HOYA CORPORATION (7741): Quality That Survives the Scrubbing, and a Gain Booked Because an Asset Got Cheaper

Verdicts

Lens	Verdict	Buy below	Most load-bearing items
Buffett	watch	¥7,700	B101; B42; B48/B34
Munger	watch	¥14,000 (range ¥13,500–14,800)	M14/M56/M61; M18/M40; M88
Pabrai	pass	¥6,000	P1; P53 with P60; P62
Li Lu	too-hard	¥10,500 published, explicitly not a trigger	L1/L12; L46; L18/L15
Claude	watch	implied ¥8,500, band ¥6,800–¥10,600	C35; C102; C33

Three watch, one pass, one too-hard — and five named prices running from ¥6,000 to ¥14,000 against a stamp of ¥23,840 F2428, discounts of 41% to 75%. Nobody argued about the business. The spread of 2.33× is entirely a disagreement about what object a buy-below price is: a discount off a private-owner appraisal, a historical multiple applied to normalised earnings, half of a faded intrinsic value, a required-return solve, or a bear case marked by a jury. They are set out and explained in the synthesis rather than averaged, because the average of five different operations is not a sixth answer.

The load-bearing observation: this company's elite returns survive normalisation, and the reader should notice how rarely a flattered year does. FY2026 carries ¥30,940百万円 of pre-tax one-off other income D14 — a ¥23,538百万円 re-measurement of the buy-out liability on a China intraocular-lens joint venture F2044, plus ¥7,402百万円 of disposal gains on the その他 segment sold during the year D13 — ¥23,743百万円 after tax at the group's own 23.26% effective rate D19D18. Strip all of it and return on equity is 23.00% D27, against the printed 25.4% F55 and a four-year mean of 21.00% D33; the pre-tax margin is 31.31% D47, back inside the 29.83%–31.85% band of the four prior years D43D42. Roughly 55% of the ROE step-up is one-offs D35 — and the step-up survives anyway. Return on invested capital excluding cash, ex-one-off, is 43.44% D84.

And the price is the whole argument. ¥23,840 F2428 is 35.36× ex-one-off earnings D7 rather than the printed 32.05× D5, 7.84× book D4, on a 1.24% dividend yield D8 and a 3.39% free-cash-flow yield D110. Net cash of ¥531,851百万円 D86 is 6.65% of the market capitalisation D89 — which is why the fortress balance sheet, admired by every lens, protects the enterprise and not the purchase price.

The business

ＨＯＹＡ株式会社 began in 1941 as an optical-glass works on the western edge of Tokyo, and it still carries the name of the town E8. Eighty-five years later it does one recognisable thing: it makes the precise piece of glass that somebody else's product cannot exist without. Eyeglass lenses from 1962 E11, soft contact lenses and semiconductor mask substrates from the 1970s E12, photomasks from 1983 E13, intraocular lenses from 1987 E14, hard-disk glass substrates from 1991 E16.

Two shops sit under one roof, and they are not alike. ライフケア sells things that go in front of eyes or inside them — eyeglass lenses, contact lenses, medical endoscopes, treatment instruments and reprocessors, intraocular lenses, orthopaedic implants E34E35. It did ¥590,680百万円 of outside revenue F395, 62.32% of the group D65. 情報・通信 sells the glass plate a chipmaker prints his pattern on, the photomask itself, flat-panel photomasks, the glass disc data is written to, and precision optical lenses E36. It did ¥354,751百万円 F401, 37.43% of revenue D66 — and 58.96% of segment profit D68.

Management's own account of strategy is unusually explicit, and it is a portfolio doctrine. The stated method is to run several businesses that differ in model, cycle sensitivity and selling region; to judge which stage of its life cycle each is in; to allocate resources to high-growth areas; and to withdraw from businesses whose markets are declining E44. That is not a slogan — the withdrawals are dated: crystal ended in 2009, hard-disk glass media sold to Western Digital in 2010, the PENTAX imaging business sold to Ricoh in 2011 E23, the speech-synthesis business contracted in September and completed 31 October 2025 E300, and the contact-lens retail arm carved into a wholly owned company just after the year end E313. The competitive framing is 「小さな池の大きな魚」 — a big fish in a small pond: choose narrow markets, take leadership, keep strengthening technology, business model and value chain E103E196.

Why customers keep paying, as the filing itself puts it. Life Care is described as requiring approvals from the authorities of each country plus advanced technical capability and a highly reliable quality-control system E235; Information Technology as making components indispensable to digital devices E236. Demand is named specifically: eyesight decline among the young from prolonged screen use plus global ageing; the spread of minimally invasive medicine; high-performance, energy-saving semiconductors and data-centre investment E49. The year's growth is attributed to advanced extreme-ultraviolet and deep-ultraviolet mask-blank demand, a Chinese flat-panel photomask plant ramping, and data-centre nearline demand for the larger hard-disk format — while the smaller 2.5-inch format fell substantially E94. Production is to forecast, so there is no order backlog at all E101; every period's revenue must be won again.

The counter-shape is disclosed too, and it is not small. One customer, Seagate Technology LLC, is 11.75% of consolidated sales F436, up from 10.71% F434, and the segment note places that customer group inside 情報・通信 E241. The Life Care price threat is stated by the company in its own risk register: mass retailers expanding, joint purchasing organisations forming, online operators rising, and the plain admission that the speed of price decline may exceed its ability to absorb it E82. Some raw materials have limited or hard-to-substitute suppliers E87, and most production sits in Southeast Asia, where the group's own flood assessment found relatively high exposure E63.

