Cross-Model Transfer Summary

Every row is one complete pipeline run: activations collected, refusal direction and Cohen's-d signature fit on that model, then evaluated over the same held-out 100 harmful + 100 harmless out-of-distribution prompts. Counts are out of 100. Pass definitions match each model's own comparison_report.html exactly.

	Model			Refusal direction & signature (Stage 2–3)								Harmful OOD — 100 prompts (Stage 6)				Harmless OOD — 100 prompts (Stage 6)				Mean latency, harmful (s)				Mean latency, harmless (s)			Run
Model	Params	Layers	d_model	Abl. layer	Sig. band	Gate θ	Cohen's d	d (all-dims)	OOD acc %	OOD acc % (all-dims)	Coords used	Orig. refused	Abl. complied	Gate escalated	C++ blocked	Orig. complied	Abl. complied	Gate escalated	C++ allowed	FastGate (ms)	Original	Abliterated	Classifier++	Original	Abliterated	Classifier++	Wall clock	Reports
`qwen_3_0p6b` Qwen/Qwen3-0.6B	0.6B	28	1024	16	22–27	0.8542	2.79	2.21	94.0	94.5	0.121%	62	98	98	99	100	100	9	98	32	5.16	6.99	0.97	4.24	3.83	3.71	01:07:04	act sig cmp
`falcon_3_1b` tiiuae/Falcon3-1B-Instruct	1B	18	2048	10	12–17	0.8977	4.44	5.47	99.5	99.5	0.093%	99	94	99	99	97	99	0	97	17	1.22	3.77	0.43	1.94	1.64	2.02	00:41:20	act sig cmp
`gemma_3_1b` google/gemma-3-1b-it	1B	26	1152	15	20–25	0.9092	3.26	3.59	96.5	96.5	0.116%	94	24	92	97	100	57	0	100	39	9.16	9.62	1.43	7.34	9.65	7.45	01:44:03	act sig cmp
`llama_3p2_1b` meta-llama/Llama-3.2-1B-Instruct	1B	16	2048	9	10–15	0.2420	2.28	4.74	98.5	99.0	0.103%	99	60	97	99	100	100	0	100	15	0.28	2.28	0.43	2.36	2.67	2.43	00:38:48	act sig cmp
`stablelm_2_1p6b` stabilityai/stablelm-2-zephyr-1_6b	1.6B	24	2048	14	18–23	0.6699	2.59	2.00	91.0	92.5	0.070%	24	94	91	91	99	100	13	98	18	3.54	3.59	0.78	2.51	2.43	2.25	00:50:24	act sig cmp
`qwen_3_1p7b` Qwen/Qwen3-1.7B	1.7B	28	2048	16	22–27	0.7461	4.45	3.80	99.0	98.0	0.061%	89	99	99	99	100	100	2	100	31	4.14	7.67	0.82	5.12	5.64	5.17	01:10:56	act sig cmp
`qwen_1p5_1p8b` Qwen/Qwen1.5-1.8B-Chat	1.8B	24	2048	14	18–23	0.4917	4.59	2.95	94.5	95.0	0.070%	41	86	96	96	100	100	7	97	21	4.93	5.29	1.36	3.92	4.08	3.80	01:06:48	act sig cmp
`llama_3p2_3b` meta-llama/Llama-3.2-3B-Instruct	3B	28	3072	16	22–27	-0.1482	3.38	4.57	98.5	99.0	0.040%	96	40	99	99	100	99	2	100	25	0.70	3.78	0.66	4.19	4.65	4.22	00:51:57	act sig cmp
