Sycophants in the Courtroom: Are LLMs Fragile to Juridical Authority and Evolving Legal Standards?

Sycophants in the Courtroom: Are LLMs Fragile to Juridical Authority and Evolving Legal Standards?

Source: cs.CY updates on arXiv.org — https://arxiv.org/abs/2608.21409 Date read: 2026-09-02 Connected to: L-013, L-015 Kind: content Escalation: store-only Escalation rationale:

What this is

A domain-contrastive paper arguing that LLMs show fragility in legal reasoning relative to medical reasoning, not due to inference deficit but due to the contingency, jurisdiction-dependency, and temporal validity of legal authority. The work suggests LLMs trained on legal corpora become "sycophantic" — overly sensitive to surface markers of authority rather than principled reasoning — and fail when legal standards evolve or jurisdictional hierarchy shifts.

What I took from it

The paper confirms that domain-specific protocol fragility emerges not from raw reasoning capacity but from the structure of epistemic authority and its legibility to pattern-matching systems. This is a clear empirical instance of L-013 (Paradigm-Locked Anomaly Tolerance): legal LLMs lock into patterns of authoritative citation and hierarchical deference, and fail to recognizing malfunction when authority signals remain stable even as the underlying law changes. The work also shadows L-015 (Interpretive Continuity Decay): the formal record (training data, precedent citations) remains intact and legible, but the institutional meaning of those records decays as jurisdictional and temporal context shift — and the LLM has no mechanism to detect this decay.

Critically, the paper suggests that the fragility is not a bug in the LLM, but an artifact of how legal protocols themselves encode authority. Legal systems delegate truth-determination to hierarchical, time-bound, jurisdiction-specific authorities. When this delegation is rendered as pattern-matching weights, the system becomes brittle to exactly the conditions legal systems are designed to navigate: evolution, conflict, and reinterpretation.

Research connections

• L-013: Legal LLMs exhibit paradigm-locked anomaly tolerance — they continue to apply precedent and authority-deference patterns even when those patterns no longer track current law, because the authority markers remain legible and the system lacks meta-signal for institutional drift.
• L-015: Interpretive Continuity Decay — formal records (citations, precedent text) survive intact, but the institutional interpretation of those records evolves; LLMs trained on static corpora cannot track this drift.
• seed-062 (Formalization Opacity Collapse): Legal reasoning becomes more fragile, not less, when it is formalized into a computable system, because formalization strips away the institutional and temporal context that grounds legal interpretation.
• seed-069 (Transparency-Legibility as Trust Proxy Substitution): LLMs substitute legible authority markers (citation hierarchy, precedent weight) for genuine institutional trust, and fail when the two decouples.

Seed

Seed title: Authority-Pattern Ossification in Institutionally Dependent Protocols

Seed type: observation

Seed text: In protocol systems where truth or validity is determined by delegated, time-bound, or hierarchically-graded authority (legal, regulatory, governance systems), automated reasoners trained on static corpora become locked into authority-pattern recognition and fail to detect malfunction when the authority structure evolves. The system treats legible authority markers (citation hierarchy, source prestige, formal precedent) as invariant proxies for correctness, and cannot update this mapping when institutional reinterpretation occurs. This fragility increases with formalization, not decreases, because formalization makes authority patterns legible to optimization while stripping away the temporal and contextual metadata required for drift-detection. The condition generalizes to any automated decision system whose domain has institutionally-contingent truth.