Governance is genuinely separated, and that is a finding. HOYA is a 指名委員会等設置会社 adopted expressly so management would not be conducted by insider logic alone E130; all members of all three statutory committees are outside directors E131; there are only three executive officers — CEO, CFO and CSO — with day-to-day authority pushed down to division heads E136E194; the audit committee has held group-wide authority over the accounting auditor since a board resolution taken long ago, with subsidiaries contracting only after its consent E169; the whistleblowing helpline has run since 2003 in every country of operation and reports periodically to the audit committee E141, with reports concerning executive officers handled by that committee directly E142. There is no parent company E317 and no related-party transactions outside subsidiaries in either year E295.

And three declarations of "none" are not true as a reader would take them. 「重要な契約等 該当事項はありません」 stands in the same document as a next-generation mask-blank technology alliance with Dai Nippon Printing, four named joint ventures, a contractual obligation to buy out a joint-venture interest, and the 11.75% customer E331E110. 「重要な偶発債務はありません」 stands against three transfer-pricing assessments from the Tokyo Regional Taxation Bureau, tax paid and carried as an asset, an appeal filed with the Tokyo High Court on 24 June 2025 — and the auditor selecting exactly that uncertainty as the sole consolidated Key Audit Matter, stating there is uncertainty whether the courts will accept the company's assertions E332E322. And 企業結合 is omitted as immaterial in a year that added goodwill through business combinations and impaired goodwill because an acquisition's own plan became unachievable E333. All five lenses found all three, independently.

The numbers

Five years, consolidated, IFRS, in 百万円. 売上収益 661,466 → 723,582 → 762,610 → 866,032 → 947,749 F1F2F3F4F5; 税引前当期利益 210,706 → 215,832 → 236,564 → 259,965 → 327,668 F6F7F8F9F10; 当期利益 165,322 → 168,788 → 182,566 → 201,750 → 251,451 F11F12F13F14F15. Revenue compounded at 9.41% over four years D59, 1.4328× cumulatively D58, with no down year in the window D54D55D56D57. 親会社所有者帰属持分当期利益率 printed 22.1 / 20.8 / 20.3 / 20.8 / 25.4% F51F52F53F54F55, on a 78.4% equity ratio F50. Book value per share went ¥2,201.68 → ¥3,041.71 F31F35; basic earnings per share ¥446.45 → ¥743.93 F36F40.

The one-off, sized to the yen. その他の収益 rose ¥31,258百万円 year on year D16, and ¥30,940百万円 of the FY2026 line is identified one-offs D14: a ¥23,538百万円 change in the fair value of financial obligations F2044E288 and ¥7,402百万円 of disposal gains on the segment sold in the year D13. The balance-sheet counterpart is unambiguous — その他の長期金融負債 falls from ¥23,793百万円 to ¥614百万円 F1019F1020, a release of ¥23,179百万円 D17. At the group's own effective rate of 23.26% D18 the after-tax value is ¥23,743百万円 D19.

Ex-one-off, line by line. 税引前当期利益 ¥296,728百万円 D20 against ¥327,668 reported F10 — still the highest of the five years. 親会社の所有者に帰属する当期利益 ¥229,342百万円 D22 against ¥253,085 reported F1102. Earnings per share ¥674.14 D6 against ¥743.93 F40. ROE 23.00% D27 against 25.38% recomputed as reported D26 and a four-year mean of 21.00% D33 — a ratio of 1.0952× to that mean, against 1.2095× on the printed figure D35D34. Pre-tax margin 31.31% D47 against 34.57% reported D46, inside the prior band of 29.83% to 31.85% D43D42. Net margin 24.03% D53, inside its own band of 23.30% to 24.99% D51D48. The operating proxy D36 is ¥315,777百万円 as reported and ¥284,837百万円 ex-one-off D38, a proxy margin of 30.05% D40 against 28.24% the prior year D41.

Returns on capital, and the limit on them. Invested capital excluding cash is ¥503,153百万円 D78; NOPAT ex-one-off — struck on that same operating proxy, less tax at the group's own effective rate — is ¥218,583百万円 D76; return on invested capital excluding cash, ex-one-off, is 43.44% D84, and 20.29% including the cash D81. That is a group figure and can never be anything else, because the segment note discloses no segment assets in either year D79: what it carries is revenue, profit, interest, depreciation, impairment and capital expenditure E238E237, and nothing from which invested capital could be built. Segment return on invested capital, segment return on assets and segment asset turnover are therefore not computable from this filing D84.

The two segment margins, and where the one-off actually sits. ライフケア printed 21.93% D60 against 16.40% the prior year D61; 情報・通信 printed 54.14% D62 against 54.68% D63 — essentially flat. But セグメント利益 is pre-tax profit F1328, so the ¥23,538百万円 F2044 sits above the line, and the joint venture that produced it is a cataract intraocular-lens business E91, which is a Life Care business. Locate it there and Life Care's margin is about 17.94% — (¥129,531百万円 F398 less ¥23,538百万円 F2044) over ¥590,683百万円 of segment revenue including inter-segment F1310 — a gain of roughly 1.5 points on the year rather than 5.5. The filing quantifies the segment effect nowhere and cross-references none of it E335.

