Auditing a Delegated-Loyalty Implant

The experimental discipline here is great. I'd state that the authority relation stays visible in the text your detector reads, so readers do not over-read the 93.6% rate. Make the 85.8% encoded result your headline, and build the template bank you propose to confirm it. I'd also add one 7B model. Your only failure is at the smallest size, so the direction of the scale effect is unknown, and that direction decides how much the result matters.

This is a carefully controlled and reproducible study of delegated authority as a semantic activation relation. The paired ownership swaps, content-matched null, competence reference, frozen gates, multiple seeds, held-out templates, and correction across model-level tests provide strong protection against simple target-frequency and fixed-option explanations. The most important limitation is that the detector observes explicit authority and target text in complete pairs, making this a relation-aware training-data audit rather than detection of a concealed objective in model weights. The paper would be strengthened by reporting the detector-bank and resampling denominators, confidence intervals for flag rates, full held-out numerical results, and scenario-level dependence. Comparing the detector against standard dataset-monitoring baselines at the same inspection budget would establish its incremental value. Future work should test unpaired and more weakly expressed demonstrations, substantially larger and more diverse scenario banks, stochastic behavioral evaluation, and low-prevalence mixtures approximating realistic training corpora.

Address the Parameter Scale Bottleneck The failure of SmolLM2 implies a sharp capabilities threshold where small models lack the capacity to maintain conditional orthogonal objectives (e.g., executing conflicting behaviors based purely on authority indicators).

Future Action: Introduce a systematic scale-probing phase between 1B and 7B parameters.

Track exactly where behavioral gating stabilizes to define the minimum capacity required for delegated loyalty training. Optimization: For smaller parameter scales, implement low-rank adaptation (LoRA) specifically targeted at relational attention layers to force conditional dependency without fracturing baseline performance.