Credit Risk Committee Analyst
This skill prepares committee-ready JSON deliverables for a shared credit
office. Every task follows the same core pattern: fetch API data, apply the
policy rules encoded in the references, and produce valid JSON conforming
exactly to the provided answer template.
Core workflow
- Read
input/payloads/answer_template.json to learn the exact output shape,
enum values, ordering rules, and precision requirements.
- Fetch the API manifest at
<TASK_ENV_BASE_URL>/api/manifest to confirm
available benchmark versions and record counts.
- Fetch credit policies at
<TASK_ENV_BASE_URL>/api/policies. These encode
risk-rating thresholds, CDFI factor scores, CRE weighted-scoring weights,
stress formulas, and capacity/concentration rules. The references/policies.md
document reproduces these rules for quick lookup during computation.
- Fetch branch-level data at the relevant endpoints (branch details, metrics,
loans, sector-exposures, applications).
- Fetch benchmark data (FDIC or NCUA) when the task references benchmarks.
- For credit-union tasks, fetch the segment endpoint.
- Apply the policy rules step by step and populate the answer template.
- Output only valid JSON conforming to the template. No narrative text
outside the JSON.
Task archetypes
The API supports five committee-product archetypes. Match the task to the
closest archetype and follow its computation guide.
1. Rating migration review (branch portfolio regrade)
Re-derive risk ratings for a filtered loan population, summarize migration,
compute NPA benchmark variance, and flag material downgrades and the top
problem credit.
Fetch: branch, branch loans, branch metrics, policies, FDIC benchmark.
Computation steps:
- Filter loans to the target rating floor (e.g. current_rating >= 3).
- For each loan, re-derive the rating from three factors (see
references/policies.md -- "Risk Rating Rules"):
- DSCR-based rating from dscr_thresholds.
- LTV-based rating from ltv_thresholds.
- Delinquency-based rating from delinquency_minimums (skip if Current).
- Final rating = worst (highest numeric) of the available factor ratings.
- Compute
final_rating_exposure_totals: group loans by final_rating, sum
loan_count and exposure, sort ascending by final_rating.
- Compute
migration_from_current_rating_3: subset of loans whose
current_rating equals the exact floor (e.g. 3), grouped by final_rating,
sorted ascending.
- Material downgrade: a loan whose final_rating minus current_rating >=
material_downgrade_notches (2). List them sorted ascending by loan_id.
- NPA benchmark: NPA exposure = sum of outstanding_balance for loans with
payment_status Nonaccrual or final_rating >= 8. Compute branch NPA ratio as
NPA exposure / total_loans_outstanding. Compare against the FDIC
total_loans_noncurrent_pct. Variance = branch ratio - benchmark ratio.
Variance in bps = variance * 10000. Round to 2 decimal places.
- Top problem credit: the loan with the worst combination of final_rating
and exposure. Among loans with the worst final_rating, pick the highest
exposure.
- Watch-list action coverage: assign a recommended action to loans needing
follow-up after regrade. Loans with final_rating 7 that are 90+DPD or
Nonaccrual ->
special_assets. Loans with final_rating 8 ->
partial_chargeoff_review. Loans with final_rating 6 -> watchlist. Loans
with final_rating 7 and payment_status Current -> special_assets. See
references/reason_codes.md for the full
action-assignment rules. Group covered loans by action, sorted ascending by
action.
2. Lending-committee allocation package
Allocate a branch's quarterly lending capacity across pending applications,
identify concentration flags, and report post-approval sector concentrations.
Fetch: branch, branch metrics, branch sector-exposures, branch applications,
policies.
Computation steps:
- Read
lending_capacity_q1 from branch details.
- Read all pending applications for the branch. Prioritize by risk quality:
applications with stronger metrics (higher DSCR, lower LTV, higher FICO,
longer relationship) rank higher. Approved and conditionally approved
applications appear in
priority_ranking.