The balance sheet has almost no other side. 資産合計 ¥1,300,897百万円 F998 against 負債合計 ¥265,893百万円 F1046; 親会社の所有者に帰属する持分 ¥1,020,460百万円 F1012; equity ratio recomputed at 78.44% D91. Interest-bearing debt is ¥42,241百万円 in total F1735, of which ¥25,628百万円 is lease liabilities D82, leaving borrowings of ¥16,613百万円 D83 at a 2.01% weighted average rate F1723. Cash is ¥574,092百万円 F994, so net cash is ¥531,851百万円 D86, up from ¥496,683百万円 D87. There are no borrowings with financial covenants E266, ¥85,000百万円 of overdraft and ¥50,000百万円 of commercial-paper capacity are wholly undrawn E279, no forward exchange contracts were outstanding at either date E274, and the group is subject to no externally imposed capital requirement E271. Goodwill ¥54,405百万円 F970 plus intangibles ¥21,578百万円 F972 is ¥75,983百万円 D92 — 5.84% of assets D94 and 7.45% of parent-attributable equity D93. This is not an acquired balance sheet.

Cash conversion and what was done with the cash. Operating cash flow ¥278,446百万円 F65 against 当期利益 of ¥251,451百万円 F15 is 1.1074× as reported D107 and 1.2228× ex-one-off D108; five-year cumulative operating cash flow is ¥1,128,245百万円 D100 against cumulative investing outflow of ¥153,380百万円 D101 — 13.59% D103 — for cumulative free cash flow of ¥974,865百万円 D102. Capital expenditure was ¥65,711百万円 on the accrual basis F1336 and ¥56,581百万円 in cash F1226 against depreciation and amortisation of ¥58,218百万円 F1064, a ratio of 1.1287× D105 after 1.2541× the prior year D106. R&D was ¥36,291百万円 F440, 3.8% of revenue F5. Every planned facility in the table is funded from own funds E118.

And the capital came back. Buybacks of ¥171,970百万円 F1254 plus dividends of ¥81,903百万円 F1243 came to ¥253,873百万円 D111 — 91.17% of operating cash flow D112, 100.31% of reported profit attributable D113, and 110.70% of ex-one-off profit attributable D114. The company returned more than it earned and net cash still grew D86D87. 発行済株式総数 went 369,702,020 → 334,838,020 F106F595, down 8.46% in the window D120 and 9.43% including the 3,576,300 shares cancelled on 15 May 2026 D121D122E301 — and every single movement in that history is a cancellation, never an issue E121. Dividend policy is a progressive dividend against a 40% payout guide F714, delivered at 39.7% consolidated F717, decided by the board rather than the meeting E127.

The price the cannibal ate at. Three buyback authorisations totalling ¥250,000百万円 were live in the year D116, 68.78% executed by the year end D117. The executions: ¥99,999,084,000 over 4,708,300 shares F679F678 — about ¥21,239 apiece; ¥71,948,014,500 over 2,600,800 F686F685 — about ¥27,664; and ¥28,050,977,500 over 975,500 after the year end F692F691 — about ¥28,756. Against ex-one-off earnings of ¥674.14 D6 those are 31× to 43×, near a fiscal-year high of ¥29,590 F190 against a low of ¥14,345 F195. The stated rationale is restraining the expansion of capital to improve capital efficiency E126 and reducing the share count E301 — a capital-structure rationale. Neither of the two places the filing states a rationale ties a repurchase to a view on value E126E301.

What management is paid on. The annual bonus runs on revenue, profit attributable and earnings per share, 80% quantitative, across a 0–200% band E178; the performance share unit runs three years on revenue, EPS, ROE and an ESG indicator E179. The FY2026 bonus gate was struck on an EPS target of ¥592.43 against an actual of ¥743.93 F837F838 — the reported figure, not the ¥674.14 ex-one-off D6. The one-off moved the pay. Total officer remuneration is ¥1,291百万円 on the governance table F884 against ¥1,190百万円 for key management in the IFRS notes F2202, the two on different bases in one filing. The CEO holds 7,300 shares E155 against ¥562百万円 of remuneration as executive officer in one year F890; the CFO holds 12,800 E156; one outside director holds 400 E150 and another holds none E154.

And there is no capital-efficiency target of any kind. The only stated objective management indicator is SVA, introduced on the reasoning that value is created when profit exceeds the cost of capital — and it is never quantified; no cost-of-capital figure appears; the price-to-book term does not appear; the only ROE number in the document, 20.0%, is a performance condition inside a share-unit plan E334E46E183. The one capital-cost disclosure in the whole filing is involuntary: the 6.3% to 15.5% pre-tax weighted-average cost of capital band used in impairment testing F1537F1538.

Where the stamp sits. ¥23,840 F2428 against a fiscal-year range of ¥29,590 to ¥14,345 F190F195. The filing's own printed price-earnings series runs 31.4 / 31.0 / 36.4 / 28.9 / 35.7× F56F57F58F59F60, a five-year band running from a bottom of 28.9× F59 to a top of 36.4× F58. Stamp-day turnover was ¥26,947,106.500千円 F2432, 0.34% of the net market capitalisation D11. The company prints its own scorecard: five-year total shareholder return 210.2 against a dividend-inclusive TOPIX at 202.2 F180F185E128 — 16.02% a year against 15.12%, 0.90 percentage points a year of excess, over a window containing no revenue decline, 20–25% returns on equity and a 9.43% reduction in the share count D122.

The five lenses

Buffett — watch, buy below ¥7,700

Let me start where I always start: with the business.

