R0054/2026-03-31/C003/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.