Implementation Notes#

Status: Draft, 2026-09-03. Built from technical-notes.md; the product those notes describe is specified in product-spec.md, and every source tag on this page resolves in sources.md.

This is a mechanics demonstration, not a pricing or reserving result. The contractual mechanics are sourced — the 90일 면책기간 as a hard zero on the invasive tiers and its complete absence on 유사암, the one-year 50% 감액기간 measured 보험계약일 to 진단확정일, the 200 / 100 / 60 / 20 tier ladder read off a 보험금 지급기준표, the 최초 1회한 form of every diagnosis benefit, 고액암 paying in addition rather than instead, the waiver’s exclusion of 특정소액암 and 유사암 by name, the absence of any death benefit beside the statutory payment of the 계약자적립액, the nil surrender value of the 미지급형 form during the 납입기간, and the 표준해약공제액. The incidence basis is sourced too, which makes this the one morbidity model in krlib that does not open with an apology: 보험개발원 publishes a dated 「기타피부암 및 갑상선암 이외의 암 발생률」 grid by age and sex on the insured definition of cancer R5 REG-R61. Everything else is std — the tier decomposition of that grid, the whole post-diagnosis survival model, every care intensity, the lapse level, the expense and commission scales, the notional 보험가입금액 and the premium — and the 제10회 경험생명표 is released only as 평균수명 and 기대여명 REG-R33 REG-R34, so mort_table.csv is a construction and not a copy. Replace the assumption tables with company data and a real 산출방법서 before drawing any conclusion.

Run it#

python products/cancer/run.py            # the anchor cell, point_id = 1
python products/cancer/run.py 8          # the treatment-cost-only shape

run.py prints the model point, the derived scalars, the first twelve months of the cash flow statement, the undiscounted totals grouped four ways, and the ten check_* identities. Everything it prints is ASCII, so the output lands on a Windows console under any code page: amounts are labelled KRW, and the product, the four tiers, the two timing devices, the suppressed surrender-value form and the account paid on death are romanized in Revised Romanization. Real output, with the cash-flow rows elided — they are reproduced in full, to ten decimals, in technical-notes.md:

Cancer_KR_S - am boheom (cancer insurance), KRW, monthly grid, man nai
model point 1: KR-CA-0001 - M40, bi-gaengsin (non-renewable)
  cover to man nai 100, 20-year pay, mijigeuphyeong (no surrender value while paying),
  sum insured KRW 30,000,000
  premium = KRW 45,000/month   frame = 721 rows (t = 0 .. 720)   myeonchaek = 3 m   gamaek = 12 m
  tiers: gohaek 100% top-up / ilban 100% / soaek 60% / yusa 20% of the sum insured
  modules: diag=1 hosp=1 surg=1 treat=1   waiver = cancer_diag
  pyojun haeyak gongjeaek (standard surrender charge cap) = KRW 585,000

    [ the t = 0..11 rows of result_cf(), nineteen columns ]
... 709 further months to t = 720

undiscounted totals over the whole projection (per policy issued):
  premiums              8,586,707.28
  claims, diagnosis     8,009,869.07
  claims, care          4,192,693.32
  claims, account       1,474,174.97
  claims, all          13,676,737.37
  expenses              1,810,997.43
  commissions             565,757.97
  net_cf               -7,466,785.49

checks:
  check_canc_dur_ledger      True
  check_cancer_roll_fwd      True
  check_cv_floor             True
  check_hosp_cap             True
  check_net_cf               True
  check_pols_roll_fwd        True
  check_similar_ledger       True
  check_tier_shares          True
  check_treat_ledger         True
  check_waiting_period       True

Three lines to the same thing:

import modelx as mx
model = mx.read_model("products/cancer/Cancer_KR_S")
model.Projection[1].result_cf()      # the worked example's anchor cell
model.Projection[1].result_pols()    # counts, decrement rates, ledgers, the account

Projection takes a point_id; Projection[1] is the worked-example anchor cell. result_cf() returns a DataFrame indexed by policy month t with pols_if first and net_cf last, the three state counts published beside pols_if and the benefit side split into ten claims_* columns; expenses there is acquisition plus maintenance, with the claim handling cost in its own claim_expenses column. result_pols() publishes the four diagnosis flows, the three decrement rates, the 유사암 ledger and the 계약자적립액 with both surrender values beside it — the last because the whole of the 미지급형 form is the gap between cv_pp and cv_std_pp. model.Projection.doc carries the notes’ symbols mapped to cells names and states the age basis; model.Data.doc says what each input file is and, for the two decrement files, which is published data and which is a construction.

Three in-force states, and why the reduced tier is not one of them#

The model carries three in-force states rather than the chassis’s one, and the reason is the premium waiver rather than the benefit:

Cells	Meaning	Pays premium	Can lapse	Excess hazard
`pols_healthy(t)`	never invasively diagnosed	yes	yes	none
`pols_minor(t)`	first invasive diagnosis was a 특정소액암	yes	yes	특정소액암 basis
`pols_waived(t)`	has had a 일반암 or 고액암	no	no	general basis

The split is contractual, not a refinement. 「특정 소액암 … 은 보험료 납입을 면제하지 않습니다」 [S3 제14조제1항] [S1 제9조제1항]: a 특정소액암 life goes on paying and can still surrender, while a 일반암 life pays nothing and — on the 미지급형 form during the 납입기간 — has nothing to surrender. Folding the two together stops a premium the contract goes on charging. At t = 240 the 특정소액암 state is 0.0127869424 of the block, 1.77% of the in-force, and it has been paying ₩45,000 a month for twenty years.

The transition 특정소액암 → 일반암 is modelled: it is folded into surv_minor as a (1 − i_g · cover) factor and appears as the second limb of diag_gen. The reverse is not. A 일반암 life’s later 특정소액암 and a 고액암 after a plain 일반암 are both std omissions, and both understate.

The 유사암 tier is emphatically not a fourth state. It is a second benefit at a second rate on its own once-only ledger; it moves no life anywhere, stops no premium, and carries no excess mortality at all — 갑상선 five-year relative survival is 100.2% and lifetime 갑상선 mortality risk 0.1% R1, so the tier appears in no row of survival_table.csv. Implementing it as a discount on the main diagnosis benefit gets the amount right and the ledger, the waiver and the waiting period all wrong.

