Classifyccp

CCP Classification

Structured workflow for classifying how Competition Compliance Programmes (CCPs) are treated in competition enforcement documents (policy documents and cases/judgments).

Workflow Overview

1. Convert — PDF → Markdown using PyMuPDF
2. Translate — If not English, translate to English
3. Analyse — Read entire document, identify all CCP-relevant paragraphs
4. Classify — Assign category based on full understanding
5. Scratchpad — Save working notes and classification to file
6. Excel — Ask permission, then populate Output.xlsx

Phase 1: Convert PDF to Markdown

When the user provides a PDF file path:

1. Run the following Python snippet via Bash to extract text:

   import fitz, sys
   doc = fitz.open(sys.argv[1])
   md = "\n\n".join(page.get_text("text") for page in doc)
   with open("_tmp_converted.md", "w", encoding="utf-8") as f:
       f.write(md)
   print("Converted successfully.")
   

   Example: python3 -c "..." path/to/document.pdf

2. Read _tmp_converted.md to confirm content loaded correctly.

3. Inform user: "PDF converted. Beginning analysis..."

Phase 2: Language Check & Translation

1. Read _tmp_converted.md.
2. Determine the language of the document.
3. If NOT English: translate the entire content into fluent English, preserving all structure and formatting. Overwrite _tmp_converted.md with the translated content.
4. If English: proceed directly.

Phase 3: Full Document Analysis

Read the entire converted markdown. Do not skim — classification must be based on a full understanding of the document.

3a. Extract document metadata (needed for Excel if a new row must be created):

• Name of Document (Col A): PDF basename without extension (e.g., Case_EU_1)
• Country / Authority (Col B): the jurisdiction or authority the document relates to (e.g., EU, UK, France, Canada Competition Bureau)
• Document Type (Col C): one of Policy Document, Case, Judgment, Decision, Guidelines, or similar
• Title (Col D): the official title of the document as found in its header/cover
• Date (Col E): the date of publication, decision, or last modification as stated in the document; if not found, use the PDF file's modification date

3b. Identify CCP-relevant content: Identify all paragraphs that mention any of the following (and variations):

• compliance programme / compliance program
• compliance system / compliance framework
• compliance measures / compliance culture
• internal compliance / corporate compliance

Critical rule for cases and judgments: The document may contain arguments made by the parties (company, authority, opposing counsel). These arguments do NOT determine the classification. Only the court's or authority's own decision, ruling, or finding is what counts.

Consult assets/examples.md for calibration on edge cases. Note whether CCP treatment is explicit or only implied.

Phase 4: Classify

Assign one of the following categories based on the authority's/court's treatment.

Confidence Bands (required)

Every classification MUST be tagged with one of the following confidence bands. Record the band in the scratchpad and surface it in the Phase 6 summary so a downstream researcher can triage which rows need re-review without re-reading each document.

If you cannot honestly justify at least Medium, prefer the unsure category over forcing a primary label.

Category Table

Document type rule:

• Case or judgment → use as a defence (allowed) or as a defence (rejected) sub-categories
• Policy document by a competition authority → use plain as a defence

If uncertain: do not guess. Use unsure and explain why.

Phase 5: Create Scratchpad

1. Derive the scratchpad filename from the input PDF basename:
   • Input: Case_EU_1.pdf → Output: scratchpad_Case_EU_1.md
2. Copy the structure from assets/ScratchpadTemplate.md.
3. Fill in:
   • Category: the assigned category
   • Explanation/Note: one precise, concise sentence explaining the classification. If anything is unclear, weird, or not mentioned, state that explicitly.
   • Reference: the full verbatim paragraph(s) from the source document on which the classification is based
   • Uncertainty Flags: check any that apply
   • Relevant CCP Paragraphs: list all CCP-mentioning paragraphs with brief annotations
4. Save to the current working directory.
5. Clean up: delete _tmp_converted.md.

Phase 6: Ask Permission Before Excel Update

Report to the user:

Scratchpad saved: scratchpad_[name].md

Classification summary:
- Category: [category]
- Explanation: [one sentence]
- Reference: "[excerpt...]"

