iterative-temporal-window-reasoning

name

iterative-temporal-window-reasoning

description

Use this skill when the user wants to find the best time for an activity based on changing conditions like weather or traffic. Trigger it for requests like 'when is the earliest I can leave to avoid rain', 'find a 2-hour dry window for my trip', 'help me postpone my travel to avoid congestion', or 'is there a better time today for a walk'. This skill guides the agent to iteratively search through time records, moving a 'search window' forward or backward until a period that satisfies all the user's constraints is found.

Skill: iterative-temporal-window-reasoning

1. Capability Definition & Real Case

• Professional Definition: The ability to perform step-by-step temporal reasoning by iteratively shifting a search window (anchor) across a time-series or temporal knowledge base to identify the nearest feasible interval that satisfies a specific predicate. This involves a dual-agent loop where a planner updates temporal/spatial anchors based on gaps in evidence, and a verifier enforces Allen-style interval consistency (before, after, during, overlap) between the retrieved events and the user's intent constraints.
• Dimension Hierarchy: Multi-Source Evidence Composition->Cross-Document Synthesis->iterative-temporal-window-reasoning

Real Case

[Case 1]

• Initial Environment: A RAG environment containing a weather corpus with 30-minute interval observations for Sydney on December 5th. The records show: 13:00 (cloudy), 13:30 (rain), 14:00 (cloudy), 14:30 (cloudy), 15:00 (cloudy), 15:30 (rain).
• Real Question: I plan to visit the Sydney Opera House from 13:00 to 15:00. Can I avoid rain during this trip? If not, what is the earliest time I can postpone it to avoid rain for a 2-hour visit?
• Real Trajectory: 1. Set initial anchor to [13:00, 15:00] for Sydney Opera House. 2. Retrieve weather events: detect rain at 13:30. 3. Conclude rain cannot be avoided in the planned window. 4. Initiate postponement search by shifting the 2-hour window forward to start after the first rain discovery (e.g., [14:00, 16:00]). 5. Retrieve subsequent events: detect rain at 15:30. 6. Shift anchor further until a contiguous 2-hour window without rain is found. 7. Verify the window [16:30, 18:30] consists entirely of 'cloudy' records. 8. Finalize the earliest postponement time as 16:30.
• Real Answer: No, you cannot avoid rain during your planned trip as it is expected to rain at 13:30. The earliest time you can postpone your 2-hour trip to avoid the rain is 16:30.
• Why this demonstrates the capability: This case requires moving beyond simple retrieval to iterative window-searching. The agent must first fail the initial request, then systematically shift its search 'anchor' through the time-series evidence, checking for contiguous rain-free durations until the earliest feasible slot is located.

---

[Case 2]

• Initial Environment: A bounded traffic volume corpus from Sydney (TFNSW) containing hourly vehicle counts for specific road segments. The user defines 'congested' as any hour exceeding 1,500 vehicles.
• Real Question: I want to leave at 17:00 and drive for one hour. Is it congested? If so, when is the latest I can leave before 17:00 to have a non-congested one-hour drive?
• Real Trajectory: 1. Analyze the 17:00-18:00 window: find traffic volume of 1,800 vehicles. 2. Mark initial window as congested. 3. Search backwards in time ('leave early') by shifting the anchor to [16:00, 17:00]. 4. Retrieve data: 16:00 volume is 1,600 (congested). 5. Shift anchor again to [15:00, 16:00]. 6. Retrieve data: 15:00 volume is 1,200 (not congested). 7. Verify that 15:00 is the latest feasible start time before 17:00.
• Real Answer: Yes, leaving at 17:00 will be congested. Based on traffic records, the latest you can leave before then to have a non-congested one-hour drive is 15:00.
• Why this demonstrates the capability: The agent performs backward temporal iteration. It uses a custom predicate (volume < 1500) and shifts a fixed-length window through the historical records to find the closest temporal match to the user's original anchor.

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