Six select-duration cohorts, and a thirteen-month delay#

Relative survival is steeply select — most of a cancer diagnosis’s excess mortality falls in the first two years, and 62.1% of Korea’s prevalent cancer population is more than five years out from diagnosis R1. A flat hazard fitted to the five-year point therefore kills long survivors far too fast, and long survivors are exactly whom the inpatient, surgery and treatment limbs are paid for. Each diagnosed state is resolved into six cohorts by elapsed duration — select years 1 to 5 and an ultimate — and the excess hazard is read per cohort.

The cohorts are tracked exactly, as a delay on the entry flow rather than as a transfer rate. waived_grad(t, k) and minor_grad(t, k) carry the entrants of a month twelve months earlier forward on that cohort’s own decrements, and the graduation terms telescope out of the sum — which is what check_cancer_roll_fwd() asserts. The delay is thirteen months, not twelve: waived_grad(t, 1) reads diag_gen(t − 13), because the entry month is itself a full month of cohort-1 exposure. That off-by-one is invisible to the roll-forward checks, which close either way because the graduation terms cancel, so check_canc_dur_ledger() rebuilds cohort 1 independently from the entry history and is the only identity that catches it.

Two waiting periods, one of which is zero#

cover(t) multiplies every invasive-tier benefit and both invasive transitions, and is zero for t < 3. In months 0, 1 and 2 the model diagnoses nobody with an invasive cancer, pays nothing for one, and still collects the premium — so pols_minor(3) and pols_waived(3) are exactly 0.0000000000 and every care line at t = 3 is exactly zero. check_waiting_period() asserts the whole of that, gating the transition and not merely the claim: a model that suppresses only the benefit leaves a diagnosed population the contract says does not exist.

cover_similar(t) is the second start date and it is 1 from t = 0 — 「유사암의 보장개시일은 계약일임」 [S1], with the 면책기간 table marking 유사암 진단비 with a cross [S1] [S2]. claims_diag_similar(0) = 177.9475000000 is the only non-zero benefit in the first three rows of the anchor cell. Reading one wait_months off the model point and applying it to all four tiers is the commonest way to break this product, and at young ages it is not a small error: at female 만나이 30 the model’s 유사암 incidence, 0.001136 a year, exceeds the invasive base rate of 0.001005 that it is a ratio of.

The horizon is set by the contract, not by a table’s terminal age: proj_len() = 12 × (100 − issue_age) + 1 is the number of projected months — the frame’s exclusive end, 721 on the anchor cell, so result_cf() has 721 rows indexed t = 0 … 720 and the frame is range(proj_len()). t is 0-based: t = 0 is the first projected month, the one beginning at the 보험계약일, and the policy year is the contractual label t // 12 + 1. At t = proj_len() − 1 the cover ends at the 100세 계약해당일, pols_maturity takes the whole remaining exposure — 0.0103446076 — every cash flow is zero and claims(t, "MATURITY") is zero, because nothing is paid at expiry [S8]. The 만기환급형 2종 variant returning 5% of 보험가입금액 [S8] is a different product and is out of scope.

The 감액기간 is a first-year phenomenon, not a benefit scaling#

reduction_factor(t) is 0.50 for t < reduction_months() and 1.00 after, and it multiplies the four diagnosis lines and nothing else [S1] [S6] R6. Baking 0.50 into benefit_ratio halves the liability for sixty years instead of one; on a 갱신계약 the device is disapplied altogether — 「※ 갱신계약의 경우 감액지급을 적용하지 않습니다」 [S2] [S4] — which is what model point 3 exercises. The observed designs are 0, 12 and 24 months and reduction_months is a model point column carrying all three.

One refinement is deliberately not implemented and the direction is stated. The clock’s second endpoint differs by benefit: it runs to the 진단확정일 for a diagnosis benefit [S3 별표 1 주2] but to the 수술일 for a surgery or treatment benefit [S4] [S5]. So a cancer diagnosed in month 10 and operated on in month 14 really does draw a reduced diagnosis benefit and a full surgery benefit — and so does the model. reduction_factor multiplies the four diagnosis lines only and never the care limbs, so the model gives the contract’s own answer whenever the treatment date falls outside the 감액기간, which includes every case in which the diagnosis itself does, and overstates the care limbs of a treatment falling inside it.

Two columns that are deliberately zero, and one that is not a death benefit#

Three of the ten claims_* columns are published as zeros rather than dropped, because a column of zeros states a product fact where a missing column hides it.

claims_maturity is zero at every t. Nothing is paid at the 100세 계약해당일 on the 순수보장형 composite [S8], and the only retrieved surrender-value illustration shows the value returning to nil at maturity [S8].
claims_lapse is identically zero for the whole 납입기간. cv_pp(t) = 0 for every t < 240 on the 미지급형 form [S3 제41조], so twenty years of lapses cost nothing in cash. On a 전기납 미지급형 contract — model point 7 — it is zero at every duration, because the payment period never ends.
claims_death is not a death benefit. The composite carries none. What that column holds is the 계약자적립액, which 감독규정 제7-63조제1항제1호 requires a 제3보험 product to pay when the insured dies of a cause the policy does not cover REG-R17, which the 표준약관 implements — 「회사가 적립한 사망 당시의 계약자적립액」 REG-R25 제22조 — and beneath which 상법 제736조 is the floor REG-R50. It is av_pp(t) × pols_death(t): zero at t = 0 because the account is nil, and zero from t = 447 because the account is exhausted. LTC_KR_S, Child_KR_S and Medical_KR_S inherit the same requirement.

The account itself is a retrospective recursion floored at zero, av_pp(t+1) = max(0, (av_pp(t) + prem_alloc_pp(t) − risk_prem_pp(t)) · (1 + i)^(1/12)), and the floor binds on the anchor cell — because the anchor’s ₩45,000 is the figure the specification states rather than the shipped basis’s own equivalence level of ₩66,289, 32.1% below it. Every other model point carries its own equivalence premium, and only points 1, 2, 4, 5, 6 and 9 exhaust the account at all. A recursion allowed to go negative would carry a fictitious asset and pay claims_death out of it, which check_net_cf() would not catch because the ledger identity balances either way. risk_prem_pp deliberately excludes the DEATH and LAPSE lines: they are payments out of the account, so including them makes the recursion self-referential — modelx raises rather than silently mis-answering, but a hand implementation will not.

Two surrender values, and a face amount this product does not have#

cv_pp(t) is what the contract actually pays and cv_std_pp(t) the 표준형 comparator that cannot be bought on this form; check_cv_floor() asserts 0 ≤ cv_pp ≤ cv_std_pp and that cv_pp is nil for the whole 납입기간 on the 미지급형 base. Two independent mechanisms move them and conflating the two is the classic error here:

the 해약공제기간 is capped at seven years by 제7-66조제1항제2호 REG-R19, so surr_chg_pp(t) = 0 from t = 84;
the 미지급형 cliff is at 납입완료, t = 240, where the value steps from nil to ₩4,078,536.79 [S3 제41조] and the lapse rate steps from 0.1% to 0.8% REG-R27 in the same row, taking claims_lapse from ₩0.00 to ₩1,891.71.