There are two shops under this roof. The first sells things that go in front of your eyes or inside them — eyeglass lenses, contact lenses, endoscopes, the little plastic lens the surgeon puts in after a cataract E35. That shop did ¥590,680百万円 of outside business F395, 62.32% of the whole D65. People's eyes get worse. They buy again. Children stare at screens and the world keeps getting older, and management says so plainly E49. This company has been selling eyeglass lenses since 1962 E11 and contact lenses since 1972 E12. That is a toll bridge, and toll bridges are what we like.

The second shop sells the glass plate a chipmaker prints his pattern on, the glass disc your data is written to, and precision optical lenses E36. ¥354,751百万円 F401 — only 37.43% of revenue D66 but 58.96% of segment profit D68, earned at a 54.14% margin D62 that barely budged from the prior year's 54.68% D63. When a margin like that holds steady, it is telling you nobody else can easily do the work.

Now the record. Revenue went from ¥661,466百万円 to ¥947,749百万円 F1F5, 9.41% a year compounded D59. Return on equity printed 22.1, 20.8, 20.3, 20.8, then 25.4% F51F52F53F54F55. That last figure is dressed up. Inside it sits a ¥23,538百万円 re-measurement of a buy-out obligation on a China joint venture F2044 and ¥7,402百万円 of gains on a business they sold D13 — ¥30,940百万円 pre-tax D14 — and it is the re-measurement, not the disposal gains, that management itself labels 一過性 E91. Take it all out and the year was 23.00% D27 against a four-year average of 21.00% D33.

Notice something: the record survives the scrubbing. Most of the time when you scrub a headline year, the shine comes off with it. Here it doesn't. Pre-tax margin ex-one-off is 31.31% D47, right inside the 29.83% to 31.85% band the four prior years occupied D43D42; net margin ex-one-off 24.03% D53, inside its band too D48. Somebody has been minding the store for a long time.

And then the fact that made me sit up. This company is the mirror image of the cash hoarders. It paid out ¥171,970百万円 of buybacks F1254 and ¥81,903百万円 of dividends F1243 — ¥253,873百万円 D111, 110.70% of what it earned after scrubbing D114 and 91.17% of the cash the business generated D112. It returned more than it earned and its net cash still grew, from ¥496,683百万円 to ¥531,851百万円 D87D86. Shares outstanding went 369,702,020 to 334,838,020 F106F595, down 9.43% D122, and every single movement in that capital history is a cancellation, never an issue E121. Equity ratio 78.4% F50. No covenants on any borrowing E266. You could take that balance sheet through anything.

So: a wonderful business. Now let's count.

Owner earnings. Take profit to owners scrubbed of the one-offs, ¥229,342百万円 D22, add back depreciation and amortisation of ¥58,218百万円 F1064, and subtract what it costs to stand still. I triangulated three ways, as one must. Cash spent on plant was ¥56,581百万円 F1226; total capital spending on the accrual basis ¥65,711百万円 F1336; and five years show only ¥153,380百万円 of investing outflow D101 against ¥286,283百万円 of revenue added F1F5. Call owner earnings ¥225,000百万円 F1064D22. Against a market value of ¥7,998,053百万円 D2, that is an owner's yield of about 2.8% D2. Free cash flow yield 3.39% D110. Dividend yield 1.24% D8.

Here is the plainest way I can put it. Net tangible assets are equity of ¥1,020,460百万円 F1012 less goodwill of ¥54,405百万円 F970 less intangibles of ¥21,578百万円 F972 — ¥944,477百万円, about ¥2,815 a share on 335,488,796 shares D1. You are paying ¥23,840 F2428. So ¥21,025 of every ¥23,840 — 88% of the price — is economic goodwill, and against ¥229,342百万円 of scrubbed earnings D22 that premium takes roughly thirty-one years to earn back, undiscounted. I have paid up for goodwill before and been glad of it. Never on that arithmetic.

Now the other side, because a man who only lists virtues is selling something. The segment assets are not disclosed, in either year D79. So I cannot compute return on capital for either shop — only for the company whole, where it is 43.44% ex-cash D84. The moat is asserted where I can see it and unmeasurable where it lives, and one shop carries 59% of the profit D68. Three "none" declarations are not true as a shopkeeper would read them E331E332E333. And nobody is eating their own cooking: the chief executive holds 7,300 shares E155 against ¥562百万円 of pay in one year F890; the buybacks were executed at about ¥21,239, ¥27,664 and ¥28,756 F679F686F692, near 7.8 times book D4, justified as restraining the expansion of capital to improve capital efficiency E126 — a ratio argument, not a value argument — while the bonus turns on earnings per share E178 and the long-term award on earnings per share and return on equity E179, the two numbers a buyback lifts by arithmetic alone.

The verdict, plainly. On my conservative appraisal — operating owner earnings capitalised with no growth credited at the 6.3% floor of the company's own disclosed pre-tax cost-of-capital band F1537, with the cash haircut for ¥475,474百万円 of undistributed subsidiary earnings that carry no recognised deferred tax E256 — this is worth ¥11,600 to ¥18,000 a share, and on the friendliest reading I can defend, twenty-five times operating owner earnings with the most generous maintenance estimate, about ¥20,650 — still short of ¥23,840 F2428. I want a third off the bottom of that range, which is ¥7,700. At that price you buy the operating business for under ten times its ¥218,583百万円 of scrubbed after-tax operating profit D76, with a fortress behind it. It has not been offered there and may never be. There are no called strikes.

