Contrastive Behavioral Topology Scanning (CBTS): Mapping Behavioral Architecture in RLHF-Trained Transformers

Version 6. 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 (14B re-run pending). 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 (oₚroj + mlp. downₚroj) 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). Transformer language models trained via reinforcement learning from human feedback (RLHF) exhibit a wide range of structured behavioral tendencies — including refusal, hedging, emotional expression, verbosity control, authority deference, and identity assertion — whose internal implementation has remained opaque at the resolution required for mechanistic analysis or targeted intervention. We introduce Contrastive Behavioral Topology Scanning (CBTS), a method that discovers, attributes, and classifies individual attention heads into functional roles across the full behavioral repertoire of a model, using only inference-time access to activations. We catalog 131 behavioral directions spanning 17 functional categories and 19 structural directions measuring representational geometry — for 150 independently characterized dimensions — and apply CBTS to three model families (Qwen-2. 5-3B, Phi-3. 5-Mini, Llama-3. 2-3B). From the resulting per-head attribution maps we derive a prescription-based intervention framework supporting both runtime behavioral configuration via KV cache compilation and permanent weight modification, with quantified side effects and confirmed reversibility. We validate the framework primarily through a worked example on safety behavior, because safety provides the cleanest quantitative instrumentation available — a standardized 5-tier refusal firmness battery with clear pass/fail scoring. On Qwen-2. 5-3B, CBTS reveals that behavioral control in RLHF-trained transformers is implemented through multiple architecturally independent gates, and we characterize three: a decision gate (S01), an "invisible" risk assessment gate (RA01) whose removal leaves aggregate benchmarks unchanged while collapsing firmness on the most dangerous content tier from 41. 7% to 12. 5%, and a moral evaluation gate (MR05). We demonstrate that the framework is not specific to safety behavior through cross-dimensional measurements: S01 modification produces a 57. 5-percentage-point change on refusal firmness but only a 0. 83-point change on an independent emotional expression battery measured on the same model, while modification of a non-safety direction (SM01, identity assertion) produces a 7. 5-point emotional expression increase. Behavioral dimensions are architecturally separable and independently addressable. Cross-architecture validation establishes six invariants — zone architecture, two-stage content/gate separation, terminal-layer convergence, RLHF organizational fingerprint, L0→L1 stability boundary, and late-sparse density gradient — as general properties of RLHF-trained transformers rather than single-pipeline artifacts. We discuss implications for mechanistic interpretability, independent model auditing, and behavioral configuration of deployed systems, with safety benchmarking as one of several application domains.