Negative Monitor

• companies[].company_name

Optional:

• companies[].jurisdiction
• companies[].company_number
• companies[].aliases
• search_window_days
• language

Workflow Decision Tree

1. Layer 1 retrieval and evidence logging always runs first.
2. Layer 2 extraction and classification runs on logged evidence.
3. Never discard Layer 1 evidence, even when low relevance.
4. Only Layer 2 summaries are sent to warning channels.

Layer 1: Taxonomy-Guided Retrieval And Logging

Goal

Use the taxonomy.md as the retrieval backbone to maximize recall of public negative signals relevant to repayment risk.

For each company, Layer 1 should:

1. use the taxonomy as the search anchor,
2. expand each risk signal into slightly generalized public wording,
3. search across registry, legal, news, and company-announcement sources,
4. preserve all retrieved evidence with full traceability.

Do not search only the literal taxonomy labels.

Examples:

• liquidation → winding up, administration, insolvency notice
• new charge → registered charge, debenture, new facility
• shutdown → ceased trading, operations suspended, business closure

Retrieval Rule

For each company and each taxonomy category, generate 2–3 search variants:

• exact risk term
• generalized public wording
• source-targeted query where relevant

Examples:

• "<company> liquidation OR winding up"
• "<company> charge OR debenture OR refinancing"
• site:gov.uk "<company>" insolvency
• site:thegazette.co.uk "<company>" petition

Use jurisdiction-specific public sources when jurisdiction is available.

Examples:

• UK: Companies House, Gazette, gov.uk notices
• HK: Companies Registry, HKEX announcements, Gazette
• SG: ACRA, SGX announcements, Gazette
• US: SEC filings, state business registry, Chapter 11 notices
• CN: National Enterprise Credit Information Publicity System, court announcements, exchange filings
• Global: Reuters, Bloomberg, company press releases, major news websites

If jurisdiction is unavailable, prioritize global news and company announcements.

Logging Rule

Keep every retrieved record in raw_evidence, even if relevance is uncertain.

Each row must include:

• company_name
• taxonomy_category
• taxonomy_seed_signal
• search_query
• source_website
• title
• publish_date
• url
• snippet
• raw_text
• initial_relevance
• retrieval_timestamp
• dedup_key

Deduplication Rule

Within the same company, duplicate articles / filings referring to the same underlying event must be deduplicated in deduped_events.

Deduplication applies only within the same company.

Deduplication key: company_name + normalized_title + publish_date + source_website

Rules:

• keep all rows in raw_evidence
• keep only one row per event in deduped_events
• populate source_count
• retain matched queries in support_queries

If the same article mentions multiple monitored companies, keep one record for each company.

Excel Output Rule

Layer 1 must generate exactly one Excel workbook for each monitoring run.

File name: negative_event_monitoring_<run_date>.xlsx

Example: negative_event_monitoring_20260409.xlsx

Workbook sheets:

• raw_evidence
• deduped_events
• run_summary

Layer 2: Classification And Summary

Goal

Convert logged evidence into structured risk intelligence.

Processing Steps

1. Extract factual event statements.
2. Classify event into the taxonomy.
3. Assign severity (Low, Medium, High).
4. Assign signal_stage (early_warning, pre_default, default).
5. Generate per-company structured summary and alert text.

Taxonomy

(See references/taxonomy.md.)

Output Contract

For each monitoring run, produce:

1. Excel evidence workbook (all companies, all Layer 1 evidence retained).
2. Structured JSON summary file:

• negative_event_summary_YYYYMMDD.json or negative_event_summary_{run_id}.json

3. Alert-ready concise text per company.

Chat Output Rules When triggered from chatbox like Slack/Codex mention like @codex:

• send final result back to the same Slack or chatbox thread
• message body contains only alert related text
• attach exactly one file: the generated xlsx workbook
• do not include any other sections Alert text template:

[Company] <company_name>
Risk Level: <overall_risk_level>
Key Signals: <signal_1>; <signal_2>; <signal_3>
Implication: <risk_implication>
Source Websites: <site_1>; <site_2>

Operating Rules

• Keep taxonomy stable across runs for consistent downstream scoring.
• Keep retrieval (Layer 1) and classification (Layer 2) logically separated.
• Do not post raw evidence directly to warning groups.
• Preserve and export all raw evidence for audit and review.
• If relevance is uncertain, keep the record and lower initial_relevance; do not drop it.