Thirteen years apart, and implementing one without the other gives a plausible-looking row that is wrong by the whole of the other factor.

The 표준해약공제액 needs a 보험가입금액, and this product has no death benefit to supply one. [별표 15] 제3호 covers only 일반사망을 보장하는 보장성보험, so a cancer contract falls into 제9호 and takes a notional face amount by scaling a term assurance’s by a risk-premium ratio, computed at the 기준연령 요건 REG-R21 REG-R9. Reproducing that ratio needs a term assurance’s risk-premium scale the model does not carry, so notional_sa_ratio = 0.60 std stands in for it. Feeding the ₩30,000,000 headline in instead gives 459,000 + 300,000 = 759,000 against 459,000 + 180,000 = 639,000 — and because the 13-month cap of ₩585,000 binds either way REG-R29, the error is invisible on the anchor cell and visible wherever the sum insured is low relative to the premium — only then does the [별표 14] formula fall below the 13-month cap. Model point 10, the sum-insured floor of ₩10,000,000 against ₩23,000 a month, is where it shows: the formula gives ₩294,600 against a cap of ₩299,000, so the notional ratio decides the answer where the headline face amount would have put it back on the cap. This is the route by which a Korean 제3보험 product with no face amount acquires one, and LTC_KR_S and Child_KR_S inherit it — with the difference that 제9호’s third bullet excludes long-term-care risk premium from the ratio REG-R21.

Processing order within month `t`#

Stated because nothing in any retrieved document states one, and because two of the six steps change the answer if they are swapped.

Premium on pols_payer(t) and maintenance expense on pols_if(t), at the start of the month. Two weights, two lines, one row.
Diagnoses, at the end: diag_gen_h and diag_minor out of pols_healthy, diag_gen_m out of pols_minor, diag_high as a subset of the general flow, and diag_similar on the whole in-force against the once-only ledger.
Diagnosis benefits, scaled by reduction_factor(t) and gated by cover(t) for the three invasive tiers and cover_similar(t) for 유사암.
Care benefits on the six duration cohorts of both diagnosed states — on the stock, never on the flow.
Decrements, in the order transition → mortality → lapse, each state on its own basis. A life diagnosed in month t takes its new state’s mortality for the rest of it.
The 계약자적립액 recursion, then claims(t, "DEATH") releasing av_pp(t) and claims(t, "LAPSE") releasing cv_pp(t).

Two consequences fall straight out of the ordering and both are pinned by tests. Diagnosis lines ride on flows and care lines ride on stocks: claims_diag_*(t) uses the month’s new diagnoses and claims_hosp/surgery/treat(t) uses pols_diag_dur(t, k), the stock at the start of the month; multiplying a care intensity by a diagnosis flow understates the care limbs by the mean diagnosed duration — 57 at 만나이 50, 80 at 60 and 148 at 80. And the care limbs start one month after the diagnosis limbs: claims_hosp(3) = 0 while claims_hosp(4) = 13.6639954092, because a life diagnosed in month 3 is in the diagnosed stock from month 4.

The tier algebra: a partition, a subset and an addition#

Four monthly incidences come off one published rate, and their algebra differs line by line. check_tier_shares() asserts exactly this and nothing more:

i_g + i_m = inc_rate / 12       일반암 and 특정소액암 PARTITION the base rate
i_h <= i_g                      고액암 is a SUBSET, paid in ADDITION
i_z >= 0                        유사암 is ADDITIVE to the base rate

고액암 pays in addition, not instead [S3], so claims_diag_gen(3) = 1,360.1145372945 and claims_diag_high(3) = 49.7602879498 are both paid on overlapping flows: a leukaemia collects 200% of S and a stomach cancer 100%. Treating 고액암 as a fifth slice of a partition halves it. And 유사암 is additive because the published grid excludes it: the bureau’s rate is stated on invasive cancer excluding 기타피부암 (C44) and 갑상선암 (C73) R5 REG-R61, which is exactly the 유사암 boundary the 약관 draw, so the grid and the reduced tier fit together rather than needing reconciliation and similar_share can and does exceed 1.0 — 1.60 at female 만나이 20. A model that constrains the four shares to sum to one prices the reduced tier out of existence at precisely the ages where it dominates.

Two ledgers that are per policy and per diagnosed life#

similar_avail(t) is the 유사암 tier’s once-only availability and it rides on pols_if, not on pols_healthy: a life who has already had an invasive cancer can still collect a 유사암 benefit, because no payment terminates or exhausts the contract [S1] [S3] [S4]. check_similar_ledger() asserts similar_avail(t) + similar_used(t) = 1 with the used side accumulated off the published claim line rather than off the recursion, so the identity cannot close by construction. At the anchor cell similar_avail(720) = 0.9172290909: 8.28% of policies consume the tier.

treat_avail(k) is the 최초 1회한 anti-cancer treatment benefit’s availability, and it is a per-diagnosed-life quantity evaluated at the midpoint of select year k, so treat_avail(1) = exp(−1.20 × 0.5) = 0.5488116361 rather than 1.0 or exp(−1.20). Reading it at the start of the year pays every entrant at full availability; reading it at the end understates. Weighting it by pols_cancer would measure the block’s consumption rather than the individual’s and defer exhaustion forever. The ultimate first-treatment hazard is exactly zero, which is what makes the once-only bound hold at any horizon: a life reaching the ultimate cohort without having drawn the benefit never draws it, so treat_cum_pp(t) converges — to 0.7516253263 on the anchor cell. check_treat_ledger() asserts it never passes 1.

One simplification is stated rather than hidden: the model runs one aggregate 유사암 ledger where the contracts pay each of the five members once. That understates the tier, and it is a std scope decision rather than an approximation.

Modules that are off in the base run#

Four constructions are implemented and held at an inert base value, so the base run reproduces the worked example while the machinery stays visible and testable.

