Implementation Notes

Status: Draft, 2026-08-20. Built from products/endowment/technical-notes.md; the product it implements is specified in product-spec.md. The savings machinery underneath it — the policy value (hokenryō tsumitatekin, 保険料積立金), the surrender value (kaiyaku-henreikin, 解約返戻金), policy loan (keiyakusha kashitsuke, 契約者貸付), automatic premium loan (jidō furikae kashitsuke, 自動振替貸付) and the grace, 失効 and 復活 chain — is specified once in whole life technical notes (終身保険) and implemented in WholeLife_JP_A, whose cells names this model carries wherever the two express the same quantity — pol_val_pp, cv_pp, surr_charge_pp, reserve_pp, loan_pp, pols_if_at, check_net_cf. This file states the deltas; where it sets an inherited parameter to a different number — the loan rate — it names the difference as a deviation and gives its reason rather than leaving it silent.

This is a mechanics demonstration, not a pricing or reserving result. Three things here are sourced and everything else is constructed. Both annual premiums are twelve times a monthly premium published for exactly the model point they sit on — the 養老保険 (yōrō hoken, endowment assurance) anchor cell [S9] and the education endowment (gakushi hoken, 学資保険) cell [S11]; the assumed interest rate (yotei riritsu, 予定利率) of 1.00% that the cash-value construction runs on is published by product group, before and after a dated revision [S9]; and the staged education money (gakushikin, 学資金) grid is one carrier’s published S型 schedule [S10] [S11]. The surrender-value formula, the expense basis and the surrender curve are published by no carrier for either cell — a sharper gap than the savings chassis faced — so α, every expense level and the whole lapse table are std. The shipped mortality table is a std construction whose provenance column points at the Institute of Actuaries of Japan (日本アクチュアリー会); it is not a copy of 生保標準生命表2018（死亡保険用）, which the publisher’s site terms forbid this library to redistribute REG-R21. Replace the tables and the cash-value basis with a company 算出方法書 before drawing any conclusion from the numbers.

Run it

python products/endowment/run.py            # the 養老 anchor cell (point_id = 1)
python products/endowment/run.py 2          # the 学資 cell

Three lines to the same thing:

import modelx as mx
model = mx.read_model("products/endowment/Endowment_JP_A")
model.Projection[1].result_cf()

Projection takes a point_id; Projection[1] is the worked example’s anchor cell and Projection[2] its second cell. result_cf() returns a DataFrame indexed by policy year t with one column per cash flow line, pols_if first and pols_if_pay and pols_wv beside it. Two further result tables sit beside it: result_pols() for the in-force probabilities and decrement rates, and result_val() for the per-policy value construction.

The model and its Projection Space both carry docstrings: model.doc describes the product and the projection basis, and model.Projection.doc holds the full mapping between the technical notes’ symbols and the cells names.

Two cells, one model

cell is a model point column and it selects between two genuinely different constructions, not two parameter sets.

On the endowment cell the contract is an endowment assurance, so the death benefit sits inside the expected present value:

pi     = S x A(x, n) / a-due(x, m)
W(t)   = S x A(x + t, n - t) - pi x a-due(x + t, max(m - t, 0))
DB(t)  = S - L(t)

On the education cell the death payment releases the value rather than adding to it, so the value is built from the survival benefits alone and the death benefit is a return of premiums floored at that value:

EPV(t) = S x [ sum over s > t of g(s) v^(s-t) (s-t)p  +  v^(n-t) (n-t)p ]
pi_g   = EPV(0) / a-due(x, m)
W(t)   = EPV(t) - pi_g x a-due(x + t, max(m - t, 0))
Wb(t)  = W(t) + S x g(t)
DB(t)  = max( P x min(t, m) - S x G(t - 1) - L(t),  Wb(t) )

Both live in pol_val_at(t, i), which branches on cell(); pol_val_pp(t) is that at i_cv and reserve_pp(t) the same at i_std. Excluding the death benefit from the education cell’s EPV is the std step, and it is one carrier’s own wording read literally: its 死亡払戻金 is the 責任準備金相当額 — the amount equivalent to the policy reserve (sekinin-junbikin, 責任準備金) [S10] — so on that design the decrement is exactly value-neutral and the composite’s max-form dominates it [S3] [S13].

