Credit Office Review Skill
Overview
This skill covers five committee-task families served by a shared credit office REST API. The solver fetches live branch data, applies centralized credit policy rules, and writes structured JSON answers in the requested template shape.
The API base URL is always supplied by the task runner as <TASK_ENV_BASE_URL>.
All data is read-only; the solver does not write back to the API.
API Discovery
Start by calling GET <TASK_ENV_BASE_URL>/api/manifest to confirm the available
endpoints and record counts. Then use these endpoints as needed.
All endpoints:
| Endpoint | Returns |
|---|---|
/api/health |
Service status, record counts |
/api/manifest |
Full endpoint list, benchmark versions, seed, policy version |
/api/policies |
Credit policy rules: risk rating, CDFI factors, CRE scoring, stress, concentration |
/api/branches |
All branches list with capacities, ceilings, state, institution type |
/api/branches/{id} |
Single branch detail |
/api/branches/{id}/metrics |
Branch quarterly metrics (NPA, delinquency, deposits, charge-offs) |
/api/branches/{id}/loans |
Loan portfolio for a branch |
/api/branches/{id}/sector-exposures |
Per-sector exposure, limit, grandfathered flag |
/api/branches/{id}/applications |
Pending credit applications for a branch |
/api/benchmarks/fdic/q4-2024 |
FDIC benchmark ratios |
/api/benchmarks/ncua/q1-2025 |
NCUA state-level benchmark rows |
/api/credit-union-segments/{segment_id} |
Credit union segment profile |
Always fetch /api/policies early. The policy rules govern every task.
Credit Policy Rules
Risk Rating Re-Derivation
The policy rule is dominant factor: the final rating is the worst (highest
numeric value) among all available factors. Each factor maps to a rating via
its own threshold table. If a factor value is null, skip that factor.
DSCR → rating:
| DSCR range | Rating |
|---|---|
| ≥ 1.50 | 3 |
| ≥ 1.25 | 4 |
| ≥ 1.05 | 5 |
| ≥ 1.00 | 6 |
| < 1.00 | 7 |
LTV → rating:
| LTV range | Rating |
|---|---|
| ≤ 0.65 | 3 |
| ≤ 0.75 | 4 |
| ≤ 0.85 | 5 |
| ≤ 1.00 | 6 |
| > 1.00 | 7 |
Delinquency floor (payment status → minimum rating):
| Payment status | Minimum rating |
|---|---|
| Current | (no floor) |
| 30 Days Past Due | 4 |
| 60 Days Past Due | 5 |
| 90+ Days Past Due | 7 |
| Nonaccrual | 8 |
Material downgrade: two or more notches (final_rating - current_rating ≥ 2).
Watch-List and Workout Actions
Map the final re-derived rating to a recommended action:
| Final rating | Recommended action |
|---|---|
| 6 | watchlist |
| 7 | special_assets |
| 8 | partial_chargeoff_review |
Loans that stay at rating 3 or upgrade do not populate the watch list. Loans downgraded to rating 4 or 5 are monitored but do not trigger a watch-list action entry unless additional qualitative flags exist.
When the task asks for workout-queue ordering, sort by descending exposure and then ascending loan_id as tiebreaker.
CDFI Factor Scoring (Risk Classes)
Score each of the four factors from the policy table, then sum to get a total factor score. Classify from the total.
Factor tables (from /api/policies > cdfi_factor_scores):
Debt-to-asset: <0.40 → 0, 0.40-0.60 → 2, 0.60-0.80 → 4, >0.80 → 6
FICO: >720 → 0, 680-720 → 1, 580-679 → 3, <580 → 5
Liquidity months: >12 → 0, 6-12 → 1, 3-6 → 3, <3 → 5
LTV: <0.40 → 0, 0.40-0.60 → 2, 0.60-0.80 → 4, >0.80 → 6
Class from total score:
| Score range | Class |
|---|---|
| 0 – 5 | Prime |
| 6 – 9 | Desirable |
| 10 – 13 | Satisfactory |
| 14 – 18 | Watch |
| ≥ 19 | Doubtful |
| ≥ 19 and LTV > 1.0 | Projected Loss |
The Projected Loss class always overrides Doubtful when LTV exceeds 1.0.
DSCR Stress Tests
Two stress formulas are used depending on the task:
- Watch-list parallel shock (+200bp):
stressed_dscr = base_dscr / 1.18 - CRE dual stress (rate + vacancy):
stressed_dscr = base_dscr * 0.85 / 1.18
Breach threshold for both: stressed_dscr < 1.00.
Concentration Analysis
Branch-level sector concentration limit is in branches[].sector_ceiling_pct.
CRE-specific limit is branches[].cre_policy_limit_pct. Compare sector
exposures and post-approval concentrations against these limits.
