<!-- ⚠️ GENERATED FILE — DO NOT EDIT MANUALLY Source: extensions/workflows/fdic-failure-forensics/ Generator: scripts/extensions/build-adapters.mjs Edit the canonical extension definition and re-run: npm run extensions:build -->

FDIC Failure Forensics

Kind: Workflow (Codex Skill)

Retrospective failure-forensics workflow for a single FDIC-insured institution. Reconstructs the pre-failure financial timeline from public FDIC data, identifies the earliest visible warning signals, and explains likely drivers of deterioration.

FDIC Failure Forensics Workflow

Skill Type: RIGID

Follow every phase in order. Do not skip phases. Do not produce any report output before completing Phase 3.

Composition

This workflow uses:

• Persona: fdic-skill-builder — enforces FDIC data rules, dependency-tier modeling, and three-outcome-state discipline
• Tools: fdic-core-mcp + fdic-analysis-mcp — full FDIC data retrieval and analysis surface

When to Activate

• Analyze or reconstruct a bank failure
• Review what public data showed before a specific institution failed
• Perform a post-mortem or forensic review of a failed bank
• Identify early warning signals that preceded a failure

Do not activate for active institution deep dives or multi-institution screening.

Phase 1 — Institution Resolution and Failure Confirmation

Step 1: Resolve to CERT

Path A — CERT provided: fdic_search_institutions with filters: "CERT:<cert>". Path B — Name provided: Search with filters: "NAME:\"<name>\"", limit 10, include inactive. If multiple candidates, present disambiguation list and wait. Zero results: Stop and ask the user to refine or provide a CERT.

Step 2: Confirm failure

Call fdic_get_institution_failure with confirmed cert. If no failure record: hard-stop. Suggest using a deep-dive workflow instead. Extract: FAILDATE, RESTYPE, COST, QBFASSET, QBFDEP, BIDNAME, CHCLASS.

Step 3: Derive date parameters

• last_repdte: last quarter-end on or before FAILDATE
• lookback_start: quarter-end N quarters before last_repdte (default N=8)

Hard Boundary: Do NOT proceed without confirmed CERT with failure record and validated dates.

Phase 2 — Data Collection

Step 4: Financial timeline (Hard)

fdic_search_financials — cert, fields: CERT,REPDTE,ASSET,DEP,NETINC,ROA,ROE,NETNIM,EQ,LNLSNET, sort_by: REPDTE DESC, limit: lookback+2. Zero records = hard-stop.

Step 5: Risk signals (Hard)

fdic_detect_risk_signals — certs: [cert], repdte: last_repdte, quarters: lookback, min_severity: warning. Zero signals = analytically significant finding.

Step 6: Health assessment (Soft)

fdic_analyze_bank_health — cert, repdte, quarters. Failure → note and continue.

Step 7: Structural history (Soft)

fdic_search_history — cert, limit 20, sort PROCDATE DESC. Failure → note and continue.

Step 8: Domain follow-up (Conditional)

• Funding stress OR deposits down >15% → fdic_analyze_funding_profile
• Credit signals OR net loans up >20% with declining earnings → fdic_analyze_credit_concentration

Step 9: Regional context (Soft)

fdic_regional_context — cert, repdte. Failure → omit section.

Phase 3 — Report Assembly

Fixed 8-section order. Core sections (1-5, 8) always present. Sections 6-7 only if data collected.

1. Institution Identification — name, CERT, location, assets/deposits at failure
2. Failure Event Summary — FAILDATE, RESTYPE, acquirer, DIF cost, date range
3. Pre-Failure Financial Timeline — quarterly table + trend narrative, inflection point
4. Earliest Warning Signals — signals by severity, proxy summary if available; absence = key finding
5. Likely Failure Drivers — every statement tagged [Observed], [Inferred], or [Unknown from public data]
6. Domain Analysis (conditional) — funding and/or credit concentration
7. Regional Context (if available)
8. Caveats and Limitations — data source, temporal gap, proxy disclaimer, retrospective framing

Output Rules

• Neutral, forensic, supervisory-safe tone
• Every Section 5 statement must be tagged
• Exact dates always
• Three outcome states: No data / Not applicable / Tool failure
• "the data showed" not "the bank should have"
• No supervisory impersonation