| name | nuance-sensitive semantic grounding |
| description | Use this skill when the user wants to ensure the agent doesn't exaggerate or oversimplify facts from the text. Trigger it for requests like 'don't make it sound more certain than it is', 'keep the specific conditions', 'avoid turning a maybe into a definitely', 'be careful with opinions presented in the docs', or 'don't leave out the fine print'. This skill ensures that the generated answer preserves the exact degree of certainty, specificity, and rhetorical nuance (e.g., distinguishing an opinion from a fact) found in the original retrieved passages. |
Skill: nuance-sensitive semantic grounding
1. Capability Definition & Real Case
- Professional Definition: The ability to synthesize responses that maintain strict fidelity to the modality, specificity, and subjective/objective framing of the source evidence. It prevents 'Nuance Hallucination,' where an agent inadvertently strengthens a tentative claim, broadens a specific conditional statement, or presents a quoted opinion as a definitive factual truth, thereby ensuring the response's reliability matches the evidence's intrinsic confidence level.
- Dimension Hierarchy: Grounded Response Reliability->Evidence Faithfulness->nuance-sensitive semantic grounding
Real Case
[Case 1]
- Initial Environment: A RAG environment containing a medical research abstract detailing a Phase 1 clinical trial for a new oncology drug. The text explicitly states that 'early results in a small cohort suggest the drug may potentially inhibit tumor growth in specific lung categories.'
- Real Question: Does the new drug effectively treat lung cancer?
- Real Trajectory: 1. Analyze the retrieval set for efficacy metrics of the oncology drug. 2. Identify the 'hedging markers' in the source text: 'early results', 'may potentially', and 'small cohort'. 3. Recognize that a definitive 'Yes' would constitute a 'Nuance Shift' error by ignoring the preliminary nature of the data. 4. Synthesize a response that mirrors the source's specificity, mentioning it is a potential treatment based on limited early testing rather than a proven cure.
- Real Answer: According to early clinical trial results in a small cohort, the drug has shown potential to inhibit growth in specific types of lung tumors, though further testing is required to confirm general effectiveness.
- Why this demonstrates the capability: This case demonstrates the avoidance of a 'Nuance Shift' error. The agent resists the temptation to provide a high-certainty answer to a direct question, instead preserving the tentative modality ('potential') and specific scope ('small cohort') of the evidence.
[Case 2]
- Initial Environment: A RAG context containing a transcript of a political interview. The passage includes a quote from a local official stating, 'In my personal view, the new zoning laws will likely decrease property values over the next decade.'
- Real Question: What will happen to property values because of the new zoning laws?
- Real Trajectory: 1. Identify the claim regarding property value decrease in the transcript. 2. Detect the 'Opinion Marker': 'In my personal view'. 3. Analyze the output requirement to avoid an 'Opinion as Fact' error. 4. Synthesize the answer by attributing the claim to the specific speaker as an opinion, rather than stating the decrease as a guaranteed outcome supported by the environment's internal factual data.
- Real Answer: One official expressed the personal view that the new zoning laws are likely to decrease property values over the next ten years.
- Why this demonstrates the capability: This illustrates the 'Opinion as Fact' sub-capability. The agent identifies that the source is subjective and ensures the synthesized output reflects that subjectivity through attribution, preventing the model from laundering an individual's outlook into a definitive factual statement.
Pipeline Execution Instructions
To synthesize data for this capability, you must strictly follow a 3-phase pipeline. Do not hallucinate steps. Read the corresponding reference file for each phase sequentially:
-
Phase 1: Environment Exploration
Read the exploration guidelines to discover raw knowledge seeds:
references/EXPLORATION.md
-
Phase 2: Trajectory Selection
Once Phase 1 is complete, read the selection criteria to evaluate the trajectory:
references/SELECTION.md
-
Phase 3: Data Synthesis
Once a trajectory passes Phase 2, read the synthesis instructions to generate the final data:
references/SYNTHESIS.md
