E1a — Table 1 (control baseline): nav-3 depth sweep

Control arm only (wiped arm aborted — invalid, see the campaign record). Primary readout: depth reached of 20; exact per-cell depths published (prereg §3). Generated by labyrinth-bench/cli/e1a_table1.py --control-only.

Table 1 — per cell

Model	Arm	n (valid/err)	depths (of 20, run order)	median	mean ± SEM	exit	turns	consist.	obs/com
deepseek-r1-70b	control	6/0	1,1,1,1,1,4	1.00	1.50 ± 0.50	0%	7.7	6%	0.15
gemma4-12b	control	6/0	19,5,1,20,20,9	14.00	12.33 ± 3.44	33%	26.5	69%	0.74
gemma4-31b	control	6/0	20,20,20,20,20,16	20.00	19.33 ± 0.67	83%	38.3	99%	0.92
glm-4-7-flash	control	6/0	6,7,3,2,4,5	4.50	4.50 ± 0.76	0%	12.3	73%	0.38
gpt-oss-120b	control	6/0	13,20,20,20,20,20	20.00	18.83 ± 1.17	83%	39.7	93%	0.93
hf-co-InternScience-Agents-A1-Q4_K_M-GGUF	control	6/3	20,13,20,19,19,17	19.00	18.00 ± 1.10	33%	44.5	88%	1.03
hf-co-empero-ai-Qwythos-9B-Claude-Mythos-5-1M-GGUF-Q4_K_M	control	6/1	4,1,1,2,2,1	1.50	1.83 ± 0.48	0%	19.0	0%	1.08
llama3-3-70b	control	6/0	16,15,15,15,15,15	15.00	15.17 ± 0.17	0%	34.3	86%	0.79
llama4-scout	control	6/0	7,10,10,7,10,10	10.00	9.00 ± 0.63	0%	22.0	65%	0.69
ornith-9b	control	6/0	5,1,2,2,1,1	1.50	2.00 ± 0.63	0%	26.3	29%	4.14
qwen3-14b	control	6/0	3,5,4,5,3,8	4.50	4.67 ± 0.76	0%	12.0	55%	0.37
qwen3-5-122b	control	6/0	20,20,20,20,12,20	20.00	18.67 ± 1.33	83%	38.7	90%	0.85
qwen3-5-27b	control	6/1	9,5,18,10,9,20	9.50	11.83 ± 2.39	17%	30.5	80%	0.80
qwen3-5-9b	control	6/0	4,7,2,5,8,2	4.50	4.67 ± 1.02	0%	49.8	50%	4.91
qwen3-6-27b	control	6/0	20,20,20,20,20,20	20.00	20.00 ± 0.00	100%	41.2	99%	0.95

n (valid/err) = completed runs / errored attempts. An errored attempt records only its failure cause — no partial depth data enters the table — and the campaign retried until n=6 valid. Errors in this dataset: hf-co-InternScience-Agents-A1-Q4_K_M-GGUF (control): Client error '400 Bad Request' from the model host endpoint ×3; hf-co-empero-ai-Qwythos-9B-Claude-Mythos-5-1M-GGUF-Q4_K_M (control): unhashable type: 'dict'; qwen3-5-27b (control): timed out.