Switch	Base	What it does
`void_adjust`	`False`	Scales `pols_if_init()` by `1 − void_prob()`, de-recognising the 0.0003357124 of policies diagnosed inside the 90-day window. An in-window diagnosis makes the affected cover 무효, not merely unpayable [S1 제28조제2항] R7 제644조 — a de-recognition, not a decrement — so it releases the premium already collected as well as the future benefit and belongs in a validity adjustment at outset. Putting it in the lapse column keeps premium income the insurer never earned
`inc_be_factor`	`1.0`	The best-estimate adjustment to the sourced incidence basis. The shipped rate is a 참조순보험요율, a net premium rate with a safety loading inside it, not a best estimate REG-R4. The claim that the loading is about 10% was seen only in a search summary and is unverified, so the factor is left at the identity rather than resting the model on an unconfirmed number. What is sourced is that the rate carries no trend allowance at all — 「현재도 예정위험률 산출 시 미래의 추세를 반영하지 않고 있음」 R4 — while Korea’s crude incidence has risen 161% since 1999 R1. Two errors, opposite signs, neither quantified
`renew_reprice_rate`	`0.0`	Steps the premium at each ten-year renewal on the 갱신형 chassis flag. Setting `chassis = "gaengsin"` already removes the 면책기간 and the 감액기간 [S2] [S4]; the base run holds the issue rate flat and records the contract-boundary tension rather than resolving it, which is a K-IFRS 1117 question REG-R60 this model does not answer. Live on model point 3
`lapse_canc_factor`	`1.0`	Inert rather than off: wherever the waiver fires a diagnosed life has no premium to miss and no surrender value to take, so `lapse_rate_canc_mth` returns zero whatever the factor is. It reaches a cash flow only on the `waiver_trigger = "none"` design, model point 9

What is not modelled, and is named so it is not mistaken for absent#

Each of these is specified in product-spec.md and switched off here, because its rate cannot be sourced even where its mechanic can:

재진단암. The two-year clock and the four-limb definition (새로운 원발암 / 전이암 / 재발암 / 잔여암) are sourced [S1] [S8], and no public source gives a cancer re-diagnosis incidence R1 R4. Understates.
The 요양병원 limb. Excluded from the inpatient benefit and its separate 90-day rider not carried [S2] [S8] — the market’s own structural answer to the most disputed benefit in Korea, 2,125 암입원비 complaints in 2018 alone R3.
암 사망, the 다빈치로봇 surgery limb and 비흡연체 rating, all three of which appear in the retrieved wordings [S2] [S3] [S5] with no rate or differential published behind them.
Care benefits on the 유사암 tier. Real contracts pay the inpatient and treatment limbs at 20–25% on 유사암 [S1]; the model pays nothing, because attaching invasive care intensities to a tier whose survival is 100.2% would credit it with an exposure no retrieved statistic measures R1. Understates.
부활. Lapse is absorbing. A reinstated Korean cancer policy re-runs the 90 days from the 부활일 [S1] [S3] [S7] REG-R25 제27조, so it is not the policy that lapsed; modelling reinstatement as a negative lapse restores cover the contract does not restore and deletes a real anti-selection control. Conservative.
The stage drift. Survival is a stage story far more than a site story and the mix is moving in the policyholder’s favour — 국한 45.6% (2005) → 51.8% (2023) R1 — which raises the cost of every post-diagnosis limb. The model holds the survival basis flat.

And one boundary in the other direction: do not reuse Medical_KR_S’s machinery here. There is no 급여/비급여 split, no 자기부담금, no annual limit and no 재가입 in this product, and no benefit here reimburses a cost. The one shared mechanic is the 제3보험 requirement to pay the 계약자적립액 on death REG-R17.

Inputs are external files#

Eight CSVs sit beside run.py, in the model folder’s parent; the model folder holds __init__.py, _system.json and its two Space folders and nothing else — no _data/, no IOSpec, no embedded values — so a diff of the model shows logic changes only. This is the annuallife.TradLife_A layout; contrast basiclife.BasicTerm_S, which keeps its inputs inside the model. The consequence worth knowing: the model is not portable on its own. Copying Cancer_KR_S without its parent’s CSVs produces a model that reads and then fails on first evaluation.

products/cancer/
    model_point_table.csv     <- inputs live here
    mort_table.csv            incidence_table.csv
    tier_share_table.csv      tier_table.csv
    survival_table.csv        care_table.csv
    lapse_table.csv
    run.py
    model.md  product-spec.md  technical-notes.md  sources.md
    Cancer_KR_S/              <- formulas only
        __init__.py  _system.json
        Data/__init__.py          (reads the CSVs, once per model)
        Projection/__init__.py    (the by-policy projection)

File	Reference	Reader	Index	Contents
`model_point_table.csv`	`model_point_file`	`Data.model_point_table()`	`point_id`	ten model points; point 1 is the anchor
`mort_table.csv`	`mort_table_file`	`Data.mort_table()`	`sex`, `age`	all-cause mortality of the never-diagnosed, 만나이 15–100
`incidence_table.csv`	`incidence_table_file`	`Data.incidence_table()`	`sex`, `age`	암 발생률 excluding C44 and C73, on the published ten-year grid
`tier_share_table.csv`	`tier_share_file`	`Data.tier_share_table()`	`sex`, `age`	the three tier shares at anchors 20 / 40 / 60 / 80
`tier_table.csv`	`tier_table_file`	`Data.tier_table()`	`tier`	the benefit ladder, each tier’s 면책기간, and which tiers waive
`survival_table.csv`	`survival_table_file`	`Data.survival_table()`	`sex`, `tier`, `dur_year`	post-diagnosis excess hazard, select years 1–5 and ultimate
`care_table.csv`	`care_table_file`	`Data.care_table()`	`dur_year`	admissions, days, operations and first-treatment hazard per diagnosed life
`lapse_table.csv`	`lapse_table_file`	`Data.lapse_table()`	`segment`	three segments, not a policy-year grid

Read once, in `Data`#

Projection is parameterized by point_id, so every Projection[N] is a separate ItemSpace with its own cells cache; readers placed there would re-read every file for every model point. They live in the unparameterized Data Space instead, and test_inputs_are_read_once_not_once_per_model_point asserts the property against the file set registered in kr_registry.INPUT_FILES. input_dir() returns _model.path.parent, resolved at run time and never hard-coded, so the model works from any checkout.

Every assumption CSV carries a provenance column and every cell in it begins with a citation tag. model_point_table.csv is the only exemption, a model point being a configuration rather than an assumption. On this product the column does more work than usual, because the eight files sit at three quite different levels of authority and a reader must be able to tell which by eye: one is published data reproduced verbatim, one is a construction fitted to public summary statistics, one is read off a 약관, and four are standardizations with nothing published behind them at all.