Sector over-concentration mitigations from policy: participation_required,
reduced_amount, board_exception. Existing over-ceiling exposure may be
grandfathered (grandfathered: 1), but new approvals must not worsen that
sector without a mitigation applied.
CRE Weighted Score (5 C's)
Weighted CRE score uses five components from policy cre_weighted_score.weights:
| Component | Weight |
|---|---|
| Capacity | 0.45 |
| Capital | 0.03 |
| Character | 0.05 |
| Collateral / Exposure | 0.36 |
| Conditions | 0.11 |
Score each component from application data (lower is better). Multiply by its weight and sum. Classify:
| Score range | Class |
|---|---|
| ≤ 2.0 | approve_quality |
| 2.0 < score ≤ 3.0 | conditional |
| > 3.0 | weak |
Benchmark Metrics
FDIC Q4 2024: field total_loans_noncurrent_pct for NPA comparison.
FDIC field total_real_estate_30_89_pct for CRE delinquency comparison.
NCUA Q1 2025: a table of states with delinquency_bps, loan_to_share_pct,
roaa_bps, positive_net_income_pct. The US row is the national aggregate.
Compute branch NPA ratio as nonperforming_loans / total_loans_outstanding.
Variance bps = (branch_ratio - benchmark_ratio) * 10000.
Decision and Reason Codes
Decision enum: approve, conditional_approve, decline, defer, participation_required
Reason codes for declines:
capacity_limit— branch has insufficient remaining Q1 capacitysector_breach— approval would breach sector ceiling without mitigationweak_dscr— DSCR is below acceptable thresholdshigh_ltv— LTV exceeds policy ceilinglow_fico— FICO score is too lowrecent_bankruptcy— applicant has recent bankruptcystartup_risk— business age under 2 years with other weak factorsunderwater_collateral— LTV > 1.0policy_floor_missing— documentation or policy requirement not metdocumentation_gap— required documentation incompletefdic_adverse_variance— branch FDIC benchmark performance is materially worse than benchmarkncua_peer_weakness— NCUA state metrics trail peers
Condition codes: participation_required, reduced_amount, board_exception,
sba_guaranty_required, startup_monitoring, none
Task-Type Workflows
1. Rating Migration Review
Used when the task asks to review and re-derive risk ratings for a branch's loans in a specified rating band, summarize migration, and produce NPA benchmark analysis.
- Fetch
/api/policies,/api/branches/{id},/api/branches/{id}/metrics,/api/branches/{id}/loans,/api/benchmarks/fdic/q4-2024. - Filter loans by
current_rating >= target_current_rating_min. - Re-derive each loan's risk rating using the dominant-factor rule (DSCR, LTV, delinquency floor).
- If a loan has no available factors, retain its
current_rating. - Build
final_rating_exposure_totals: group byfinal_rating, sum loan count and exposure; sort ascending by rating. - Build
migration_from_current_rating_3: for loans that werecurrent_rating == 3, show where they migrated to (only rows where final_rating > 3). Each row hasfinal_rating,loan_count,exposure,loan_idssorted. - Build
material_downgrades: loans withfinal_rating - current_rating >= 2. Includeloan_id,current_rating,final_rating,downgrade_notches,exposure. Sort ascending byloan_id. - Build
watch_list_action_coverage: loans with final_rating ≥ 6. Group by action (final_rating 6 →watchlist, 7 →special_assets, 8 →partial_chargeoff_review). Sortby_actionascending by action name. - NPA benchmark: use
nonperforming_loansfrom Q1 2025 branch metrics (thequartervalue2025Q1). Benchmark metric istotal_loans_noncurrent_pct. Branch NPA ratio =nonperforming_loans / total_loans_outstanding. Variance bps =(branch_ratio - benchmark_ratio) * 10000. - Top problem credit: the loan with the poorest
final_ratingand, as tiebreaker, the largestoutstanding_balance. Its action follows the rating-to-action mapping above.
2. Lending Allocation Package
Used when the task asks to allocate branch Q1 lending capacity across pending applications.
- Fetch
/api/policies,/api/branches/{id},/api/branches/{id}/metrics,/api/branches/{id}/sector-exposures,/api/branches/{id}/applications. lending_capacity_q1comes from the branch record.- For each application, evaluate credit quality:
- DSCR: under 1.0 is a hard decline signal (
weak_dscr). - LTV: above 1.0 →
underwater_collateral; above 0.85 without compensating strength →high_ltv. - FICO: below 580 →
low_fico. - Startup:
years_in_business < 2with weak supporting factors →startup_risk. - Bankruptcy:
bankruptcy_months_agowithin recent window →recent_bankruptcy. - Documentation:
documentation_complete == 0→documentation_gap. - Capacity: sum of approved amounts must stay within
lending_capacity_q1. Applications that would exceed remaining capacity getcapacity_limit. - Concentration: after approval, sector exposure must not exceed
sector_ceiling_pctwithout a mitigation. Breach →sector_breach.