- For each application, decide: approve (fully within capacity and no policy
issues), conditional_approve (needs mitigation like participation or SBA
guaranty), decline (hard policy failures), or defer.
bank_capacity_used for a participation_required application = the bank's
retained portion (approved_amount minus the participation share). For
sba_guaranty_required, the SBA-guaranteed portion does not consume bank
capacity.
gross_approved_amount = sum of approved_amount for approve +
conditional_approve decisions. committed_capacity_amount = sum of
bank_capacity_used for those same decisions. remaining_capacity =
lending_capacity_q1 - committed_capacity_amount.
- Concentration flags: for each sector with a pending application, compute
post_approval_pct = (existing sector exposure + approved_amount) /
total_loans_outstanding. Compare against the sector's limit_pct from the
sector-exposures endpoint. Flag true when post_approval_pct > limit_pct.
Set handling per references/reason_codes.md.
- Decline reasons: assign controlled reason codes from the template's enum
per the rules in references/reason_codes.md.
- Post-approval concentrations: for every sector with any existing
exposure, add the approved amounts and recompute post_approval_pct. Flag
over_limit when post_approval_pct > limit_pct.
3. Credit-union segment posture page
Produce a controlled posture recommendation for a credit-union segment backed
by NCUA state benchmarks, peer comparison, operating controls, and escalation
triggers.
Fetch: manifest, policies, NCUA benchmark, credit-union segment.
Computation steps:
- Read the segment endpoint to get state_code, peer_states, quarterly_capacity,
minimum_checklist, internal_context, and risk_tolerance.
- From the NCUA benchmark table, extract the row for the segment's state_code
and the US row. Compute peer median: for each metric, take the median of the
peer_states' values.
- Compare NC vs US and NC vs peer_median on all four metrics: delinquency_bps,
loan_to_share_pct, roaa_bps, positive_net_income_pct. Direction is higher,
lower, or equal.
- Posture: choose based on capacity, external risk, and risk_tolerance.
- If internal delinquency is approaching or exceeding 90bps, or external
delinquency_bps is substantially above both US and peer median ->
temporarily_pause.
- If external risk is weaker but capacity available and risk_tolerance moderate
->
continue_with_tighter_conditions.
- If all metrics stronger ->
continue_approving.
See references/policies.md for the full decision matrix.
- Controls:
required_checklist_gates come from the segment's
minimum_checklist. added_operating_controls include at minimum:
lien_perfection_prior_to_funding, monthly_segment_delinquency_watch,
pre_close_insurance_binder_verification, quarterly_state_benchmark_monitoring,
senior_underwriter_second_review.
- Escalation triggers: assign trigger_ids and conditions per
references/reason_codes.md. The three standard
triggers are segment recent delinquency >= 90bps, missing insurance/lien
exceptions, and quarterly capacity exceeded.
- Interpretation: map capacity_status (capacity_available when
quarterly_capacity > 0), external_risk_status (weaker_than_national_and_peers
when NC metrics are directionally worse than both US and peer median),
risk_tolerance from the segment, and committee_message from the
posture/risk combination.
4. Watch-list stress packet
Identify adverse-rated loans, assign CDFI risk classes, stress DSCR at +200bp,
and queue workout actions.
Fetch: branch, branch loans, policies.
Computation steps:
- Filter loans to the adverse rating floor (current_rating >= 6).
- CDFI risk class: for each loan, compute a factor score by summing
contributions from fico, liquidity_months, ltv, and debt_to_asset using the
tables in references/policies.md -- "CDFI Factor
Scores". Map the total score to a risk class:
- 0-5: Prime
- 6-9: Desirable
- 10-13: Satisfactory
- 14-18: Watch
=19 and ltv <= 1.0: Doubtful
=19 and ltv > 1.0, or payment_status Nonaccrual: Projected Loss
Sort risk_classes by ascending loan_id.
- Monitoring cadence: monthly when any loan's risk_class is Watch or worse
or any loan is 90+DPD or Nonaccrual; quarterly otherwise; semiannual
only when all loans are Desirable or better and all are Current.