Both limbs of that max must be evaluated. On the composite’s basis the value limb dominates at every duration on the shipped education cell, so the max never switches — but that is a property of that cell’s negative loading, not of the contract, and a point with a positive loading binds the other way. Hard-coding either limb passes on one cell and fails on the next, so a test raises the premium on a copy of the cell and watches the refund limb take over.

The policy value must land on its own maturity benefit

At t = n there is no future benefit beyond maturity and no future premium on either construction, so W(n) = S exactly. That is check_pol_val_terminal(), and it is the one thing a whole life chassis can never check: a whole life reserve that drifts can hide for decades, while an endowment reserve that does not converge on its own maturity benefit is wrong on the first run.

pol_val_db_pp(t) names the death benefit inside the EPV — S on the endowment cell, zero on the education cell — so that a single recursion covers both:

(W(t-1) + pi 1{t <= m}) (1 + i_cv) = q(t) DB_val(t) + (1 - q(t)) Wb(t)

That is check_pol_val_roll_fwd(), and it pins the timing as well as the arithmetic: premium credited at the start of the year, interest for the whole year, death benefit and staged benefit at the end. Its q(t) is the unadjusted table rate, because the policy value is a contractual quantity on the pricing basis: a best-estimate adjustment to the projection must not move it. That is testable rather than asserted — model point 4 carries mort_be_factor = 1.25 and its pol_val_pp series is identical to model point 2’s to the last bit.

A third identity falls out of i_std defaulting to i_cv, which it does because the numeric standard valuation rate (hyōjun riritsu, 標準利率) could not be established from any retrieved official document R4 R5: reserve_pp(t) - cv_pp(t) = surr_charge_pp(t) exactly, wherever the value exceeds the deduction. That is check_surr_charge(); the one branch it does not assert is where the deduction would exhaust the value and surr_val_pp floors at zero.

Two lives, two decrements, one policy

The education cell carries waiver of premium (hokenryō haraikomi menjo, 保険料払込免除) on the policyholder (keiyakusha, 契約者) — a second decrement on a second life who is not the insured (hihokensha, 被保険者). It has no analogue in uslib or uklib.

mort_rate(t) reads the table at the insured’s x + t - 1; mort_rate_ph(t) reads it at the policyholder’s y + t - 1. Reading one table at one age for both is the most likely implementation error on this product, and on the anchor cell the two ages coincide, so it would not show there. Model point 2 has an insured aged 0 and a policyholder aged 30, and the two rates differ by a factor of five in year 5.

mort_rate_ph(t) is zero for t > m. Every waiver trigger in the retrieved 約款 is conditional on the event falling during 保険料払込期間 [S1] [S10], and the termination-without-waiver path is the failure mode of that same provision, so after 払込満了 there is nothing for the provision to act on and the composite treats the contract as continuing through the 契約者’s death by succession [S1] [S7] [S10] [S13]. Carrying the decrement past m would terminate a further 0.8367% of policies and delete their maturity benefits, in exactly the years in which 86% of the second cell’s receipts fall.

The waiver produces no outgo line at all. What it produces is the absence of premium income: premiums(t) is carried on pols_if_pay(t) alone while maint_expenses(t) runs on pols_if(t), the total, and commissions(t) pays renewal commission on the paying state only. A waived policy costs the insurer administration and pays the distributor nothing. Because omitting the waiver altogether leaves every claim column unchanged, neither booking a phantom “waiver benefit” nor dropping the module is visible in a benefit reconciliation — so the test runs a copy of the cell with waiver = False and asserts that every claim column is bit-identical while the premium line moves.

Premiums on a waived policy are deemed paid, so prem_cum_pp(t) keeps growing on a policy that pays nothing and cv_pp(t) is the same series in both states [S1] [S10] [S13]. A model that keeps two value series is modelling a contract nobody wrote.