- DSCR: under 1.0 is a hard decline signal (
- Assign a decision for each application:
- Clear credit + capacity →
approve. - Good credit but sector concentration at ceiling →
conditional_approvewithparticipation_required. - SBA-eligible startup with marginal credit →
conditional_approvewithsba_guaranty_requiredandstartup_monitoring. - Multiple hard reasons →
decline.
- Clear credit + capacity →
priority_rankinglistsapplication_idvalues forapproveandconditional_approvedecisions, highest priority first. Order by application with the strongest credit profile (highest DSCR, strongest relationship, largest deposit balance) first, then conditional approvals.bank_capacity_usedforparticipation_requiredis the retained portion (typicallyapproved_amount * (1 - participation_pct)where participation is the fraction the bank does not retain). If no explicit participation split is given, use a standard 75/25 split (bank retains 75%).post_approval_concentrations: re-compute each sector's exposure including newly approved amounts. Comparepost_approval_pcttolimit_pct.
3. Credit Union Segment Posture
Used when the task asks for a segment-level posture recommendation backed by NCUA benchmarks.
- Fetch
/api/policies,/api/credit-union-segments/{segment_id},/api/benchmarks/ncua/q1-2025. - Extract the state row from NCUA for the segment's
state_code. - Extract the
US(national) row from NCUA. - Compute peer median from the
peer_stateslisted in the segment data: for each of the four metrics, take the median across the peer states. If an even number of peer states, average the middle two values. - Compare NC values to US and peer median on
delinquency_bps,loan_to_share_pct,roaa_bps,positive_net_income_pct. Directions arehigher,lower, orequal. - Posture: decide based on segment capacity and external risk.
- Capacity available + risk weaker →
continue_with_tighter_conditions. - Risk stronger than peers + capacity →
continue_approving. - No capacity or severe risk →
temporarily_pause.
- Capacity available + risk weaker →
- Controls:
required_checklist_gatesfrom the segment'sminimum_checklist.added_operating_controlsare the full set from the template when tighter conditions are needed:pre_close_insurance_binder_verification,lien_perfection_prior_to_funding,senior_underwriter_second_review,quarterly_state_benchmark_monitoring,monthly_segment_delinquency_watch,committee_exception_for_capacity_overrun. - Escalation triggers: use the template enum triggers, assign owners based on
the condition domain (risk →
credit_risk_manager, ops →operations_control_manager, capacity/committee →lending_committee_chair). - Interpretation: match
capacity_status,external_risk_status,risk_tolerance(from segment), andcommittee_messageconsistently.
4. Watch-List Stress and Workout
Used when the task asks to review adversely rated loans, apply CDFI risk classes, stress DSCR, and queue workout actions.
- Fetch
/api/policies,/api/branches/{id}/loans,/api/branches/{id}/metrics. - Filter loans by
current_rating >= adverse_rating_min. - For each loan, compute CDFI factor score:
- Sum scores from available factors (debt_to_asset, fico, liquidity_months, ltv).
- Skip null values.
- Classify using the CDFI class table above.
- Monitoring cadence: if any loan in the watch list is
DoubtfulorProjected Loss→monthly; if any isWatch→quarterly; otherwisesemiannual. - DSCR stress: for loans with DSCR available, compute
stressed_dscr = base_dscr / 1.18. Breach whenstressed_dscr < 1.00. Sort results ascending byloan_id. - Workout queue: all adverse loans, sorted descending by
exposure, then ascendingloan_id. Recommended action from rating-to-action map above.projected_lossistrueforProjected Lossclass. - Severe bucket counts: group loans with
current_rating ≥ 6bycurrent_ratingandpayment_status. For each bucket:loan_count,exposure. Sort ascending bycurrent_ratingthenpayment_statuslexicographically.
5. Competing CRE Decision
Used when the task asks to compare two CRE applications and recommend a path.
- Fetch
/api/policies,/api/branches/{id},/api/branches/{id}/metrics,/api/branches/{id}/loans,/api/branches/{id}/sector-exposures,/api/branches/{id}/applications,/api/benchmarks/fdic/q4-2024. - For each of the two applications, compute the weighted CDFI CRE score:
- Each of the five components is scored from the application's data.
- Use the policy weights to produce a weighted sum.
- Classify using the CRE score class table.
- Dual stress: for each application, compute
stressed_dscr = base_dscr * 0.85 / 1.18. Breach when stressed falls below 1.00. - Concentration analysis:
cre_policy_limit_pctfrom branch data.- Sum up existing CRE exposure: all branch loans with
loan_type == "CRE". - Compute
existing_cre_concentration = existing_cre_exposure / total_loans_outstanding. selected_post_approval_cre_concentration = (existing_cre_exposure + selected_app_requested_amount) / total_loans_outstanding.selected_policy_variance_bps = (post_approval_concentration - cre_policy_limit_pct) * 10000.- FDIC comparison uses
total_real_estate_30_89_pctagainst branchdelinquency_30_plus_pct.