What a student should take from this: scrub the headline year before you admire it, then check whether what's left still looks good — here it did, which is exactly why this company earns the file it gets. But quality settles whether you may buy, never at what price; when 88% of the price sits above net tangible assets F1012F970F972, the payback arithmetic tells you the truth faster than any story about moats. And when the filing withholds the one number that would let you test the moat where it lives — segment returns D79 — treat the moat as asserted rather than proved, and widen the margin you demand for the part you cannot see.

Munger — watch, buy below ¥14,000

Start by inverting, because that is the only honest way to open a memo on a business that looks this good.

Both sides look like textbook moats — a regulatory-and-quality barrier in Life Care E235, technology leadership inside a deliberately narrow niche in Information Technology E236E103. The figures back it up: 情報・通信 at 54.14% D62, essentially flat on 54.68% D63; ライフケア up from 16.40% to 21.93% D61D60; revenue compounding 9.41% D59; net cash of ¥531,851百万円 D86 with no covenant, no maturity wall and no forced-seller trigger anywhere E266E271. This company could survive the quotation halving without blinking. On the numbers alone it clears most of the quality gates cleanly, and I want that said before anything else.

Now invert further, because that is the job, not a formality. Read the filing against itself and it contradicts its own risk register in three separate places. No significant contracts E110 — against an alliance with Dai Nippon Printing, four live joint ventures and a customer at 11.75% of sales E331F436. No material contingent liabilities E298 — against three transfer-pricing assessments spanning nearly two decades, ¥20,460百万円 sitting on the balance sheet as a prepaid tax asset the company may never recover F1679, an appeal filed with the Tokyo High Court on 24 June 2025 E258, eleven months before this filing went in, and — the detail that should stop you — the auditor's sole consolidated Key Audit Matter is precisely that uncertainty E332E322. Business combinations omitted as immaterial E297 — against goodwill added by acquisition in the same year and a goodwill impairment on a deal that missed its own plan E333F1473F1513. Three "nones" is not a typo. It is a pattern, and my own rule on complexity is blunt: more than one flag shifts the burden of proof against the company.

This year's earnings carry a related wrinkle. Printed return on equity is 25.4% F55, but ¥30,940百万円 of the pre-tax gain behind it is one-off D14, mostly a remeasurement of a China joint-venture buyout liability that came in under estimate E91. Strip it and return on equity is 23.00% against a remarkably flat four-year mean of 21.00% D27D33 — a real improvement, but roughly half the size the printed number implies D35D34. And the company states no corporate return target of any kind: SVA is never quantified, and the 20.0% ROE that recurs three grant-years running is a share-plan hurdle E334, cleared on the one grant whose actual is disclosed — 22.2% against 20.0% on the FY2023 award E183 — with the FY2024 and FY2025 awards still in flight E185. That is exactly the sort of thing an analyst who already likes the story walks past without reading twice.

None of this makes HOYA a bad business, and I want to be precise about what it does to the case. The moat mechanisms are named, not inferred backward from returns. Incentive design is reasonably sober — clawback and malus including on restatement E176, no repricing history, a genuinely thin corporate centre of three executive officers over a heavily delegated divisional structure E136E194. The balance sheet buys all the patience in the world. But I cannot verify the moat where competition actually happens: no segment assets are disclosed, so segment return on invested capital is not computable, and I am left asserting quality at the consolidated level while the disaggregated evidence stays locked away D79D84. And the price is not fair: 35.36× ex-one-off earnings D7, 7.84× book D4, on a business now returning more cash than it earns — 110.70% of ex-one-off profit D114 — rather than compounding retained capital at its own 43.44% ex-cash return D84. The reinvestment engine that would justify paying up for growth is mostly not running. What you buy at this multiple is the dividend, the buyback-driven shrink, and a bet that the margin story holds.

So: watch, buy below ¥14,000 — and let me be plain about what that number is, because a reader would otherwise take it for an appraisal. It is not. Under lens isolation I had no opportunity-cost comparator available at all, so my required-alternative test was unanswerable and I substituted the napkin: ¥674.14 of ex-one-off earnings D6 at a fair-not-great 20× to 22× gives ¥13,500 to ¥14,800, and I publish ¥14,000 from the middle of it. That is gravity math, not an opportunity cost, and my frontmatter records it as such. The economics clear on their own; the character gate does not — not because I have found fraud, but because a filing that misstates its own risk register three times in one document forfeits the benefit of the doubt on everything else it declares clean, and a price this rich leaves no room to be wrong about it.

What a student should take from this: when a filing declares "none" three times and you can find the "something" each time inside the same forty pages, do not average that finding against the good numbers — it is evidence about the filer's candour, and candour is a cost you pay before you get to price the business at all. A reported return built substantially from a one-off remeasurement should always prompt the question of what is left standing once the balance-sheet miracle of the year reverses. And when a company will not show you segment assets, remember that a moat you cannot measure at the level where competition happens is a moat you are taking on faith, however good the consolidated arithmetic looks.

Pabrai — pass, buy below ¥6,000

Let me start where I always start, which is with the tail I lose on.

