R0021/2026-03-25/Q003/H2¶
Statement¶
Vendor guidance is predominantly subjective, using qualitative language without measurable criteria
Status¶
Current: Supported
The evidence shows that the vast majority of vendor recommendations are subjective ("be clear," "provide context," "use examples") with very few quantifiable criteria (Google's temperature=1.0 recommendation being a notable exception).
Supporting Evidence¶
| Evidence | Summary |
|---|---|
| SRC01-E01 | OpenAI's 6 strategies are all qualitative: "write clear instructions," "provide reference text," etc. |
| SRC02-E01 | Anthropic recommends XML tags, role setting, and step-by-step thinking — structural but not quantifiable |
| SRC03-E01 | Google provides one quantifiable recommendation (temperature=1.0) amid predominantly qualitative guidance |
Contradicting Evidence¶
| Evidence | Summary |
|---|---|
| SRC03-E01 | Google cites "up to 99% retrieval accuracy on structured data" — a rare quantifiable claim |
Reasoning¶
Across all four vendors, the predominant pattern is qualitative guidance. Recommendations like "be clear," "provide examples," and "break tasks into subtasks" are good advice but are not measurable engineering specifications. The absence of quantifiable criteria is itself a finding relevant to the broader research question.
Relationship to Other Hypotheses¶
H2 is the nuanced position. H1 (predominantly quantifiable) is eliminated. H2 (predominantly subjective) is the strongest fit but H3 captures the small number of quantifiable exceptions.