`incidence_table.csv` — published data, reproduced#

This is the file that makes the product an exception in this library. 보험개발원, the statutory 보험요율 산출기관 of 보험업법 제176조 REG-R4, displays its 장기손해보험 참조순보험요율 in force 적용시점 2024년 4월 1일 이후, and that display carries a 「기타피부암 및 갑상선암 이외의 암 발생률」 grid by age and sex R5 REG-R61:

만나이	0	10	20	30	40	50	60	70	80
남자	0.000297	0.000148	0.000230	0.000531	0.001343	0.003567	0.008540	0.019206	0.027892
여자	0.000318	0.000152	0.000250	0.001005	0.003382	0.004962	0.006239	0.008626	0.011452

It is dated, it has a stated effective date, and its definition is the insured one — invasive cancer excluding C44 and C73, classified by 원발부위 — so it already embodies the tier carve-out the 약관 make and the 원발부위 rule the supervisor imposed from 2011-04-01. Two things sit on top of it and both are std: inc_rate(t) interpolates log-linearly between the published ten-year ages, which reproduces every published value exactly and is locally the exponential family the curve follows; and the two rows above 80, at 90 and 100, are the age-80 rate × 1.15 held flat, on the deceleration R1’s crude bands show. Those two rows are not decoration — 22.6% of the anchor cell’s diagnosis benefit is earned at 만나이 80 and above, and their provenance says so.

A unisex basis is materially wrong at every age and wrong in opposite directions either side of about 55: at 만나이 40 the published female rate is 2.52× the male, at 80 the male is 2.44× the female R5. The registry states the crossing in terms — 「50대 초반까지는 여자의 암발생률이 더 높다가, 50대 후반부터 남자의 암발생률이 더 높아지는」 R1 — and the published reference rate crosses in the same place. Model point 2 is the anchor’s female twin for exactly this reason.

The independent check is worth carrying: interpolating R1’s male all-site crude rates to exact age 40 and deducting thyroid gives 0.001365 against the bureau’s 0.001343, a difference of +1.6% with the right sign, because C44 could not be deducted — a published net premium rate and an independently derived crude rate agreeing to within two per cent.

`mort_table.csv` — a construction, not a copy#

The opposite case, and the contrast is the point. Korea’s industry table, the 제10회 경험생명표 applied from 2024-04, is not published in full: 보험개발원 releases the 평균수명 and the 기대여명 and not the rates REG-R33 REG-R34. There is no Korean equivalent of the freely downloadable Japanese 標準生命表, so there is no published rate to anchor on. What is shipped is a Makeham

q(x) = 1 − exp( −( A + B·c^(x + 0.5) ) )                                       [std]

whose two free parameters are solved so the table reproduces the 국가데이터처 생명표’s 2024 기대여명 at 40 and 65 exactly — 남 41.9 / 19.5 and 여 47.4 / 23.7 REG-R38 — and which then returns 기대수명 at birth of 80.80 and 86.88 against the published 80.8 and 86.6. That last is a check, not a target, and it is the only external validation available. Ages 15 to 100, the composite’s issue-age range through to its 100세 만기. Drop a licensed extract in over the same schema — sex, age, mort_rate — and no formula changes.

mort_be_factor = 1.0, and unlike jplib’s 1.25 that is right: the shipped table is a population all-cause basis calibrated to public 기대여명, not a valuation table with a prudential margin, so there is no margin to unwind and scaling it would be inventing one.

`care_table.csv` — the weakest file, and it says so on every row#

No Korean source publishes cancer admissions, bed-days, operations or treatment courses per diagnosed patient. R1 publishes incidence, survival and prevalence and nothing about treatment volume, and the one published utilisation series on the 보험개발원 display is a 질병입원율 for all disease R5. So the file is std throughout: its shape standardized on the clinical ordering the contracts’ own design implies — treatment front-loaded into the first two years and decaying to a maintenance level — and its level on the 180-day-per-stay cap those contracts carry [S1] [S4] R3. Three properties are worth stating. No row’s hosp_days_adm approaches the 180-day cap, so check_hosp_cap() asserts that the contractual cap is respected, not that it bites; surg_open_yr + surg_closed_yr in select year 1 is 0.90, about one operation per newly diagnosed life, which is the sanity check; and the ultimate first-treatment hazard is exactly zero, which is what makes the 최초 1회한 bound hold at any horizon.

`lapse_table.csv` and `tier_table.csv`#

lapse_table.csv carries three segments rather than a policy-year grid, because the functional form is prescribed rather than observed. 감독규정 제7-66조제4항 permits the 미지급형 form only where the premium or benefit was calculated on a 최적해지율 REG-R19, and the FSS’s November 2024 계리가정 ruling then fixes the shape: among models converging to zero lapse at 완납 the 로그-선형 모형 is the 원칙모형, converging to 0.1%, with a post-완납 ultimate of 0.8% REG-R27. Those two endpoints are the ruling’s own numbers and are sourced; only the 4.6% starting level is std, and it has no observed range, because no public Korean lapse or persistency figure for 암보험 exists R3. The instrument-level caveat is real and is carried at the point of use: the 「IFRS17 주요 계리가정 가이드라인」 attachment was never converted from HWP, so the values are verified from the 보도자료 and the functional form is unverified at instrument level REG-R27.

tier_table.csv is the benefit ladder itself — 200 / 100 / 60 / 20 per cent of the 보험가입금액 — read directly off the one retrieved 약관 that states every tier as an amount at 보험가입금액 1,000만원 [S3 별표 1]. It also carries each tier’s own wait_months, which is where the product’s two start dates come from, and a waives_premium flag that is 1 for 일반암 and 고액암 alone [S3 제14조제1항] [S1 제9조제1항]. It is the most heavily sourced file in the directory and the only one whose every row is [S#] rather than std.

No input file is keyed by the projection index#

Not one of the eight CSVs carries the model’s t, so the move to the 0-based frame left every input file untouched. The time-like columns and what each of them is:

survival_table.csv and care_table.csv — dur_year, values 1 to 6. Elapsed select year since diagnosis, with 6 standing for the ultimate, on a clock that starts at the 진단확정일 and not at the 보험계약일. It is a 1-based contractual duration label of the same kind as a policy year, read by excess_hazard(tier, k), treat_avail(k) and the care limbs of claims(t, kind) with k = 1 … 6, and the only place it is derived from the frame at all is treat_cum_pp(t)’s k = min(6, (t − 1) // 12 + 1) for a life diagnosed at t = 0. Left as written.
mort_table.csv, incidence_table.csv, tier_share_table.csv — age. Attained 만나이, reached through age(t) = issue_age() + t // 12. Not a time index. Left as written.
model_point_table.csv — wait_months, reduction_months, pay_term_y. Elapsed counts, already 0-based by nature: wait_months = 3 means cover attaches once three whole months have elapsed, which on the 0-based frame is t = 3, and reduction_months = 12 means the 감액기간 covers t < 12. Neither is a point on the frame’s axis that has to move with it. Left as written.
tier_table.csv — wait_months. The same elapsed count, per tier. Left as written.
lapse_table.csv — segment, three named segments and no policy-year grid; the policy-year mapping happens in lapse_rate(t) through policy_year(t) = t // 12 + 1. Left as written.