I am asked to pay ¥23,840 F2428. What do I own if everything goes wrong? Book value per share is ¥3,041.71 F35 — the stock is 7.84 times book D4, so the whole equity account is 12.8% of the price. Now strip that book the way my own item makes me: goodwill and intangibles of ¥75,983百万円 D92 go to zero. What a liquidator gets is the cash, and cash net of every yen of debt is ¥531,851百万円 D86 — ¥1,585.30 a share D88, 6.65% of what I am asked to pay D89. Haircut the ¥132,482百万円 of inventories F984 and the ¥209,613百万円 of receivables F986 and you might reach a tenth of the price. Ninety percent of my purchase price is a bet that this business keeps earning what it earns.

That is a fail on my first item, and my first item is the whole game. Heads I win, tails I lose eighty-seven cents on the dollar is not a trade I make. A P1 failure routes to pass on my checklist no matter what sits above it — and what sits above it here is very good.

Because there is almost nothing on the wrong side of this balance sheet. Interest-bearing debt of ¥42,241百万円 F1735 against ¥574,092百万円 of cash F994; of that debt ¥25,628百万円 is leases D82, so real borrowings are ¥16,613百万円 D83 at 2.01% F1723. And the sentence I read three times: 「財務制限条項付の借入金はありません」 — no borrowings carry financial covenants E266. Delta Financial was solvent on paper and died anyway because it could not roll its funding. This company never needs a lender's goodwill.

The business fits in five sentences, which is the test. HOYA melts optical glass and makes two families of thing from it: things that go in front of eyes, and things that go into chips and disk drives. The eye half sold ¥590,680百万円 F395 at 21.93% D60, up from 16.40% D61. The chip half sold ¥354,751百万円 F401 at 54.14% D62. One caution, because this filer reports under IFRS: there is no 営業利益 and no 経常利益 here, and セグメント利益 is pre-tax profit F1328 — so those margins sit after finance income and after the one-offs and are not operating margins. Where I need an operating number I use the ledger's one explicit proxy, ¥284,837百万円 ex-one-off D38, a 30.05% proxy margin D40 against 28.24% D41. I invent no line the filing does not print.

So: a near-monopoly, a fortress, and an owner-friendly policy — ¥253,873百万円 returned D111, 110.70% of ex-one-off profit D114, the count down 9.43% D122. You are genuinely paid to wait. So why am I passing?

First, the price. Strip the one-offs D14 and earnings are ¥674.14 D6, not ¥743.93 F40. That is 35.36 times D7. I do not pay fancy prices for great businesses. The Microsoft chart from 1999 to 2015, dividends included, returned zero, and it was a wonderful company the whole way.

Second, the payback. Enterprise value is ¥7,466,202百万円 D90. Five years of actual free cash flow came to ¥974,865百万円 D102 — thirteen percent of today's enterprise value. I hunt for hidden price-earnings ratios of one. Thirteen percent recovered in five years is a payback of roughly thirty-eight years on that five-year average D90D102 — and even on the FY2026 run-rate of ¥270,860百万円 D99, the best of the five, it is 27.6 years.

Third — and this is the one that would actually hurt me — the majority of the profit sits in an industry that changes fast. Information Technology is 37.43% of revenue D66 but 58.96% of segment profit D68, and this year's growth was three simultaneous highs: extreme-ultraviolet and deep-ultraviolet mask blanks, a Chinese flat-panel plant ramping, and data-centre demand for the larger hard-disk format E94. In the same breath the filing tells me the smaller format fell substantially E94 — a product line HOYA has made since 1991 E16, dying while I read. And I cannot measure the moat where it lives: segment assets are not disclosed D79D84.

Two more things. Seagate is 11.75% of sales F436, up from 10.71% F434, sitting inside the high-margin segment E241 — while the same filing says there are no significant contracts at all E110. And the buybacks I just praised were executed at about ¥21,239, ¥27,664 and ¥28,756 F679F686F692 — thirty-one to forty-three times ex-one-off earnings D6. A cannibal eating itself at forty times is not compounding my ownership; it is handing value to the sellers. I want the Singleton pattern — issue high, buy low. This is buying high, near a five-year high of ¥29,590 F190.

My crayon, then. Fade earnings to the four-year mean return on equity of 21.00% D33 on average parent-attributable equity of ¥997,241.5百万円 D25; that is ¥209,421百万円, or about ¥615.58 on 340,199千株 weighted average F2091. Strip the ¥13,153百万円 of interest income F2069, about ¥29.67 a share after the group's own 23.26% rate D18 — call it ¥585.91 of operating earnings. Fifteen to twenty times that, plus the ¥1,585.30 of net cash D88, gives ¥10,374 to ¥13,303. Half the midpoint is ¥6,000. At ¥6,000 this is 8.9 times ex-one-off earnings D6, under two times book F35, with net cash covering a quarter of the price D88. That is a Dhandho setup. It is also 75% below today's price F2428, and I do not expect to see it. Most companies are a pass. This is one, and that is the system working.

What a student should take from this: a fortress balance sheet is not a floor — the floor is only as deep as the assets you can actually mark, and at 7.84 times book D4 those assets cover about a tenth of the price no matter how much net cash sits inside them D89. When a filing hands you a 54% segment margin D62 but no segment assets D79, it has shown you the reward and withheld the risk; that asymmetry of disclosure belongs in your sizing, not in your enthusiasm. And always check the price the cannibal eats at: a shrinking share count D122 is a gift only when the shares are bought below value, and at thirty-one times earnings on the cheapest authorisation — ¥99,999,084,000 F679 over 4,708,300 shares F678 — rising to forty-three on the dearest — ¥28,050,977,500 F692 over 975,500 F691 — against ¥674.14 D6, it is a transfer to the sellers.