- DSCR stress: for every adverse loan with a DSCR value, compute
stressed_dscr = dscr / 1.18. Breach threshold = 1.0. Flag loans where
stressed_dscr < 1.0. Sort results ascending by loan_id (only loans with
DSCR available). List breached loan_ids sorted ascending.
- Workout queue: all adverse loans, sorted descending by exposure then
ascending loan_id. Assign recommended_action per
references/reason_codes.md. Set projected_loss
= true for Projected Loss risk class or Nonaccrual status, false otherwise.
- Severe bucket counts: for loans with current_rating >= 6 (the original
adverse filter), group by current_rating and payment_status, sorted
ascending by current_rating then payment_status. Sum loan_count and
exposure per bucket.
5. Competing CRE decision
Compare two CRE applications with weighted scoring, dual-factor stress, and
concentration analysis.
Fetch: branch, branch metrics, branch loans, branch sector-exposures,
branch applications, policies, FDIC benchmark.
Computation steps:
- Fetch both applications. Compute the weighted CDFI score using the CRE
weights from references/policies.md -- "CRE
Weighted Score". The five dimensions are capacity, capital, character,
collateral_exposure, and conditions. Map application fields to dimension
scores using the factor tables (a higher scored dimension is worse).
- Map each application's weighted score to a score_class: approve_quality
(<=2.0), conditional (<=3.0), weak (>3.0).
- Assign an individual decision per application based on score_class,
concentration checks, stress results, and FDIC benchmark comparison.
- Recommended path: select the stronger application (lower weighted
score). If both have the same score_class, prefer the one that passes the
stress test and triggers fewer concentration flags.
- CRE dual-factor stress: stressed_dscr = dscr * 0.85 / 1.18.
Coverage breach threshold = 1.0. Report base_dscr, stressed_dscr, and
breaches_threshold for each application.
- Concentration: use cre_policy_limit_pct from branch details. Compute
existing_cre_exposure as sum of outstanding_balance for all branch loans
where loan_type is CRE. existing_cre_concentration = existing_cre_exposure /
total_loans_outstanding. selected_post_approval_cre_concentration =
(existing_cre_exposure + selected approved_amount) /
total_loans_outstanding. selected_policy_variance_bps = (post_approval_pct -
cre_policy_limit_pct) * 10000.
- FDIC benchmark: use total_real_estate_30_89_pct. branch_delinquency_ratio
= delinquency_30_plus_pct from branch metrics. Variance = branch ratio -
FDIC ratio. Variance in bps = variance * 10000.
- Conditions: assign per conditions enum in the answer template, guided by
references/reason_codes.md. Always include
at minimum: committee_cre_exception, minimum_dscr_covenant_1_25,
quarterly_financial_reporting, and no_additional_cre_without_committee_review.
Add tenant_roll_and_lease_review and updated_appraisal_before_close when the
property type warrants it.
Computation discipline
- Always use the exact field names, enum values, and ordering rules from the
answer template. Do not invent field names or values.
- Round numeric values to the precision specified in the template (typically
2 decimals for currency, 4 decimals for ratios/percentages).
- When the template says "ascending" or "descending", sort exactly that way.
- Do not skip null/empty fields; include them with appropriate null or 0 values
as the template requires.
- Compute values from API data; do not guess or hardcode them.
- For NCUA peer median computation: sort the peer states' values and take the
middle value. For an even number of peers, average the two middle values and
round to the nearest integer.
Reference files
- references/policies.md -- Full policy rules: risk
rating thresholds, CDFI factor scores, CRE weighted scoring, stress formulas,
concentration/capacity rules. Load this before computing any committee output.
- references/api_surface.md -- API endpoint catalog,
data shapes, and field descriptions. Load this when you need to understand
what data each endpoint returns.
- references/reason_codes.md -- Decision reason
codes, watch-list action assignment rules, condition codes, decline triggers,
and escalation trigger mappings. Load when assigning codes or actions.
Common pitfalls
- Risk rating: the dominant-factor rule means final_rating is the worst
(highest numeric) of DSCR, LTV, and delinquency-derived ratings, not an
average. If a factor is unavailable (null), skip it.
