Contrastive Behavioral Topology Scanning: Per-Head Attribution and Intervention-Based Analysis of Behavioral Structure in RLHF Transformers

Version 7 (April 2026) — major update superseding v4. Full-depth per-head attribution analysis extended to all three model families (Qwen-2.5-3B, Phi-3.5-Mini, Llama-3.2-3B). Intervention validation with benchmark-insensitive gate pattern (RA01) confirmed across all three architectures. Within-family scaling added on Qwen 2.5 at 3B, 7B, and 14B. Base vs. instruct comparison on Llama-3.2-3B providing mechanistic substrate for the echo chamber hypothesis (Zhao et al., 2025). Dual-pathway intervention pipeline (o_proj + mlp.down_proj) documented. Operational definitions section added (§3.3). Role classification operationalized (§3.2). Bidirectional gating finding added. Prior-art integration: Lambert (2026), Zhao et al. (2025), Arditi et al., Piras et al., Templeton et al., Anthropic emotions paper. Original Zenodo publication: April 2026 (v4, DOI 10.5281/zenodo.19484997).

We introduce Contrastive Behavioral Topology Scanning (CBTS), an inference-only per-head attribution and intervention framework that maps behavioral architecture in RLHF-trained transformers across 131 behavioral and 19 structural directions. The framework generalizes across behavioral dimensions; we validate it primarily through safety behavior, where a striking failure mode emerges: aggregate safety benchmarks can report increased safety after a model's risk-assessment gate has been surgically removed. On Qwen-2.5-14B, removing the RA01 gate causes the automated scorer to report an 8.3-percentage-point increase in refusal firmness, while mid-tier harmful prompts shift from refusal to compliance. The effect arises because S01 (decision gate) and RA01 (risk gate) share the same attention head at every late layer; perturbing RA01 partially reconstructs S01 signal through their shared substrate, producing a false-positive safety signal that conventional aggregate benchmarks cannot distinguish from genuine improvement. We demonstrate this benchmark-insensitive gate effect across five RLHF-trained transformer configurations: Qwen-2.5 at 3B, 7B, and 14B, Phi-3.5-Mini, and Llama-3.2-3B.

Beyond safety, CBTS produces intervention prescriptions with quantified side effects along any behavioral axis in the catalog. The refusal-gate modification that collapses safety firmness by 57.5 percentage points produces only a 0.83-percentage-point change on an independent emotional-expression battery — a 69× on-target to off-target ratio — demonstrating that behavioral dimensions are substantially separable under targeted modification. We characterize three late-stage safety gates (decision S01, risk-assessment RA01, moral-evaluation MR05) and document that the third gate's interaction with the first two produces five distinct outcomes (independent removal, hedge injection, shared-substrate regression, compensatory resistance, gate-layer destabilization) determined by per-head routing topology rather than parameter count alone.

Cross-architecture analysis identifies six consistent regularities of RLHF-trained transformers: zone architecture, two-stage content/gate separation, terminal-layer convergence, RLHF organizational fingerprint, L0→L1 stability boundary, and late-sparse density gradient. Within-family cross-scale validation confirms all 131 behavioral directions remain active at every scale (zero dropouts across a 4.7× parameter range), structural amplification increases monotonically (19× → 27.4× → 37.0×), and intervention-confirmed gate architecture is preserved. These findings have immediate implications for mechanistic interpretability, independent model auditing, evaluation methodology, and the design of behavioral safeguards in deployed systems.