The staged schedule is data

benefit_schedule() reads the whole grid from benefit_schedule_table.csv, keyed by the model point’s schedule_id, and benefit_pct(t) is a lookup into it. Total receipts range from 100% to 400% of 基準保険金額 across the six carriers in the source set [S1] [S3] [S7] [S10] [S13], so a model that treats the grid as anything but data is modelling one carrier. Model point 3 runs the degenerate J variant — one payment of 100% at t = 18, then maturity — without touching a formula.

schedule_id = none on the endowment cell is a product fact and not a missing value: the survival benefit there is a single payment at t = n. The 満期保険金 is never a schedule row; it is present on both cells and held separately, so a schedule with no rows still matures.

The staged benefit is not a claim and not a decrement. It is paid on survival at a fixed anniversary to everything in force in both states and it terminates nothing. check_staged_value() asserts that it comes out of the policy value rather than beside it — Wb(t) - W(t) = S g(t) — which is the sourced constraint that each 祝金 reduces the 解約返戻金 [S7]. On the second cell the surrender value falls from ¥1,738,755.93 to ¥1,056,811.08 across the 70% payment; a model that pays the benefit beside the value inflates every later surrender.

The composite’s seam is printed, not smoothed

prem_net_level_pp() is the net level premium on i_cv and it is a derived output, never an input: the gross premium is sourced and the loading is what falls out. On the anchor cell that loading is ¥35,243.66, or 19.457% of the gross premium — plausible for a thirty-year endowment, and coherent because the premium and the rate come from the same carrier and the same release [S9]. On the education cell the same calculation gives a net premium above the gross one, a loading of −1.745% that no real product carries, because that premium is a different carrier’s [S11] and that carrier does not publish its 予定利率.

implied_rate() restates each loading as a rate, solved by bisection on prem_net_level_at(i) — −0.4239% p.a. on the anchor cell and +1.1592% on the education cell. Both are published rather than suppressed: the seam belongs to a composite built from several carriers’ documents, and hiding it would make the library less honest rather than more accurate. The bisection costs about eighty extra evaluations of the EPV cells the first time it runs, so it is deliberately not on the result_cf() path.

henreiritsu() is the 返戻率 the products are actually sold on, and it is a contractual ratio: (S x sum of g(t) + S) / (P x m), on one policy that survives, pays every premium, takes every benefit in cash and receives no dividend. It reads the contractual premium term and never the projected premium income, because it is undefined on a policy that surrenders and unbounded on a waived one. It returns 92.0099% on the anchor cell and 113.7849% on the education cell, against the “approx. 113.7%” published for exactly that plan [S11] — the carrier truncates where the model rounds. It is not a rate of return, not probability weighted, not discounted and not net of expenses, so it is not the ratio the cash-flow statement produces.

What is absent, and why that is a product fact

There is no cliff. No retrieved document offers a suppressed-surrender-value (tei-kaiyaku-henreikin-gata, 低解約返戻金型) form of either cell, so there is no k multiplier, no step at 払込満了 and no surrender spike, and no lapse-rate spike to go with them. surr_val_pp(t) and cv_pp(t) are the same series and both are published, so that the absence is stated rather than inferred. Importing the savings chassis’s k = 0.70, its step at m or its 15% surrender spike would model a product that does not exist here. Model point 7 is the only shipped point with m < n, so it is where such a step would show; its value crosses 払込満了 smoothly.

There is no tail and no terminal age. proj_len() is the 保険期間 exactly, everything closes at t = n, and the closing cash flow is a certain payment of S to the survivors rather than a decrement. lapse_rate(n) is zero std: a surrender at the end of the final year and the maturity payment fall on the same anniversary at the same amount, so running both double-counts the terminal payment and running the surrender instead of the maturity misclassifies 61% of the anchor cell’s undiscounted outgo. pols_maturity names the survivors so that the roll-forward closes in the final year, where they neither die nor surrender.

