp
regressed → broke (0.01×/traj)	infra_confusion	—	100.0%	—	—	—	0.0pp
	implement	50.0%	—	—	—	—	0.0pp
	fix_bug	50.0%	—	—	—	—	0.0pp
regressed → improved (0.00×/traj)	fix_bug	—	100.0%	—	100.0%	—	0.0pp
regressed → end_limit (0.00×/traj)	fix_bug	—	—	—	100.0%	—	0.0pp
stuck → end_submit (0.00×/traj)	infra_confusion	—	—	100.0%	—	—	0.0pp
churn → improved (0.00×/traj)	fix_bug	—	100.0%	—	—	—	0.0pp
improved → churn (0.00×/traj)	fix_bug	—	100.0%	—	—	—	0.0pp
regressed → breakthrough (0.00×/traj)	fix_bug	—	100.0%	—	—	—	0.0pp
churn → churn (0.00×/traj)	fix_bug	—	100.0%	—	—	—	0.0pp
regressed → stuck (0.00×/traj)	cosmetic_refactor	100.0%	—	—	—	—	0.0pp
broke → end_limit (0.00×/traj)	verify	100.0%	—	—	—	—	0.0pp
stay → regressed (0.00×/traj)	rewrite	100.0%	—	—	—	—	0.0pp