Sign convention#

net_cf is income positive — premiums less every benefit line, less expenses, less claim_expenses, less commissions — which is the library-wide sign, so there is no outgo-positive liability_cf companion to publish: one stream, one sign, one name.

The identity check_net_cf() asserts, stated in one line: premiums less the ten claims_* columns, less expenses, less claim_expenses, less commissions, equals net_cf, in every projected month. It is rebuilt from the columns of result_cf() rather than from the net_cf cells, so a column wired to the wrong cells, a benefit line dropped from the table or a double-counted claim expense shows up as a non-zero residual in the very table a reader is looking at. The benefit side is swept as every column whose name begins claims_, taken as a group rather than enumerated — which it can be, because result_cf() publishes the ten splits and no bare claims subtotal beside them. The claims(t, kind) cells still returns the total when kind is omitted; it is the column that is retired, and RETIRED_COLUMNS in the conventions suite keeps it retired.

The asymmetry that defines this product’s cash-flow signature sits inside that one line: premiums are weighted by pols_healthy + pols_minor and three of the benefit lines by the diagnosed cohorts, because the waiver fires on the same first invasive diagnosis that starts every care limb. Meanwhile expenses rides on pols_if, because a waived policy is still administered. At t = 4 the premium weight and pols_if differ by ₩4.03; by 납입완료 the gap is 3.69% of the block, and by 만나이 80 the waived state alone is 17.7% of the in-force. It is invisible for the first four rows, where the two are equal, which is exactly why a first-year test does not catch it. pols_payer() is where the choice is made, and it is a product choice: on waiver_trigger = "none" — model point 9 — the diagnosed keep paying and can lapse.

Naming#

Cells names follow lifelib’s basiclife/BasicTerm_S wherever it has an analogue — pols_* for policy counts, plural nouns for cash flows, *_rate for annual rates and *_rate_mth for their monthly counterparts, claims(t, kind) with an uppercase kind string. The full notes-symbol-to-cells mapping lives in the Projection docstring, headed Notes symbol, and is not repeated here; test_the_projection_docstring_carries_the_symbol_map in tests/test_model_conventions_kr.py asserts it carries at least proj_len and model_point.

Names that needed care#

Notes	Cells	Why
`l_n(t)`, `l_w(t)`	`pols_minor` / `pols_waived`	The two diagnosed states are named for what the contract does to the premium, not for the tier that caused them — because that is the only difference that reaches a cash flow. `pols_cancer` is their sum and is published for the roll-forward alone
`D_n(t,k)`, `D_w(t,k)`	`pols_minor_dur` / `pols_waived_dur`	The `_dur` suffix marks a count keyed by elapsed duration since diagnosis, distinct from `policy_year`. `pols_diag_dur(t, k)` is the union the care limbs ride on
`Z(t)`	`similar_avail` / `similar_used`	A once-only ledger per policy, with `similar_used` accumulated off the published claim line and not off the recursion, so `check_similar_ledger` cannot close by construction
`A(k)`	`treat_avail`	Keyed by the select year `k`, not by `t`: a per-diagnosed-life availability, which is why it is not a function of the projection period at all
`w(t)` vs `w_c(t)`	`lapse_rate_mth` / `lapse_rate_canc_mth`	Two monthly lapse rates, because the waiver makes one of them identically zero. `lapse_rate` stays the annual rate, which the conventions suite asserts
`CV(t)` vs `CV_std(t)`	`cv_pp` / `cv_std_pp`	What is paid against the 표준형 comparator that cannot be bought. Publishing only one of the two hides the whole of the 미지급형 form
`alpha_cap`	`surr_chg_cap_pp`	The 표준해약공제액, a statutory cap on `surr_chg_pp`; named for the cap and not for the charge, because the two differ at every duration inside the 해약공제기간
(no symbol)	`claims(t, "MATURITY")`	Exists and returns zero at every `t`. A column of zeros states the product fact where a missing column would hide it

What this product argued in the cross-model naming review#

Three settlements in RETIRED_NAMES came out of the review and this model observes all three. yejeong_rate → prem_int_rate: 예정이율 is the pricing rate and must not share a name with the declared crediting rate decl_rate — this model has a 예정이율 and, being 금리확정형, no 공시이율 at all, and the distinction is only visible if the names differ. surr_charge_pp → surr_chg_pp, with the Korean cap spelled surr_chg_cap_pp; this product is one of the four that computes the cap from [별표 14] rather than reading a schedule. cv_ratio → cv_floor_ratio, because a bare ratio name said nothing about which of the two ratios on the 무·저해지 chassis it was; here cv_floor_ratio = 0.0 is the 미지급형 fraction during the 납입기간 and cv_post_pay_ratio = 0.5 the fraction afterwards, and one name could not have carried both.

Two names this product argued for and kept: pols_maturity rather than pols_expiry, because the count whose cover ends at the scheduled end of the contract is a maturity in the library’s vocabulary whether or not a benefit attaches; and mort_be_factor rather than mort_ae_factor, since the library adjusts a table to a best estimate rather than computing an actual-to-expected ratio, and a name meaning A/E could not have said why this model holds the factor at 1.0.

Standardizations used#

Every row is std. The sourced contractual parameters — the 90 days and the 유사암 carve-out from it, the 감액기간, the four tier ratios, the 최초 1회한 form, the waiver triggers, the 7-year 해약공제기간, the [별표 14] coefficients and the 미지급형 fractions — are in product-spec.md and technical-notes.md and are not repeated. “Observed range” is what the retrieved documents actually bound; several of them bound nothing at all, which is said rather than papered over.