Model points

point_id	Cell	What it exercises
1	endowment	The worked example’s anchor cell: M30, 30/30, ¥5,000,000, ¥181,140 [S9]
2	education	The worked example’s second cell: child 0 / 契約者 30, 22/17, S型 grid, waiver [S10] [S11]
3	education	The degenerate J grid — the schedule is data
4	education	wv_frac = 0.90, wv_load = 1.50, mort_be_factor = 1.25 — the carve-out and both margins
5	education	wv_lapse_mult = 0.50, female lives — the waived state persists
6	endowment	F40, 20/20, dynamic surrender on and inert
7	endowment	M35, 25/15 — 短期払, where m < n
8	endowment	The automatic premium loan module on
9	endowment	A 契約者貸付 drawn at half the first year’s surrender value

Inputs are external files

Four CSVs live beside run.py, not inside the model folder. The model folder holds __init__.py and _system.json per Space and nothing else, so a diff of the model shows logic changes only. This is the annuallife.TradLife_A layout; basiclife.BasicTerm_S keeps its inputs inside the model instead. The consequence worth knowing is that the model is not portable on its own: copy Endowment_JP_A/ without its parent’s CSVs and it reads cleanly, then fails on first evaluation.

File	Contents	Provenance
model_point_table.csv	Nine model points indexed by point_id, both cells, every module switch	Points 1 and 2 sourced [S9] [S10] [S11]; the other seven std
mort_table.csv	Mortality by sex and attained age, 0 to 60, one provenance tag per row	The library-wide canonical std construction anchored to REG-R18 R1: 22 male and 18 female ANCHOR rows read at quoted ages, the rest log-linearly interpolated in ln q to five decimals. Both sexes built from their own anchors
lapse_table.csv	Surrender and premium-default rates by policy year, 4 / 3 / 2 percent and 1.0 / 0.8 / 0.6 percent	std; the industry 解約・失効率 of 5.6% is an amount-weighted all-product ceiling only R9 REG-R31
benefit_schedule_table.csv	The staged 学資金 grids, one row per payment, keyed by schedule_id	S型 grid [S10] [S11]; J degenerate variant [S10]; timing std

The readers and the four *_file References live on Data, which takes no parameters, so each file is read once per model however many model points are projected; Projection reaches them through its data Reference. Data.input_dir() resolves to _model.path.parent at run time, so the model works from any checkout location, and an input can be swapped by repointing a filename Reference with no formula change.

The mortality table is a construction, not a copy

生保標準生命表2018（死亡保険用）is published by 日本アクチュアリー会 at a stable public URL, free and in full REG-R18 — a real contrast with uklib, whose CMI tables cannot be read at all without a subscription — so anyone can retrieve it and check a rate. But the publisher’s site terms prohibit reproduction and transmission without written consent REG-R21. This library therefore cites the table by URL, quotes the individual rates its worked example needs, and ships a table whose provenance column says of each row which of two things it is: an ANCHOR row read at a quoted age, or an INTERPOLATED row filled log-linearly in ln q between the two neighbouring anchors and rounded to five decimals.

mort_table.csv is the canonical construction shared across the library, identical row for row in every jplib product that ships it, so the same cell carries the same value and the same provenance string wherever it appears. Both sexes are built the same way from their own sourced anchors — there is no ratio and no derivation of one sex from the other, and model points 5 and 6, which read the female column, read published female anchors rather than a multiple of the male ones. The file is restricted to attained ages 0 to 60, which is every age these nine model points reach: the 養老 anchor cell matures at attained age 60.

Two further distinctions survive that one, both in the Data docstring. The shipped rates trace a valuation table carrying a margin sized to roughly a 2σ level R2 REG-R20, not best-estimate experience — which is what mort_be_factor is for. And the table is read at attained age (man-nenrei, 満年齢) while the published table is built on a nearest-birthday (hoken-nenrei hōshiki, 保険年齢方式) basis REG-R20, an understatement of up to half a year of age that the notes name rather than hide.

Modules that are off in the base run

Five of the notes’ optional constructions are implemented and switched off, so the base run reproduces the worked example while the machinery stays visible and testable. Each is exercised in both positions by the tests.