Li Lu — too-hard (arithmetic published at ¥10,500; explicitly not a trigger)

I have spent this study on a company that is easy to admire and hard to know, and the distance between those two things is the whole of my answer.

Start with what HOYA is. An optical-glass works from 1941 E8 that eighty-five years later still makes the precise piece of glass somebody else's product cannot exist without. Management calls this 「小さな池の大きな魚」 E103, and the phrase is not marketing — it describes a firm that has repeatedly chosen narrow, technically defended positions and then walked away from the ones that stopped being defensible: crystal in 2009, hard-disk media to Western Digital in 2010, PENTAX imaging to Ricoh in 2011 E23, the speech-synthesis business in October 2025 E300, the contact-lens retail arm carved out in April 2026 E313. That is a seventeen-year trail, 2009 to 2026, left by an operator who knows what he owns.

The numbers behave the way that history predicts. Revenue ¥661,466百万円 → ¥947,749百万円 F1F5. Return on parent-attributable equity 22.1, 20.8, 20.3, 20.8 and 25.4% F51F52F53F54F55 — a four-year mean of 21.00 D33 inside a range of 1.8 points, flatter than most businesses manage across a single cycle. Net cash ¥531,851百万円 D86, no financial covenants E266, both facilities undrawn E279. Group return on invested capital excluding cash, normalised, 43.44% D84. And a management that returns what it cannot use: ¥253,873百万円 D111, 110.70% of normalised profit D114, the share count falling in every year of the window — 369,702,020 F106, 356,960,520 F107, 350,958,720 F108, 345,859,220 F109, 338,414,320 F110 — with every dated movement a cancellation E121.

That deserves respect. The Japanese failure mode I have watched for two decades is the cash pile that never comes home; HOYA does not have that disease. Nor the other one — no parent E317, no controller, a register 61.03% foreign F616 whose four largest names are custodian and trustee nominees E123 and whose top ten hold 41.57% between them F642, cross-holdings cut to ¥321百万円 unlisted and ¥98百万円 listed E189, both earmarked for sale E190, and no related-party transactions in either year E295. Structurally, on every ownership test I apply to this company, the register comes back clean.

So why do I decline?

Because when I ask my own question — could I own the whole of this for ten years and be content — the single fact that would decide it is not in the record. Two-fifths of revenue and three-fifths of profit come from 情報・通信 D66D68, at a segment margin of 54.14% D62 that was 54.68% the year before D63. A margin like that is either a moat or a mirage, and the way to tell is to know what capital sits behind it — what a competitor would have to spend to take it. HOYA discloses no segment assets, in either year, in any form D79; segment return on invested capital, turnover and capital intensity therefore do not exist D84. There is no market share, no capacity figure, and no order book at all, because production is to forecast and 受注実績 is omitted outright E101. On the variable that decides the next decade I am reduced to inferring a moat from a margin, and inferring a moat from a margin is what a man does when he does not know.

I would carry that gap if I could trust that what is withheld is genuinely small. I cannot, and that is the second reason. Three times in this one document the company declares a category immaterial while describing material things inside it E331E332E333. None of the amounts is large. All three are the same act: the company judging materiality, and judging it as a careful reader would not. Having watched a firm do that three times in one filing, I can no longer treat its silences as evidence of smallness — which is precisely what accepting the segment-asset gap would require of me.

Then the price, which settles it even for a reader more forgiving than I am. Strip the ¥30,940百万円 of one-offs D14 and the printed 25.4% return becomes 23.00% F55D27, still excellent, while 32.05 times earnings D5 becomes 35.36 D7. 7.84 times book D4. A free-cash-flow yield of 3.39% D110. Buy a business earning 21% on its equity D33 at 7.84 times that equity D4 and you own, on your own capital, something close to 2.7% — everything else must arrive as growth, and the exit multiple must hold for a decade against a printed five-year band whose top is 36.4× F58 and whose bottom is 28.9× F59.

One further thing troubles me as an owner. This company cannot reinvest at its own returns. Capital expenditure was ¥65,711百万円 F1336 against ¥278,446百万円 of operating cash flow F65; the remainder goes back to shareholders D114. A 43% return on invested capital D84 deployed across a quarter of the cash is not a 43% compounder — it is a 9% grower D59 paying out the rest. That is honest allocation and I prefer it to hoarding. It is not the machine the multiple is pricing.

So: too-hard. Not because the business is poor — on the evidence in this ledger it is an excellent one — but because the thing I must know I cannot know, because the company has demonstrated it will call material things immaterial, and because the price leaves nothing at all for being wrong. And because a student should always be made to see the arithmetic, here it is: normalised owner earnings of ¥209,421百万円 — the 21.00% four-year mean return D33 on ¥997,241.5百万円 of average parent-attributable equity D25 — requiring a 12% ten-year owner return with no re-rating, at a durable growth rate of 6% against a delivered 9.41% D59 and payout near 100% D114, gives a required entry yield of about 6% and an entry multiple near 16.7×, or roughly ¥3,497,000百万円 over 335,488,796 shares D1 — about ¥10,400. Half the ¥23,840 stamp F2428 is ¥11,920. I take the lower and publish ¥10,500, 44% of the stamp and below the fiscal-year low of ¥14,345 F195. This is not a trigger. A too-hard verdict is a knowledge problem, and ¥10,500 only cures the price problem. It becomes live solely on the disclosure condition in my falsifier.