Parameter	Value	Rationale	Observed range
age basis	만나이 throughout	every decrement the model uses is published on 만나이 — R1’s bands, R5’s age grid, REG-R38’s life table — and converting to the contract’s 보험나이 needs a distribution of issue dates within the policy year that no source supplies	the two differ for roughly half of all issue dates; on the steep part of the curve half a year is worth about 3.5% of the rate R5
`inc_rate` interpolation	log-linear in age between published grid ages	reproduces every published value exactly and is locally the exponential family the curve follows	linear interpolation of a grid that rises by a factor of 20.8 across the projection understates the mid-decade rate materially
incidence above age 80	age-80 rate × 1.15, flat	the published grid stops at 80 and the anchor projects to 100; the deceleration is R1’s own crude bands	log-linear extrapolation of the 70→80 slope reaches 0.0405 at male 90, a decadal step of 1.45 where R1’s male 80+ band is only 1.24x its 70-79 band
`inc_be_factor`	1.0	the shipped rate is a net premium rate, not a best estimate REG-R4; the identity is a decision, not an omission	the “about 10%” loading was seen only in a search summary and is unverified
`tier_share_table.csv`	age-graded, sex-split shares at anchors 20 / 40 / 60 / 80, linearly interpolated	the crude all-ages site rates the registry publishes mix age distributions that differ violently; ungraded shares misprice the reduced tier in both directions	at female 만나이 30 the 유사암 rate is 0.001136 against a general-tier 0.000593; by 60 the ratio is 0.20
유사암 share	a floor, not an estimate	R1 does not cover 경계성종양, does not identify 대장점막내암 inside D010–D012, and does not carry 기타피부암 in its top-ten table	10% [S6] [S7], 20% [S3] [S4] and 70% [S8] on the benefit ratio; nothing published on the incidence share
고액암 share	0.030 at male 40, graded	none of 골, 뇌 or 백혈병 is in the retrieved top-ten table	none published at all — the weakest of the three shares
one aggregate 유사암 ledger	one, not five	the contracts pay each of the five members once; the model pays the tier once	understates, and by construction
no care benefits on 유사암	zero	attaching invasive care intensities to a tier at 100.2% relative survival credits it with an exposure no statistic measures R1	real contracts pay the inpatient and treatment limbs at 20–25% on 유사암 [S1]
`mort_table.csv`	Makeham fitted to two 기대여명 anchors per sex REG-R38	the 제10회 경험생명표 is not published in full REG-R33 REG-R34, so there is no rate to anchor on	the fit returns 기대수명 80.80 / 86.88 against the published 80.8 / 86.6 — a check, not a target
`mort_be_factor`	1.0	a population all-cause basis has no prudential margin to unwind	jplib carries 1.25 against a valuation table; the two are not comparable
`survival_table.csv` grading	five select years plus a non-zero ultimate of 0.020 / 0.008	a flat hazard at 0.0834063 kills long survivors far too fast when 62.1% of the prevalent population is beyond year five R1	the five-year totals are sourced exactly; the shape between them is not
`care_table.csv`	std on every row	no Korean source publishes cancer utilisation per diagnosed patient; the only published series is a 질병입원율 for all disease R5	none; the level is anchored on the 180-day cap [S1] [S4] and the year-1 operation count of 0.90
`treat_avail(k)`	mid-cohort, `exp(−Σ hazard × span)` at month `12(k−1)+6`	reading at the start of the year pays every entrant at full availability; at the end, understates	none; the ultimate hazard is set to exactly zero so the 최초 1회한 bound holds at any horizon
lapse starting level	4.6% p.a. in policy year 1	set so the log-linear path from 4.6% to the prescribed 0.1% averages about 1.2% a year over the 20-year 납입기간	no public Korean lapse figure for 암보험 exists R3; carried from a disclosed 적용해지율 envelope (4.6% at one carrier, 8.4% at another)
lapse is absorbing	부활 not modelled	a reinstated policy re-runs the 90 days [S1] [S3] [S7]	conservative; nothing published on Korean reinstatement rates
`prem_int_rate`	2.50% p.a., 금리확정형	anchored on the 2026 평균공시이율, which the FSS Governor computes under 감독규정 제1-2조제13호 REG-R9 REG-R48	연복리 1.5% on one retrieved product [S8]; a 0.5% 최저보증이율 floor on another [S1]
`notional_sa_ratio`	0.60 of the headline sum insured	[별표 15] 제9호’s ratio needs a term assurance’s risk-premium scale the model does not carry REG-R21	working back from the 13-month cap on the specification’s illustrative 25% loading gives about 0.60 (footnote 30); on the model’s own 15% gross-to-net it gives 0.42, and the cap binds either way at the anchor, so the choice is invisible there and visible on point 10, where the sum insured is low relative to the premium
`expense_acq` / `comm_init_rate`	₩300,000 at issue / 0.6 × annualised premium	together ₩624,000, 13.9 months of premium, sitting against the FSC’s 13-month statement of the [별표 14] cap for a 보장성보험 REG-R29 REG-R20	no Korean carrier publishes an expense basis: [S1] names 계약체결비용 and 계약관리비용 and quantifies neither; [S8] states the deduction without quantifying it
`expense_maint` / `inflation_rate`	₩2,500 a month / 2.0% p.a.	on `pols_if`, because a waived policy is still administered	none published
`expense_claim_diag` / `_hosp`	₩150,000 per diagnosis / ₩30,000 per admission	on the event counts, not on `pols_if` — different weights, so publishing one column would hide a real movement	none published
`prem_load_acq` + `prem_load_maint`	10% + 5% = 15% gross-to-net	drives `prem_alloc_pp` = ₩38,250 a month while `t < 240`	the 산출방법서 is a 기초서류, filed and never published REG-R2
anchor premium	₩45,000 a month	the figure `product-spec.md` states; a modelling input, not a quoted rate	the shipped basis’s own equivalence premium is ₩66,289 — ₩45,000 is 32.1% below it — and that gap is why the anchor’s account is exhausted at `t = 447`. Scaling by the 1.2352 PV ratio gives ₩55,586 and is only a lower bound: `claims_death` and `claims_lapse` ride on the 계약자적립액 and so rise with the premium
processing order	premium → diagnosis → benefits → transition → mortality → lapse	the transition must precede mortality, or a life diagnosed in month `t` carries the healthy hazard through the month it was diagnosed in	nothing in any retrieved document states a processing order
`roll_fwd_tol`	1e-10, scaled by `sum_assured()` in the money checks	one tolerance closes identities between policy counts; the money identities compare won amounts of order 1e8	`roll_fwd_tol × S` is ₩0.003 at the anchor, far below the smallest error a reader adding up the statement could see

Tests#

tests/test_cancer_kr.py holds the notes’ worked example hard-coded as module-level tables, so a reviewer can lay it beside technical-notes.md and compare by eye rather than by re-running the model. Money is asserted to the ten decimal places the notes print, in-force counts and rates to ten, and the ledgers to ten.