Module	Switch	Off	On at	What it does
Automatic premium loan (自動振替貸付)	apl_default_mult	0.0	point 8	Scales the table premium-default rate. The advance is applied to the premium, so it moves premiums(t) and loan_pp(t) and leaves the in-force recursion alone: a default is not a lapse, and a policy does not lapse while the cash value can carry the premium [S1] [S10]. apl_advance_pp(t) is capped at the value still free of loan. Two of the six carriers do not offer it at all [S6] [S13], so the off position is a product variant and not merely a switch
Policy loan (契約者貸付)	pol_loan_util	0.0	point 9 (0.50)	Draws that fraction of cv_pp(1) at outset std, rolling up at i_loan = 2.40% [S9] — a deviation from WholeLife_JP_A’s 2.75%, and a deliberate one. The chassis picks 2.75% off a different carrier’s vintage 貸付利率 schedule, the band a contract written under the older 予定利率 falls in, and marks the pick std. Here the rate is the 契約貸付利率 this product’s own anchor carrier publishes — 2.00% → 2.40% for contracts dated on or after 2025-01-02, moved in the same release as the 予定利率 the model runs on [S9] — so the loan rate and the pricing rate come from one document. Both sets of notes give the same reason the two products need not agree: the loan rate tracks the contract’s vintage 予定利率, not the market. No cash flow of its own: it nets off death_ben_pp(t) and the surrender benefit, which is why every benefit in the base run is gross
The refused waiver	wv_frac	1.00	point 4 (0.90)	Splits the 契約者 decrement between a transition to the waived state and a termination paying pol_val_pre_pp(t) to the 契約者’s legal heirs [S1] [S7] [S10]. claims_ph_death is a column of zeros while this is 1, and that zero is the product fact
Dynamic surrender	dyn_lapse	False	point 6	w x min(3, max(1, 1 + β (CV / cumprem − 1))), β = 2.0 std. Inert wherever it is switched on, which is the finding rather than a defect: the value never reaches cumulative premiums on either cell, peaking at 92.0% at maturity on the anchor cell, so an owner is never given a value reason to surrender
Mortality margins	mort_be_factor, wv_load	1.00	point 4 (1.25, 1.50)	Two inputs and not one, because the margin points opposite ways on the two lives: on the 契約者 the waiver is a cost, so an overstated rate is prudent, while on an insured child whose death benefit is roughly the reserve already held it is nearly neutral. wv_load is the one place a separate disability decrement is right — 高度障害 is inside the table rate R2 REG-R20, the waiver’s third trigger 身体障害 within 180 days of a listed accident is not [S1] [S10] [S16] — so 1.00 understates the waiver
Waived-state persistency	wv_lapse_mult	1.00	point 5 (0.50)	Multiplies the surrender rate in the waived state. Almost certainly too high at 1.00: a waived policy receives every benefit for no further premium and has a strictly dominant reason to persist. Named so that it can be moved

Three constructions are named and deliberately not implemented, and the model says so by name rather than by silence. dividend_type is validated and the value five_year is rejected: the ５年ごと利差配当 variant [S1] [S10] needs a 配当基準 that sits in the filed but unpublished 算出方法書 REG-R2, and the notes’ cash flow equation carries no dividend term. net_cf evaluates the validator, so such a model point fails on its first cash flow rather than being projected silently under a 有配当 label. Reinstatement (復活) is not modelled either, and it costs more here than on a protection product: two carriers pay a 学資金 whose payment date fell while the policy was lapsed once the policy is reinstated [S1] [S10], so treating every exit as terminal understates later-duration in force, premium income, staged benefits and the maturity benefit together.

Reduction of the sum assured (gengaku, 減額) is the third, and it belongs to the chassis rather than to this product: reducing 基準保険金額 is treated as a partial surrender, the reduced portion releasing its own 解約返戻金 and the future premium being re-rated [S1] [S2] [S6] [S10]. It is in the composite specification and not implemented — there is no reduction year, no sum_assured_at and no partial-surrender cash flow, so every shipped model point carries one sum_assured() for the whole term. The absence is stated because it is not free here: on the education cell a 減額 re-scales the whole staged 学資金 grid, every payment of which is a percentage of 基準保険金額, and one carrier refuses it outright once the 学資年金開始日 has arrived [S6]. It is the chassis’s mechanic, so WholeLife_JP_A is where it should be built first.