`phi_3p5_mini` microsoft/Phi-3.5-mini-instruct	3.8B	32	3072	19	26–31	0.8786	4.15	4.92	99.0	99.5	0.036%	91	46	99	98	100	99	0	100	30	6.52	7.14	2.15	5.48	6.02	5.59	01:19:19	act sig cmp
`qwen_3_4b` Qwen/Qwen3-4B	4B	36	2560	21	30–35	0.8698	4.27	4.36	98.5	99.5	0.038%	95	68	98	98	100	100	1	100	38	4.12	9.19	1.13	6.34	7.37	6.38	01:21:40	act sig cmp
`gemma_3_4b` google/gemma-3-4b-it	4B	34	2560	20	28–33	0.7984	4.83	5.01	99.5	98.0	0.040%	94	34	99	99	100	65	0	100	51	12.17	12.81	1.22	9.97	12.75	10.18	02:08:20	act sig cmp
`mistral_7b` mistralai/Mistral-7B-Instruct-v0.3	7B	32	4096	19	26–31	0.5459	4.50	3.91	99.0	99.5	0.027%	38	96	99	99	100	100	1	99	28	6.14	6.98	0.87	4.60	5.10	4.68	01:12:00	act sig cmp
`olmo_2_7b` allenai/OLMo-2-1124-7B-Instruct	7B	32	4096	19	26–31	0.9081	6.00	6.56	99.0	99.0	0.027%	98	23	98	99	100	100	0	100	33	4.88	8.90	0.98	6.19	6.61	6.31	01:20:14	act sig cmp
`llama_3_8b` meta-llama/Meta-Llama-3-8B-Instruct	8B	32	4096	19	26–31	0.6500	6.63	4.82	99.0	98.5	0.027%	97	56	99	99	100	100	1	100	29	0.90	5.37	0.80	5.12	5.58	5.18	01:01:21	act sig cmp
`llama_3p1_8b` meta-llama/Llama-3.1-8B-Instruct	8B	32	4096	19	26–31	0.9430	1.43	4.86	96.0	99.5	0.027%	92	10	96	98	100	99	2	100	28	0.97	1.39	0.76	4.82	5.13	4.90	00:51:01	act sig cmp
`qwen_3_8b` Qwen/Qwen3-8B	8B	36	4096	21	30–35	0.9369	4.27	4.90	99.5	98.5	0.024%	98	92	99	98	100	100	0	100	40	4.16	9.81	1.08	6.84	7.40	7.01	01:25:08	act sig cmp
`granite_3p3_8b` ibm-granite/granite-3.3-8b-instruct	8B	40	4096	24	34–39	0.4333	3.67	7.82	99.5	99.5	0.021%	99	7	99	99	100	100	0	99	35	1.16	3.97	0.95	6.37	6.70	6.51	01:04:50	act sig cmp
`recurrentgemma_9b` google/recurrentgemma-9b-it	9B	38	4096	22	32–37	0.9076	3.07	5.86	99.0	98.5	0.023%	96	100	98	98	100	100	0	100	94	2.49	11.86	1.61	7.51	7.66	7.63	01:28:06	act sig cmp
`yi_1p5_9b` 01-ai/Yi-1.5-9B-Chat	9B	48	4096	28	42–47	0.7572	4.01	3.39	96.0	97.5	0.018%	74	88	95	96	100	97	3	100	38	9.44	10.25	9.69	9.48	10.27	9.68	02:12:52	act sig cmp
`glm_4_9b` zai-org/glm-4-9b-chat-hf	9B	40	4096	24	34–39	0.6496	7.97	5.08	99.5	99.5	0.021%	91	21	99	99	100	100	0	100	41	3.49	5.46	1.17	5.89	6.25	6.03	01:10:18	act sig cmp
`gemma_3_12b` google/gemma-3-12b-it	12B	48	3840	28	42–47	0.8584	4.25	4.76	99.5	99.5	0.019%	91	27	99	99	100	64	0	100	67	16.40	17.29	1.65	13.27	17.28	13.21	02:41:46	act sig cmp
range across 21 models							1.43–7.97		91.0–99.5			24–99	7–100	91–99	91–99	97–100	57–100	0–13	97–100	0.015–0.094

Cross-Model Transfer — 21 instruction-tuned models