- NPA definition: nonperforming loans include Nonaccrual loans AND loans
with final_rating >= 8. Do not count 90+DPD loans as NPA unless they are also
Nonaccrual.
- Capacity: bank_capacity_used is the bank's retained portion after
subtracting SBA-guaranteed or participation shares. Do not double-count.
- Concentration percentages: always divide by total_loans_outstanding from
branch metrics, not by total_assets.
- FDIC benchmark metric choice: use total_loans_noncurrent_pct for general
NPA comparisons; use total_real_estate_30_89_pct for CRE-specific delinquency
comparisons. The task context determines which metric applies.
- Sort stability: when two items have equal sort keys, use the secondary
sort specified in the template. If none is specified, use ascending loan_id
or application_id as tiebreaker.
- Loan field name variations: the API uses different field names across
endpoints (e.g., outstanding_balance in /loans vs requested_amount in
/applications; borrower_name vs applicant_name; payment_status in loans vs
no direct equivalent in applications). Always inspect the actual API response
rather than assuming field names.
1---2name: credit-risk-committee-53description: >4 Prepare credit risk committee deliverables from the shared credit office API.5 Use this skill whenever the task involves a credit-risk committee review,6 branch loan portfolio analysis, risk-rating migration review, lending-committee7 allocation package, credit-union segment posture page, watch-list stress8 packet, CRE competing-credit decision, FDIC or NCUA benchmark comparison,9 CDFI-style risk classification, DSCR stress testing, concentration-limit10 checks, or any committee-ready credit memo. The skill encodes the full policy11 framework and API surface; always load it before computing committee outputs.12metadata:13 source: https://skillmd.com/skills/prism-shadow/credit-risk-committee-514---1516# Credit Risk Committee Analyst1718This skill prepares committee-ready JSON deliverables for a shared credit19office. Every task follows the same core pattern: fetch API data, apply the20policy rules encoded in the references, and produce valid JSON conforming21exactly to the provided answer template.2223## Core workflow24251. Read `input/payloads/answer_template.json` to learn the exact output shape,26 enum values, ordering rules, and precision requirements.272. Fetch the API manifest at `<TASK_ENV_BASE_URL>/api/manifest` to confirm28 available benchmark versions and record counts.293. Fetch credit policies at `<TASK_ENV_BASE_URL>/api/policies`. These encode30 risk-rating thresholds, CDFI factor scores, CRE weighted-scoring weights,31 stress formulas, and capacity/concentration rules. The references/policies.md32 document reproduces these rules for quick lookup during computation.334. Fetch branch-level data at the relevant endpoints (branch details, metrics,34 loans, sector-exposures, applications).355. Fetch benchmark data (FDIC or NCUA) when the task references benchmarks.366. For credit-union tasks, fetch the segment endpoint.377. Apply the policy rules step by step and populate the answer template.388. Output **only** valid JSON conforming to the template. No narrative text39 outside the JSON.4041## Task archetypes4243The API supports five committee-product archetypes. Match the task to the44closest archetype and follow its computation guide.4546### 1. Rating migration review (branch portfolio regrade)4748Re-derive risk ratings for a filtered loan population, summarize migration,49compute NPA benchmark variance, and flag material downgrades and the top50problem credit.5152**Fetch:** branch, branch loans, branch metrics, policies, FDIC benchmark.5354**Computation steps:**55561. Filter loans to the target rating floor (e.g. current_rating >= 3).572. For each loan, re-derive the rating from three factors (see58 [references/policies.md](references/policies.md) -- "Risk Rating Rules"):59 - **DSCR-based rating** from dscr_thresholds.60 - **LTV-based rating** from ltv_thresholds.61 - **Delinquency-based rating** from delinquency_minimums (skip if Current).62 - **Final rating** = worst (highest numeric) of the available factor ratings.633. Compute `final_rating_exposure_totals`: group loans by final_rating, sum64 loan_count and exposure, sort ascending by final_rating.654. Compute `migration_from_current_rating_3`: subset of loans whose66 current_rating equals the exact floor (e.g. 3), grouped by final_rating,67 sorted ascending.685. **Material downgrade**: a loan whose final_rating minus current_rating >=69 `material_downgrade_notches` (2). List them sorted ascending by loan_id.706. **NPA benchmark**: NPA exposure = sum of outstanding_balance for loans with71 payment_status Nonaccrual or final_rating >= 8. Compute branch NPA ratio as72 NPA exposure / total_loans_outstanding. Compare against the FDIC73 `total_loans_noncurrent_pct`. Variance = branch ratio - benchmark ratio.74 Variance in bps = variance * 10000. Round to 2 decimal places.757. **Top problem credit**: the loan with the worst combination of final_rating76 and exposure. Among loans with the worst final_rating, pick the highest77 exposure.788. **Watch-list action coverage**: assign a recommended action to loans needing79 follow-up after regrade. Loans with final_rating 7 that are 90+DPD or80 Nonaccrual -> `special_assets`. Loans with final_rating 8 ->81 `partial_chargeoff_review`. Loans with final_rating 6 -> `watchlist`. Loans82 with final_rating 7 and payment_status Current -> `special_assets`. See83 [references/reason_codes.md](references/reason_codes.md) for the full84 action-assignment rules. Group covered loans by action, sorted ascending by85 action.8687### 2. Lending-committee allocation package8889Allocate a branch's quarterly lending capacity across pending applications,90identify concentration flags, and report post-approval sector concentrations.9192**Fetch:** branch, branch metrics, branch sector-exposures, branch applications,93policies.9495**Computation steps:**96971. Read `lending_capacity_q1` from branch details.982. Read all pending applications for the branch. Prioritize by risk quality:99 applications with stronger metrics (higher DSCR, lower LTV, higher FICO,100 longer relationship) rank higher. Approved and conditionally approved101 applications appear in `priority_ranking`.1023. For each application, decide: approve (fully within capacity and no policy103 issues), conditional_approve (needs mitigation like participation or SBA104 guaranty), decline (hard policy failures), or defer.1054. `bank_capacity_used` for a participation_required application = the bank's106 retained portion (approved_amount minus the participation share). For107 sba_guaranty_required, the SBA-guaranteed portion does not consume bank108 capacity.1095. `gross_approved_amount` = sum of approved_amount for approve +110 conditional_approve decisions. `committed_capacity_amount` = sum of111 bank_capacity_used for those same decisions. `remaining_capacity` =112 lending_capacity_q1 - committed_capacity_amount.1136. **Concentration flags**: for each sector with a pending application, compute114 post_approval_pct = (existing sector exposure + approved_amount) /115 total_loans_outstanding. Compare against the sector's limit_pct from the116 sector-exposures endpoint. Flag true when post_approval_pct > limit_pct.117 Set handling per [references/reason_codes.md](references/reason_codes.md).1187. **Decline reasons**: assign controlled reason codes from the template's enum119 per the rules in [references/reason_codes.md](references/reason_codes.md).1208. **Post-approval concentrations**: for every sector with any existing121 exposure, add the approved amounts and recompute post_approval_pct. Flag122 over_limit when post_approval_pct > limit_pct.123124### 3. Credit-union segment posture page125126Produce a controlled posture recommendation for a credit-union segment backed127by NCUA state benchmarks, peer comparison, operating controls, and escalation128triggers.129130**Fetch:** manifest, policies, NCUA benchmark, credit-union segment.131132**Computation steps:**1331341. Read the segment endpoint to get state_code, peer_states, quarterly_capacity,135 minimum_checklist, internal_context, and risk_tolerance.1362. From the NCUA benchmark table, extract the row for the segment's state_code137 and the US row. Compute peer median: for each metric, take the median of the138 peer_states' values.1393. Compare NC vs US and NC vs peer_median on all four metrics: delinquency_bps,140 loan_to_share_pct, roaa_bps, positive_net_income_pct. Direction is higher,141 lower, or