Sign convention

The notes’ CF(t) is already income positive — premiums less every outgo — which is the library-wide sign of net_cf, so there is no outgo-positive liability_cf companion to publish here: one stream, one sign, one name.

Naming

Cells names follow lifelib’s basiclife.BasicTerm_S and savings.CashValue_SE wherever those models have an analogue: pols_* for policy counts, plural nouns for cash flows, *_rate for rates, *_pp for per-policy amounts, claims(t, kind) with an uppercase kind string, pols_if_at(t, timing) for the within-year in-force reads with pols_if_pay_at(t, timing) and pols_wv_at(t, timing) for the same read on one state. The full mapping from the notes’ symbols to the cells names is the table in the Projection docstring, headed Notes symbol. Six library-wide rulings and five local cases needed care beyond that.

The six rulings are the settled cross-library names, and this product carries its share of them: mort_rate_at_age(sx, z) is the table lookup keyed by age; mort_be_factor() is the multiplier turning the shipped valuation table into the projection basis (the model point column keeps its own name, mort_adj); prem_net_level_pp() / prem_net_level_at(i) are the actuarial net level premium, which is a pricing quantity and never a cash flow; check_pols_roll_fwd() / check_pols_roll_fwd_resid(t) is the per-period in-force roll-forward check; pols_if(t) is the total in force with the premium-paying subset at pols_if_pay(t) / pols_if_pay_at(t, timing); and expenses(t) is acquisition plus maintenance only, with claim_expenses(t) deducted explicitly in net_cf(t) and published as its own column.

Two of those change what a number means rather than only what it is called. pols_if was the premium-paying state here and is now pols_if_pay + pols_wv, so the result_cf() weight column is the whole surviving block — which is what the death, staged and maturity benefits and the maintenance expense have always run on, and what pols_if means in every other model in the library. And expenses no longer carries the claim handling expense, so the column is a per-policy servicing cost in this model exactly as it is in the other eight.

Notes	Cells	Why
W(t) vs Wb(t)	pol_val_pp / pol_val_pre_pp	They differ only where a staged benefit falls due at t. The death benefit and the refused-waiver termination read the value before the payment; the surrender value reads it after, which is the sourced fact that each 祝金 reduces the 解約返戻金 [S7]
q(t) for two lives	mort_rate / mort_rate_ph	The notes use one symbol shape for the 被保険者’s decrement at x + t − 1 and the 契約者’s at y + t − 1. They are separate cells on separate ages, and mort_rate_at_age is the raw lookup both read
CV(t)	cv_pp, not av_pp	This is a traditional guaranteed contract, not an account-value one, so it takes the savings family’s cash-surrender-value name. surr_val_pp is published beside it although the two are equal, so that the absence of a 低解約返戻金型 multiplier is stated rather than inferred
(no symbol)	pols_maturity	The notes write the closing payment as S x R(n) and give it no symbol. It is named so that the in-force roll-forward closes in the final year, where the survivors neither die nor surrender
(no symbol)	pol_val_db_pp	The death benefit inside the cash-value EPV — S on one cell, zero on the other. Not a cash flow; it exists so that one policy-value recursion covers both constructions

Standardizations used

Everything below is std, and each is a lever a user with a company basis replaces.

