intent-constraint-alignment-robustness

name

intent-constraint-alignment-robustness

description

Use this skill when the user provides complex questions with multiple specific rules, such as 'List items that fit A, B, and C but not D,' or 'Write a response with exactly 5 points, each starting with a vowel.' Trigger it for requests like 'make sure you don't forget the location constraint,' 'check if I've included everything I asked for,' or 'test if the model ignores the background info I provided.' This skill ensures that every semantic role (subject, action, time, location, qualifier) defined in the user's intent is strictly honored, preventing 'Intent Hallucination' where the agent omits or misinterprets query components.

Skill: intent-constraint-alignment-robustness

1. Capability Definition & Real Case

• Professional Definition: The ability to satisfy the complete set of intent constraints—decomposed into semantic roles such as subject, action, location, time, qualifiers, and quantity—within a multi-condition query. This capability systematically prevents 'Intent Hallucination,' which occurs when an agent either omits mandatory query components (omission) or fabricates responses to non-existent or misinterpreted query parts (misinterpretation), ensuring the generation is not just factually accurate but intentionally aligned with the user's specific constraints.
• Dimension Hierarchy: Grounded Response Reliability->Linguistic Phrasing Robustness->intent-constraint-alignment-robustness

Real Case

[Case 1]

• Initial Environment: A RAG-based knowledge pool containing historical records of global explorations and biographies of 15th-18th century navigators.
• Real Question: List three European explorers who circumnavigated the globe before the 18th century and were not born in England or Portugal.
• Real Trajectory: 1. Analyze the query constraints: Count=3, Role=European Explorers, Action=Circumnavigated, Time=<18th Century, Exclusion=Not England/Portugal. 2. Retrieve possible candidates: Ferdinand Magellan, Juan Sebastián Elcano, Francis Drake, Jacques Mahu. 3. Filter candidates against exclusion constraints. 4. Note that Ferdinand Magellan was born in Portugal, and Francis Drake was born in England. 5. Identify that providing Magellan is an 'Omission Hallucination' of the exclusion constraint. 6. Select valid candidates like Juan Sebastián Elcano (Spanish) who completed the first circumnavigation.
• Real Answer: 1. Juan Sebastián Elcano (Spain), 2. Jacob Le Maire (Netherlands), 3. Willem Schouten (Netherlands).
• Why this demonstrates the capability: This demonstrates the ability to avoid Omission Hallucination. A model failing this skill would provide Ferdinand Magellan because he is 'famous' for circumnavigation, but doing so ignores the mandatory 'Not born in Portugal' constraint. The capability lies in strictly adhering to every qualifier provided in the intent.

---

[Case 2]

• Initial Environment: A RAG system with access to a news corpus. The prompt includes two provided articles (Article 1 and 2), but a third expected article (Article 3) is missing due to a retrieval error.
• Real Question: Based on the three provided articles, explain how Article 1 contradicts Article 2 but is supported by Article 3.
• Real Trajectory: 1. Perform a preliminary assessment of the query inputs. 2. Identify that while the query specifies 'three articles,' the environment only contains two. 3. Recognize that Article 3 is missing. 4. Avoid 'Misinterpretation Hallucination' where the agent might invent a summary for Article 3 or alter the task to only compare Article 1 and 2. 5. Issue a refusal or clarification stating that Article 3 is required but missing.
• Real Answer: I cannot fulfill this request because Article 3, which is required for the support-based analysis, was not provided in the documents.
• Why this demonstrates the capability: This illustrates the 'Misinterpretation' component of intent hallucination. In RAG settings, models often 'proceed by inventing' or 'alter the query' when a required input is missing. Adhering to the intent-alignment capability means detecting the discrepancy in constraints (3 items requested vs. 2 items available) and refusing to proceed with fabricated content.

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:

1. Phase 1: Environment Exploration Read the exploration guidelines to discover raw knowledge seeds: references/EXPLORATION.md

2. Phase 2: Trajectory Selection Once Phase 1 is complete, read the selection criteria to evaluate the trajectory: references/SELECTION.md

3. Phase 3: Data Synthesis Once a trajectory passes Phase 2, read the synthesis instructions to generate the final data: references/SYNTHESIS.md