The derived scalars: proj_len() = 721 and result_cf() at 721 rows, indexed t = 0 … 720, pay_months() = 240, pols_if_init() = 1.0, surr_chg_months() = 84, and surr_chg_cap_pp() = 585,000 from the [별표 14] arithmetic in full, including that the 13-month cap binds and that the [별표 15] 제9호 notional face amount is what enters it.
The benefit ladder and the two start dates: benefit_ratio and tier_wait_months for all four tiers, and that cover(2) = 0 while cover_similar(0) = 1.
The assumption basis at the anchor: inc_rate(0) = 0.001343 read verbatim from R5, the log-linear inc_rate(12) = 0.0014808079, the three tier shares at 만나이 40 and 41, the four monthly tier incidences, mort_rate(0) = 0.0011068200 and its monthly form, both excess-hazard vectors including that the five general-tier hazards sum to −ln(0.659) and the five 특정소액암 hazards to −ln(0.8706), the six treat_avail values, and the three per-diagnosed-life-month care amounts — ₩125,000.00, ₩275,000.00 and ₩548,811.64 in select year 1, and ₩8,125.00 / ₩10,833.33 / ₩0.00 in the ultimate.
The first sixteen months of result_cf() and result_pols() row by row, the four hand traces the notes carry, policy year 1 in aggregate — ₩528,108.33 of premium against ₩19,994.28 of benefit, closing at −₩145,503.30 — and the undiscounted totals over all 721 months: ₩8,586,707.2756349239 of premium, ₩8,009,869.07 of diagnosis benefit, ₩4,192,693.32 of care benefit, ₩1,474,174.97 of account payments and −₩7,466,785.4889610466 of net cash flow, with the expected-payment counts each line implies (0.1971 일반암, 0.0062 고액암, 0.0930 특정소액암, 0.0395 유사암, 0.1960 treatment).
The equivalence premium on the shipped basis: the 152.7418594890-month premium annuity, the 1.2352479168 PV ratio and the ₩55,586 lower bound it scales to, against the solved equivalence level of ₩66,289 and the shipped ₩45,000.
The account, surrender and ledger paths: av_pp(447) = 0 and zero thereafter, surr_chg_pp(84) = 0, cv_pp(t) = 0 for every t < 240, the step to ₩4,078,536.79 at t = 240 with claims_lapse moving from ₩0.00 to ₩1,891.71 in that same row, similar_avail(720) = 0.9172290909, and treat_cum_pp converging to 0.7516253263.

Every entry in the notes’ Known modeling pitfalls list has a test of its own, named after the pitfall, because each is a way an implementation can look right and be wrong: the two waiting periods and the 면책기간 stopping the transition as well as the benefit, with the premium still charged inside it and voidness a de-recognition rather than a decrement; premiums on pols_healthy + pols_minor while expenses is on pols_if; 특정소액암 not waiving and still able to lapse where a waived life cannot; 고액암 as a subset paid in addition and 유사암 as additive with a share exceeding 1.0; the 감액기간 as a first-year phenomenon; diagnosis lines on flows against care lines on stocks, starting a month later; the thirteen-month cohort delay; the per-life treatment ledger with its zero ultimate hazard and the 유사암 ledger on pols_if; relative survival as an excess hazard rather than a survivorship multiplier, and 유사암 carrying neither excess mortality nor a care benefit; the notional 보험가입금액; the 7-year 해약공제기간 against the 납입완료 cliff and the two prescribed steps landing in one row; claims_lapse identically zero; the payment on death with no death benefit and the account floor binding; risk_prem_pp excluding the DEATH and LAPSE lines; nothing paid at expiry; the ten claims_* splits with no claims column; rounded lines not re-adding and commissions(0) = 323,999.9999999999; proj_len() as the row count rather than the last index; log-linear against linear interpolation; the std incidence rows above 80 carrying 22.6% of the diagnosis benefit; 부활 re-running the 90 days; and not reusing Medical_KR_S’s machinery.

Beyond those: all ten check_* identities on all ten model points, each optional module in both positions of its switch, the result_cf() column vocabulary, the CSVs’ encoding and every assumption file’s row-by-row provenance tags, an input swapped by repointing a filename Reference, and a read → write → re-read round trip against the same golden values.

The ten checks, and what each would catch:

Check	Identity	What breaks it
`check_pols_roll_fwd`	`pols_if(t) − pols_if(t+1) = deaths + lapses + maturities`	an exit that is not one of the three — most likely a diagnosis counted as one
`check_cancer_roll_fwd`	`pols_cancer(t+1) = Σ_k exposure × survival`, the graduation terms telescoping	a cohort losing or gaining lives at a boundary
`check_canc_dur_ledger`	both states’ cohort 1 rebuilt independently from the entry history	the thirteen-month delay off by one — the only check that catches it
`check_similar_ledger`	`similar_avail(t) + similar_used(t) = 1`, used read off the claim line	the once-only tier paying twice, or the ledger riding on `pols_healthy`
`check_treat_ledger`	`treat_cum_pp(t) ≤ 1` at every `t`	a per-block rather than per-life ledger, or a non-zero ultimate hazard
`check_tier_shares`	`i_g + i_m = inc_rate/12`; `i_h ≤ i_g`; `i_z ≥ 0`	고액암 treated as a slice of a partition, or 유사암 constrained to fit inside one
`check_waiting_period`	no invasive benefit and no invasive transition before `t = 3`	gating the claim but not the transition
`check_cv_floor`	`0 ≤ cv_pp ≤ cv_std_pp`, and `cv_pp = 0` for the whole 납입기간 on 미지급형	the suppression applied to the wrong period, or the cliff at the surrender-charge date
`check_hosp_cap`	no cohort’s mean stay passes the 180-day per-stay cap	a care intensity raised past what the contract will pay
`check_net_cf`	the published statement’s own columns rebuild `net_cf(t)`	a benefit kind missing from the table, or a claim expense counted twice

tests/test_model_conventions_kr.py adds the house style, parametrized over kr_registry.MODELS rather than restated here: the two-Space layout, the external inputs read once per model with no orphan CSV, the provenance column on every assumption CSV, the docstrings and their required phrases, the age basis in the registry metadata — MONTHLY | MAN for this model — against the Projection docstring, the retired-name and retired-column registers, the result_cf() contract (indexed by t, first column pols_if, a net_cf column, all names lower_snake_case, no NaN, length equal to proj_len() and the index running 0 … proj_len() − 1), and that every check_*() returns True on every shipped model point.

python -m pytest tests/test_cancer_kr.py -q
python -m pytest tests -q