Input	Value	Where
Cash-value basis rate i_cv	1.00%, the published 予定利率 adopted directly [S9]	Reference on Projection
Reference valuation rate i_std	1.00%, defaulted to i_cv because the numeric 標準利率 could not be established R4 R5	Reference on Projection
Acquisition deduction α	0.25 of one annual premium, grading linearly to zero at m	Reference on Projection
Loan rate i_L	2.40%, held flat [S9] — the adoption is the standardization; the rate itself is sourced. Deviates from the chassis’s 2.75%, itself a std pick off a different carrier’s vintage schedule; see “Modules that are off in the base run”	Reference on Projection
Acquisition expense	¥50,000 per policy at issue	Reference on Projection
Maintenance expense	¥8,000 p.a. inflating at 1.0%, on both states	Reference on Projection
Claim expense	¥20,000 per death claim	Reference on Projection
Initial commission	90% of the annual premium	Reference on Projection
Renewal commission	3% of premium, years 2 to m, paying state only	Reference on Projection
Surrender rates	4% / 3% / 2%, zero in the final year	lapse_table.csv
Premium-default rates	1.0% / 0.8% / 0.6%, gated to zero by apl_default_mult	lapse_table.csv
Mortality basis	The table construction, mort_be_factor, wv_load	mort_table.csv, model point
Waiver behaviour	wv_frac, wv_lapse_mult, and independence of the two lives	model point
Staged-benefit timing	Each payment resolved to the policy anniversary following its stated age	benefit_schedule_table.csv
Premiums on points 6 and 7	The net premium grossed at the anchor cell’s 19.457% implied loading, rounded to the yen — ¥170,217 and ¥350,241; points 3 to 5 carry point 2’s sourced premium and points 8 and 9 point 1’s	model_point_table.csv

α is re-based on one annual premium rather than on the sum assured, because 基準保険金額 is a benefit-scaling unit and not a sum assured on the education cell — total premiums are 1.85 times it. At α = 0.25 the deduction at issue is ¥45,285 on the anchor cell, within 0.7% of the level the savings chassis calibrated against a real published surrender-value run, so the only genuine Japanese surrender-value calibration in this library is carried across rather than discarded. It satisfies the three sourced quantitative constraints [S7] — below cumulative premiums at every duration on both cells, capped at the death benefit, reduced by each 祝金 — but not the fourth, adjectival one, that the early durations return very little: cv_pp(1) is 55.4% of the first year’s premium on the anchor cell. α is the named lever and this is a listed model risk.

Two scope limits belong in this list because they change the answer. The APL exhaustion test and the commission clawback are scoped to the savings chassis by the notes and are not implemented here; apl_advance_pp(t) is capped at the value still free of loan in their place. And the 学資金 whose payment date falls during a lapse, which two carriers pay on reinstatement [S1] [S10], is out of scope with 復活 itself.

Tests

tests/test_model_conventions_jp.py asserts the house style — the folder layout, the Data / Projection split, the read-once property, the docstrings, the naming, the result_cf() conventions, that every model point projects without NaN, and that read → write → re-read reproduces the same file set and the same numbers.

tests/test_endowment_jp.py asserts this product:

• the notes’ worked example on both anchor cells, hard-coded to the yen and to six decimals of in-force so that a reviewer can check it against the notes by eye: the two mortality vectors, A(30,30) and ä(30,30), π = ¥145,896.34 with its 19.457% loading and π_g = ¥110,458.94 with its −1.745% one, the W / SC / CV table, seven cash flow rows on each cell, every line of the notes’ five traces, the undiscounted column totals, the roll-forward closing to 1.000000000, and what the waiver is worth — 1.861464% cumulative entry probability and an EPV of ¥11,384.94 against ¥1,702,626.54 of premium;

• henreiritsu() at 92.0099% and 113.7849%, and implied_rate() at −0.4239% and +1.1592%;

• every one of the notes’ known modelling pitfalls, one test named after each: that maturity is certain and not a decrement, that lapse_rate(n) is zero, that pol_val_pp(n) is sum_assured(), that the two lives read the table at two ages, that mort_rate_ph(t) is zero past m and by how much that moves the block, that the waiver produces no benefit outgo, that waived premiums are deemed paid, that 高度障害 is inside the death rate while accident-caused 身体障害 is not, that a carve-out terminates the contract, that the staged benefit is paid to both states and terminates nothing, that it comes out of the value, that the schedule is data, that both death-benefit limbs are live, that henreiritsu() is contractual, and that there is no cliff;

• the roll-forward identities rebuilt independently of the checks, and all six check_*() cells with their per-year residuals, on every one of the nine shipped model points;

• each optional module in both positions, and the structural product facts: no tail states, the 短期払 point where m < n, the income-positive sign with no liability_cf, the result_cf() column order, that the enum accessors validate rather than propagating a typo, that dividend_type = "five_year" is rejected by name on the projection path, and that every shipped assumption row carries its own provenance.

python -m pytest tests -q