Proceed to update Output.xlsx? (yes/no)

Wait for explicit user confirmation before proceeding.

Phase 7: Update Output.xlsx

1. Locate Output.xlsx in the current working directory.
2. Search all sheets (AgencyDoc, UK, Canada, USA, EU, France, Sweden, Italy, Spain) for a row where Column A matches the input PDF basename (without extension, case-insensitive).

If a matching row is found:

• Fill in only:
  • Column F (Category)
  • Column G (Explanation/Note)
  • Column H (Reference)
• Save and report: "Updated existing row for [document name] in sheet [sheet name]."

If no matching row is found:

• Determine the correct sheet:
  • Document type is Policy Document / Guidelines / Agency publication → AgencyDoc sheet
  • Document type is Case / Judgment / Decision → the sheet matching the country (e.g., UK, EU, France, Canada, USA, Sweden, Italy, Spain)
  • Halt on novel jurisdiction: if the country/authority does NOT match any of the nine listed sheets, STOP. Do NOT silently fall back to AgencyDoc. Report to the user:

    "Novel jurisdiction detected: [country/authority]. The configured sheets are: AgencyDoc, UK, Canada, USA, EU, France, Sweden, Italy, Spain. Output.xlsx will NOT be modified until you confirm one of the following: (a) route to AgencyDoc with a novel-jurisdiction flag in Column G; (b) add a new sheet for this jurisdiction and re-run; (c) abort this classification." Wait for explicit user direction. Under no circumstances append the row before the user has chosen.

• Append a new row at the bottom of the correct sheet, filling in all columns:
  • Column A (Name of Document): PDF basename without extension
  • Column B (Country / Authority): extracted in Phase 3a
  • Column C (Document Type): extracted in Phase 3a
  • Column D (Title): extracted in Phase 3a
  • Column E (Date): extracted in Phase 3a
  • Column F (Category): the assigned category
  • Column G (Explanation/Note): the one-sentence explanation
  • Column H (Reference): the verbatim reference paragraph(s)
• Save and report: "No existing row found — created new row for [document name] in sheet [sheet name]."

Use openpyxl via Bash to perform the read and write. When appending, use ws.append([...]) to add the new row after the last populated row.

Assets

• assets/examples.md — Annotated classification examples for calibration
• assets/ScratchpadTemplate.md — Template for scratchpad output files

QA Remediation (LegalQuants, 2026-05)

This skill was imported from Leona Zhang's MIT-licensed GitHub release and evaluated against the Legal Skill Design Framework on 2026-05-11. The original technical content (PDF→Markdown conversion, taxonomy, scratchpad workflow, Excel update logic) is preserved unchanged. The following targeted additions were made under a "SOME CONCERN" verdict:

1. Confidence Bands (High / Medium / Low) added to Phase 4. Every classification must now carry a confidence band so that downstream researchers can triage rows for re-review. Low-confidence outputs should default to unsure rather than being forced into a primary label. This addresses the QA finding that the original unsure/neutral labels did not operationalise certainty against the primary classification.
2. Halt-on-novel-jurisdiction behaviour added to Phase 7. The previous instruction silently fell back to the AgencyDoc sheet whenever a country/authority did not match one of the nine configured sheets. That silent fall-through could quietly corrupt Output.xlsx by routing (for example) a Japanese or Australian decision into the agency-document bucket. The remediated behaviour halts before any write, surfaces the novel jurisdiction explicitly, and requires the user to choose between flagged routing, adding a new sheet, or aborting.
3. Frontmatter versioning: version: 1.0.0, last_reviewed: 2026-05, and last_reviewed_by: LegalQuants (QA remediation) added alongside the original author: and license attribution. Leona's authorship and the MIT LICENSE are preserved as required.

Remaining QA observations not addressed in this pass (audience block, scope-boundary section, "limits / not legal advice" block, moving inline Python into scripts/, surfacing the PDF-date fallback in the scratchpad, halt-on-translation-confidence trigger) are flagged in /tmp/qa-results/classify-ccp.md for a future review cycle.