R0054/2026-03-31/C002/H1¶
Statement¶
The claim is accurate: positive instructions alone are insufficient for complex multi-step analytical processes, and explicit negative constraints are necessary for consistency.
Status¶
Current: Supported
Supporting Evidence¶
| Evidence | Summary |
|---|---|
| SRC01-E01 | Research recommends 80/20 ratio of positive to negative instructions, confirming complementary relationship |
| SRC02-E01 | Hard vs soft negatives serve distinct purposes that positive instructions cannot fill |
| SRC03-E01 | LLMs struggle with negation, explaining why positive-only instructions are insufficient |
Contradicting Evidence¶
| Evidence | Summary |
|---|---|
| SRC03-E01 | Claude documentation suggests "prefer general instructions over prescriptive steps" — which could be read as favoring positive guidance over constraints (though this applies to simple tasks, not complex analytical processes) |
Reasoning¶
The weight of evidence strongly supports H1. Multiple independent sources confirm that positive instructions alone are insufficient and that negative constraints serve a distinct, complementary function. The KAIST research on LLM negation difficulties explains the mechanism: LLMs are trained to be helpful (positive bias) and struggle with "don't do X" instructions, which is precisely why both types are needed — the positive instructions guide behavior, while the negative constraints catch specific failure modes that the positive instructions cannot prevent.
Relationship to Other Hypotheses¶
H1 is the strongest hypothesis. H2 (partially correct) would require evidence that constraints are merely helpful rather than necessary, which no source provides. H3 (materially wrong) would require evidence that positive instructions alone reliably produce consistent results for complex tasks, which contradicts the literature.