equal.1424. **Posture**: choose based on capacity, external risk, and risk_tolerance.143 - If internal delinquency is approaching or exceeding 90bps, or external144 delinquency_bps is substantially above both US and peer median ->145 `temporarily_pause`.146 - If external risk is weaker but capacity available and risk_tolerance moderate147 -> `continue_with_tighter_conditions`.148 - If all metrics stronger -> `continue_approving`.149 See [references/policies.md](references/policies.md) for the full decision matrix.1505. **Controls**: `required_checklist_gates` come from the segment's151 minimum_checklist. `added_operating_controls` include at minimum:152 lien_perfection_prior_to_funding, monthly_segment_delinquency_watch,153 pre_close_insurance_binder_verification, quarterly_state_benchmark_monitoring,154 senior_underwriter_second_review.1556. **Escalation triggers**: assign trigger_ids and conditions per156 [references/reason_codes.md](references/reason_codes.md). The three standard157 triggers are segment recent delinquency >= 90bps, missing insurance/lien158 exceptions, and quarterly capacity exceeded.1597. **Interpretation**: map capacity_status (capacity_available when160 quarterly_capacity > 0), external_risk_status (weaker_than_national_and_peers161 when NC metrics are directionally worse than both US and peer median),162 risk_tolerance from the segment, and committee_message from the163 posture/risk combination.164165### 4. Watch-list stress packet166167Identify adverse-rated loans, assign CDFI risk classes, stress DSCR at +200bp,168and queue workout actions.169170**Fetch:** branch, branch loans, policies.171172**Computation steps:**1731741. Filter loans to the adverse rating floor (current_rating >= 6).1752. **CDFI risk class**: for each loan, compute a factor score by summing176 contributions from fico, liquidity_months, ltv, and debt_to_asset using the177 tables in [references/policies.md](references/policies.md) -- "CDFI Factor178 Scores". Map the total score to a risk class:179 - 0-5: Prime180 - 6-9: Desirable181 - 10-13: Satisfactory182 - 14-18: Watch183 - >=19 and ltv <= 1.0: Doubtful184 - >=19 and ltv > 1.0, or payment_status Nonaccrual: Projected Loss185 Sort risk_classes by ascending loan_id.1863. **Monitoring cadence**: monthly when any loan's risk_class is Watch or worse187 or any loan is 90+DPD or Nonaccrual; quarterly otherwise; semiannual188 only when all loans are Desirable or better and all are Current.1894. **DSCR stress**: for every adverse loan with a DSCR value, compute190 stressed_dscr = dscr / 1.18. Breach threshold = 1.0. Flag loans where191 stressed_dscr < 1.0. Sort results ascending by loan_id (only loans with192 DSCR available). List breached loan_ids sorted ascending.1935. **Workout queue**: all adverse loans, sorted descending by exposure then194 ascending loan_id. Assign recommended_action per195 [references/reason_codes.md](references/reason_codes.md). Set projected_loss196 = true for Projected Loss risk class or Nonaccrual status, false otherwise.1976. **Severe bucket counts**: for loans with current_rating >= 6 (the original198 adverse filter), group by current_rating and payment_status, sorted199 ascending by current_rating then payment_status. Sum loan_count and200 exposure per bucket.201202### 5. Competing CRE decision203204Compare two CRE applications with weighted scoring, dual-factor stress, and205concentration analysis.206207**Fetch:** branch, branch metrics, branch loans, branch sector-exposures,208branch applications, policies, FDIC benchmark.209210**Computation steps:**2112121. Fetch both applications. Compute the **weighted CDFI score** using the CRE213 weights from [references/policies.md](references/policies.md) -- "CRE214 Weighted Score". The five dimensions are capacity, capital, character,215 collateral_exposure, and conditions. Map application fields to dimension216 scores using the factor tables (a higher scored dimension is worse).2172. Map each application's weighted score to a score_class: approve_quality218 (<=2.0), conditional (<=3.0), weak (>3.0).2193. Assign an individual decision per application based on score_class,220 concentration checks, stress results, and FDIC benchmark comparison.2214. **Recommended path**: select the stronger application (lower weighted222 score). If both