Fewshot Attempt 01

Credit Committee JSON Packet Generation

Generate committee-ready JSON packets by querying a shared credit-office public API and strictly following a provided answer template.

Setup

1. Run env/setup.sh to initialize the environment and capture the API base_url printed to stdout.
2. Read input/payloads/answer_template.json before any data fetching to know exact required top-level keys, nested shapes, enum values, ordering rules, and numeric precision.
3. Use only the public API surfaces; do not read local environment files for business data.

API Workflow

1. Discover endpoints via the manifest/root endpoint.
2. Fetch all relevant data before computing:
   • Target branch/segment details and metrics
   • Portfolio or pending applications
   • Policies (limits, floors, concentration caps)
   • Benchmark data (FDIC, NCUA, peer comparisons)
3. Cross-reference identifiers (branch_id, segment_id, application_id, loan_id) between endpoints exactly as strings.

Output Rules

• Emit only the JSON object. No markdown, no narrative, no code fences.
• Include every required top-level key and every nested required key from the template.
• Use exact enum values from the template; never invent alternatives.

Numeric Precision

Always use standard rounding (round(value, N)).

List Ordering

Follow the template's explicit ordering directive. Common patterns:

• Ascending by ID: application_id, loan_id, sector, state_code, trigger_id
• Descending by exposure, then ascending by ID: workout queues, severity-ranked items
• Alphabetical: reason codes, enum lists, conditions
• Ascending by rating, then payment status: bucket summaries

When the template specifies a sort key, apply it before serializing.

Enum Reference (Common Values)

• Decisions: approve, conditional_approve, decline, defer, participation_required
• Reason codes: capacity_limit, sector_breach, weak_dscr, high_ltv, low_fico, recent_bankruptcy, startup_risk, underwater_collateral, policy_floor_missing, documentation_gap, fdic_adverse_variance, ncua_peer_weakness
• Conditions: participation_required, reduced_amount, board_exception, sba_guaranty_required, startup_monitoring, none
• Risk classes: Prime, Desirable, Satisfactory, Watch, Doubtful, Projected Loss
• Payment status: Current, 30 Days Past Due, 60 Days Past Due, 90+ Days Past Due, Nonaccrual
• Actions: monitor, watchlist, special_assets, workout, partial_chargeoff_review, legal_referral
• Monitoring cadence: monthly, quarterly, semiannual

Business Logic Patterns

• Capacity: remaining_capacity = lending_capacity_q1 - committed_capacity_amount; sum approved amounts to compute gross_approved_amount and bank_capacity_used.
• Concentration: post_approval_pct = (existing_exposure + approved_amount) / total_portfolio_or_assets. Compare against limit_pct; flag when over limit.
• Policy variance: in basis points or percentage points: post_approval_pct - limit_pct.
• Priority ranking: include only approved and conditionally approved applications, ordered by priority (e.g., higher CDFI score first, lower risk first, or larger facility first depending on prompt).
• DSCR stress: apply the formula given in the prompt (e.g., dscr * 0.85 / 1.18), compare to breach_threshold, emit boolean breaches_threshold.
• Decline reasons: map each declined application to a sorted list of reason codes drawn from the enum.
• Concentration flags: include entries for any sector where post-approval exceeds or is near limit; sort by sector then application_id.
• Post-approval concentrations: one entry per sector, sorted ascending by sector.

Common Pitfalls

1. Wrong precision: Do not truncate; round to the exact decimal count specified.
2. Missing keys: The template's required_keys and item_required_keys are mandatory even if a list is empty.
3. Wrong sort order: Verify whether the template wants ascending or descending.
4. Narrative leakage: The final output must be parseable JSON only.
5. Stale identifiers: Use the exact branch/segment/application IDs from the prompt, not inferred names.
6. Empty lists vs omission: If a template section allows empty lists, still include the key with [].