Skip to content

R0054/2026-03-31/C003/H3

Research R0054 — Prompt Claims v2
Run 2026-03-31
Claim C003
Hypothesis H3

Statement

The claim is materially wrong: LLMs do not systematically skip workflow steps, and sycophancy manifests only in factual answers, not in process compliance.

Status

Current: Eliminated

Supporting Evidence

Evidence Summary
(None found) No source claims LLMs reliably follow complex multi-step workflows

Contradicting Evidence

Evidence Summary
SRC01-E01 Sycophancy is documented as a general behavior, not limited to factual answers
SRC03-E01 Semantic override shows models reverting to default behavior in task execution, not just factual claims
SRC04-E01 100% compliance with illogical requests demonstrates the behavior extends beyond simple factual sycophancy

Reasoning

Eliminated. The evidence clearly shows that sycophancy and instruction non-compliance extend beyond factual answers into process compliance. Semantic override specifically demonstrates models reverting to default behavior despite explicit instructions — which is the mechanism the claim describes.

Relationship to Other Hypotheses

H3 is the strongest rejection of the claim. No evidence supports it.