have the same score_class, prefer the one that passes the223 stress test and triggers fewer concentration flags.2245. **CRE dual-factor stress**: stressed_dscr = dscr * 0.85 / 1.18.225 Coverage breach threshold = 1.0. Report base_dscr, stressed_dscr, and226 breaches_threshold for each application.2276. **Concentration**: use cre_policy_limit_pct from branch details. Compute228 existing_cre_exposure as sum of outstanding_balance for all branch loans229 where loan_type is CRE. existing_cre_concentration = existing_cre_exposure /230 total_loans_outstanding. selected_post_approval_cre_concentration =231 (existing_cre_exposure + selected approved_amount) /232 total_loans_outstanding. selected_policy_variance_bps = (post_approval_pct -233 cre_policy_limit_pct) * 10000.2347. **FDIC benchmark**: use total_real_estate_30_89_pct. branch_delinquency_ratio235 = delinquency_30_plus_pct from branch metrics. Variance = branch ratio -236 FDIC ratio. Variance in bps = variance * 10000.2378. **Conditions**: assign per conditions enum in the answer template, guided by238 [references/reason_codes.md](references/reason_codes.md). Always include239 at minimum: committee_cre_exception, minimum_dscr_covenant_1_25,240 quarterly_financial_reporting, and no_additional_cre_without_committee_review.241 Add tenant_roll_and_lease_review and updated_appraisal_before_close when the242 property type warrants it.243244## Computation discipline245246- Always use the exact field names, enum values, and ordering rules from the247 answer template. Do not invent field names or values.248- Round numeric values to the precision specified in the template (typically249 2 decimals for currency, 4 decimals for ratios/percentages).250- When the template says "ascending" or "descending", sort exactly that way.251- Do not skip null/empty fields; include them with appropriate null or 0 values252 as the template requires.253- Compute values from API data; do not guess or hardcode them.254- For NCUA peer median computation: sort the peer states' values and take the255 middle value. For an even number of peers, average the two middle values and256 round to the nearest integer.257258## Reference files259260- [references/policies.md](references/policies.md) -- Full policy rules: risk261 rating thresholds, CDFI factor scores, CRE weighted scoring, stress formulas,262 concentration/capacity rules. Load this before computing any committee output.263- [references/api_surface.md](references/api_surface.md) -- API endpoint catalog,264 data shapes, and field descriptions. Load this when you need to understand265 what data each endpoint returns.266- [references/reason_codes.md](references/reason_codes.md) -- Decision reason267 codes, watch-list action assignment rules, condition codes, decline triggers,268 and escalation trigger mappings. Load when assigning codes or actions.269270## Common pitfalls271272- **Risk rating**: the dominant-factor rule means final_rating is the worst273 (highest numeric) of DSCR, LTV, and delinquency-derived ratings, not an274 average. If a factor is unavailable (null), skip it.275- **NPA definition**: nonperforming loans include Nonaccrual loans AND loans276 with final_rating >= 8. Do not count 90+DPD loans as NPA unless they are also277 Nonaccrual.278- **Capacity**: bank_capacity_used is the bank's retained portion after279 subtracting SBA-guaranteed or participation shares. Do not double-count.280- **Concentration percentages**: always divide by total_loans_outstanding from281 branch metrics, not by total_assets.282- **FDIC benchmark metric choice**: use total_loans_noncurrent_pct for general283 NPA comparisons; use total_real_estate_30_89_pct for CRE-specific delinquency284 comparisons. The task context determines which metric applies.285- **Sort stability**: when two items have equal sort keys, use the secondary286 sort specified in the template. If none is specified, use ascending loan_id287 or application_id as tiebreaker.288- **Loan field name variations**: the API uses different field names across289 endpoints (e.g., outstanding_balance in /loans vs requested_amount in290 /applications; borrower_name vs applicant_name; payment_status in loans vs291 no direct equivalent in applications). Always inspect the actual API response292 rather than assuming field names.