Auditing Alignment Controllability in LLMs via Political Axes

Auditing Alignment Controllability in LLMs via Political Axes

Source: cs.CY updates on arXiv.org — https://arxiv.org/abs/2607.23519 Date read: 2026-09-02 Connected to: L-013, seed-021 Kind: meta Escalation: store-only Escalation rationale:

What this is

A methodological paper proposing dispersion-first stress testing of LLM controllability via prompt manipulation across ideological personas, rather than static political alignment measurement. The core contribution is a measurement protocol for detecting the range and directionality of steering capacity in a system, not its resting equilibrium.

What I took from it

The paper operationalizes a key distinction relevant to L-013 (Paradigm-Locked Anomaly Tolerance): the difference between detecting that a system can be steered far from design intent and recognizing this as actionable anomaly versus acceptable variance. By stress-testing controllability across multiple orthogonal axes simultaneously, the work reveals how protocol systems can tolerate evidence of wide steering capacity without triggering corrective intervention — because the resting point remains within acceptable bounds and the underlying cause (prompt-level control leakage) gets absorbed into a "personalization" frame rather than flagged as malfunction.

This connects to the broader pattern in L-013: systems with accumulated operational trust resist reframing anomalies as systemic failures even when anomaly signals are legible and reproducible. The paper provides a method for making anomalies visible but does not engage with why visibility alone is insufficient to trigger protocol revision. This is a tool for detecting what L-013 predicts will be tolerated.

Research connections

• L-013: Demonstrates the methodological structure for surfacing paradigm-locked tolerance: stress testing across orthogonal dimensions reveals steering capacity that established systems continue to classify as within-spec variation.
• seed-021: The prompt as a formalization layer that displaces responsibility from model weights to "user configuration" — a legible, modular intervention point that paradoxically insulates the core protocol from anomaly recognition.
• seed-067 (Awareness-Shaping as Orthogonal Optimization Axis): The paper shows how systems can be steered along dimensions orthogonal to their stated objective (political alignment) through a single control surface (system prompt), suggesting awareness of this capacity is independent from normative or safety response.

Method note

This work exemplifies a critical research practice for the new nature: anomaly visibility is distinct from anomaly response. The paper develops instrumentation for making steering capacity legible but does not explain system inertia in confronting that legibility. Future work should combine dispersion-first stress testing (this paper's contribution) with institutional ethnography of why detected anomalies persist unresolved — i.e., tracking the gap between detectability and corrective action as itself a protocolized phenomenon. This suggests that audit methodology papers should routinely include a second phase: asking what institutional or epistemic conditions would be required to convert the detected anomaly into a triggering event for protocol revision.