- Recommended path: select the application with the lower (better)
weighted score. Its path depends on its score class and concentration:
approve_quality→approve(if concentration allows).conditional→conditional_approveorparticipation_required(if sector is over ceiling).- The unselected application gets
declineordeferwith its reason codes.
- Conditions for the selected application: derive from risk profile. Always include conditions that address identified weaknesses.
Answer Shape
Every task supplies an answer_template.json in input/payloads/. Read it
first. The template defines all required top-level keys, field types, enums,
and sort orders. Produce valid JSON conforming exactly to that template.
Pay close attention to:
- Numeric precision: currency fields rounded to 2 decimal places; ratios rounded to 4 decimal places; bps rounded to 2 decimal places.
- Sort order: noted on each list field in the template (ascending by
loan_id,application_id,sector,final_rating, etc.). - Enum values: use only the allowed strings from the template or policy.
Common Calculation Patterns
Re-derive a single loan rating (Python pseudocode)
def rederive_rating(loan):
ratings = []
if loan["dscr"] is not None:
dscr = loan["dscr"]
if dscr >= 1.50: ratings.append(3)
elif dscr >= 1.25: ratings.append(4)
elif dscr >= 1.05: ratings.append(5)
elif dscr >= 1.00: ratings.append(6)
else: ratings.append(7)
if loan["ltv"] is not None:
ltv = loan["ltv"]
if ltv <= 0.65: ratings.append(3)
elif ltv <= 0.75: ratings.append(4)
elif ltv <= 0.85: ratings.append(5)
elif ltv <= 1.00: ratings.append(6)
else: ratings.append(7)
status = loan["payment_status"]
if status == "Nonaccrual": ratings.append(8)
elif status == "90+ Days Past Due": ratings.append(7)
elif status == "60 Days Past Due": ratings.append(5)
elif status == "30 Days Past Due": ratings.append(4)
# "Current" adds no floor
if not ratings:
return loan["current_rating"]
return max(ratings)
Compute CDFI factor score (Python pseudocode)
def cdfi_factor_score(loan):
score = 0
if loan.get("debt_to_asset") is not None:
dta = loan["debt_to_asset"]
if dta < 0.40: score += 0
elif dta <= 0.60: score += 2
elif dta <= 0.80: score += 4
else: score += 6
if loan.get("fico") is not None:
fico = loan["fico"]
if fico > 720: score += 0
elif fico >= 680: score += 1
elif fico >= 580: score += 3
else: score += 5
if loan.get("liquidity_months") is not None:
lm = loan["liquidity_months"]
if lm > 12: score += 0
elif lm >= 6: score += 1
elif lm >= 3: score += 3
else: score += 5
if loan.get("ltv") is not None:
ltv = loan["ltv"]
if ltv < 0.40: score += 0
elif ltv <= 0.60: score += 2
elif ltv <= 0.80: score += 4
else: score += 6
return score
def cdfi_class(score, ltv):
if score >= 19 and ltv is not None and ltv > 1.0:
return "Projected Loss"
if score >= 19: return "Doubtful"
if score >= 14: return "Watch"
if score >= 10: return "Satisfactory"
if score >= 6: return "Desirable"
return "Prime"
Watch-list action from final rating (Python pseudocode)
def watch_list_action(final_rating):
if final_rating >= 8: return "partial_chargeoff_review"
if final_rating >= 7: return "special_assets"
if final_rating >= 6: return "watchlist"
return None # no watch-list action
DSCR stress (Python pseudocode)
def stress_watch_list(dscr):
return dscr / 1.18
def stress_cre_dual(dscr):
return dscr * 0.85 / 1.18
Things to Avoid
- Do not treat
collateral_valueandltvas interchangeable for risk rating; useltvdirectly. - Do not assume the Q1 2025 branch metrics row is the first element; match on
"quarter": "2025Q1". - Do not assume loans are sorted; always sort results as the template dictates.
- Do not apply the delinquency floor to loans that are
Current; the floor only applies to past-due and nonaccrual statuses. - For the NCUA table,
"US"is the national aggregate row withstate_code: "US", not a computed average of all states. - When computing peer median from NCUA
peer_states, use only the states listed in the segment'speer_statesfield. Sort the values, pick the median. - When the policy says "worst numeric rating", that means the highest integer value (a rating of 8 is worse than 7, 7 worse than 6, etc.).
Partially_requiredcondition means the bank must place a portion of the exposure with a participant. The bank retains the remainder asbank_capacity_used.- The
grandfatheredflag on sector exposures means existing over-ceiling exposure is tolerated but must not be worsened by new approvals without mitigation.