evaluating

Evaluation

Purpose

Analyze screened patents by decomposing claims into elements and storing analysis data in the database for further processing.

Prerequisites

• patents.db must exist with screened_patents table populated (from screening skill)
• Load investigation-fetching skill for data retrieval operations
• Load investigation-recording skill for elements recording

Constitution

Core Principles

Element-by-Element Analysis (The Golden Rule):

• Every claim analysis MUST test the target invention against the reference patent element by element
• Break down inventions into Elements A, B, C
• Find references disclosing A AND B AND C for anticipation (Novelty)
• Do not rely on "general similarity"

Skill-Only Database Access:

• Use investigation-recording skill for elements recording (LLM interpretation task)
• For claims recording, use sqlite3 JSON functions directly with output_file — do NOT pass claim text through LLM generation (see Step 3)

Skill Orchestration

Execute Evaluation

Do NOT delegate to subagents (Agent tool) — invoke Skills directly from this session.

Process:

1. Get Patents to Analyze:

   • Invoke Skill: investigation-fetching with request "Get list of relevant patents without evaluation"

2. Batch Fetch Patent Data (up to 10 patents in parallel):

   • Split patents into batches of 10
   • For each batch, invoke Skill: google-patent-cli:patent-fetch for all patents in parallel

3. Record Claims (for each patent — mechanical, no LLM text generation):

   • After fetch_patent returns the output_file, use sqlite3 JSON functions to INSERT directly. Do NOT read claim text and regenerate it — LLM will summarize/compress long repetitive structures.

     sqlite3 patents.db "
     INSERT OR REPLACE INTO claims (patent_id, claim_number, claim_type, claim_text, created_at, updated_at)
     SELECT
       '<patent_id>',
       CAST(json_extract(value, '$.number') AS INTEGER),
       CASE
         WHEN CAST(json_extract(value, '$.number') AS INTEGER) = 1 THEN 'independent'
         ELSE 'dependent'
       END,
       json_extract(value, '$.text'),
       datetime('now'),
       datetime('now')
     FROM json_each(json_extract(CAST(readfile('<output_file>') AS TEXT), '$.claims'));
     "
     

   • After INSERT, verify with: sqlite3 patents.db "SELECT COUNT(*) FROM claims WHERE patent_id = '<patent_id>'"
   • Then UPDATE claim_type for each independent claim identified by reading claims from the DB:

     sqlite3 patents.db "SELECT claim_number, substr(claim_text, 1, 80) FROM claims WHERE patent_id = '<patent_id>'"
     

     Identify independent claims (those NOT starting with "前記", "The ... of claim", "請求項", etc.) and UPDATE:

     sqlite3 patents.db "UPDATE claims SET claim_type = 'independent', updated_at = datetime('now') WHERE patent_id = '<patent_id>' AND claim_number IN (<independent_numbers>)"
     

4. Analyze and Record Elements (for each patent — LLM interpretation task):

   • Read claims from the DB: sqlite3 patents.db "SELECT claim_number, claim_text FROM claims WHERE patent_id = '<patent_id>'"
   • For EACH claim, decompose into constituent elements (A, B, C...)
   • Invoke Skill: investigation-recording with request "Record elements for patent : <elements_data>"

5. Verify Results: Confirm all claims and elements are recorded in the database

State Management

Initial State

• Patents in screened_patents table marked as relevant without corresponding claims/elements entries exist

Final State

• No patents in screened_patents marked as relevant without corresponding claims/elements entries (all evaluated)