What a student should take from this: a high margin is not a moat until you know the capital standing behind it; when a filer gives you segment revenue and segment profit but no segment assets D79, the number that decides defensibility simply does not exist, and no amount of quality elsewhere substitutes for it. When a company declares a category immaterial and the same document describes material items inside that category, its materiality judgement has become a fact about the company rather than about the facts — stop reading its silences as evidence of smallness. And keep the two reasons to decline apart: "too expensive" is a price problem that a lower price cures, while "I cannot see the economics" is a knowledge problem that no price cures. Always say which one you have.

Claude — watch, implied buy-below ¥8,500, band ¥6,800–¥10,600

I registered my priors figures-blind, and this run turned on two of them landing and one band missing badly.

The inversion. Figures-blind I registered that the naive frame — semiconductors cyclical, medical durable — was probably inverted, and put P(Life Care proves the weaker pillar) ≈ 0.45, far above the consensus framing and deliberately so. The ledger says 0.45 was far too timid; on the evidence I had — every impairment in both disclosed years sitting in Life Care E248E249, goodwill concentrated in eyeglass lenses and medical devices E254, Life Care's growth credited to promotion, bolt-on M&A and new stores E104, and Life Care being the pillar whose own risk section admits the speed of price decline may outrun its ability to absorb it E82 — it should have been about 0.75. 情報・通信 earns 54.14% D62; ライフケア earns 21.93% printed D60, and about 17.94% once the one-off is located — (¥129,531百万円 F398 less ¥23,538百万円 F2044) over ¥590,683百万円 F1310. I hold that location at 0.90; the filing never states it E335.

The buyout hypothesis, registered at ~0.50 before any number. I asked, figures-blind, whether a remeasurement gain on a performance-linked buyout obligation means the expected purchase price fell — a bad operating signal booked as a good financial one — and I registered the inference rule in advance so §2 could not rationalise it afterwards. Management answers it verbatim: the buyout amount was carried as a long-term financial liability, and because the actual acquisition amount fell below the initial estimate owing to changes in the market environment, the difference was recorded as a one-off gain, and profit increased substantially E91. HOYA's largest single item of FY2026 income is the news that a Chinese asset it had contracted to buy got cheaper. The balance sheet confirms the release, ¥23,793百万円 → ¥614百万円 F1019F1020. And there is an unlinked mirror I hold at 0.75: ¥17,223百万円 charged to other components of equity for changes in interests in subsidiaries that remain controlled, against ¥17,176百万円 added to non-controlling interests E270. If those are the same event, the net addition to the parent's equity is far smaller than ¥23,538百万円 F2044. The filing cross-references none of it E335.

And the misses, at the same volume. I predicted the Information Technology segment margin at 22–38%, central 29% — the truth is 54.14% D62, 25.14 points above my central and 16.14 above the top of my band; that is my worst miss and it invalidates the "barbell dilution" reasoning I built the band on. I predicted the equity ratio at 60–75% against an actual 78.44% D91; return on equity at 14–22% against 20.3–25.4% in every year F53F55 — a miss made internally inconsistent by predicting a lower equity ratio at the same time; goodwill and intangibles above 15% of assets at 0.60 against an actual 5.84% D94; and distribution below 100% of profit at 0.72 against 100.31% reported and 110.70% ex-one-off D113D114. What I got right: Life Care's revenue share at 62.32% D65, the one-off at 5–25% of pre-tax profit D14F10, the record surviving its removal D20D27, net cash at 5–15% of capitalisation D89, price-earnings 25–45× D5D7, price-book 4–9× D4, the payout F717, the revenue CAGR D59, the segment table carrying only two years, and the segment note attributing none of the one-off E335.

The arithmetic, then. Owner earnings — operating cash flow ¥278,446百万円 F65 less maintenance capex anchored on depreciation ¥58,218百万円 F1064 — are ¥220,228百万円, a 2.75% private-owner yield on ¥7,998,053百万円 D2. The only company-specific capital cost in the entire archive is the impairment note's 6.3%–15.5% pre-tax band F1537F1538; even the low end is more than double that yield. Reversing the price at an 8% discount rate and 2% terminal growth, enterprise value ¥7,466,202百万円 D90 against NOPAT ¥218,583百万円 D76 embeds roughly 11% annual growth for a decade before any fade — against 8.9% delivered on ex-one-off pre-tax profit D20F6 and 9.41% on revenue D59, both flattered by a currency that added ¥63,810百万円 of translation to equity in this year alone F1092. Total cash returned is 3.17% of the capitalisation D111D2 — the whole of what an owner is paid while waiting.

The decisive number is one the company prints itself. Five-year total shareholder return 210.2 against a dividend-inclusive TOPIX at 202.2 F180F185E128 — 16.02% a year against 15.12%, 0.90 percentage points a year of excess, for five years containing no revenue decline D54D55D56D57, 20–25% returns on equity F51F55, 43% return on capital ex-cash D84 and a 9.43% reduction in the share count D122. Near-flawless execution at this multiple bought index performance. That is the outside view's answer to whether the quality is already paid for.

The threshold, and what moved it. My own bear case — half of Seagate's ¥111,393百万円 lost at the 54.14% segment margin F435D62, Life Care reverting to its own observed 16.40% D61, and 情報・通信 conceding 1,000bp on the rest — gives bear pre-tax of about ¥227,501百万円 against ¥296,728百万円 ex-one-off D20, a 23% earnings haircut rather than a collapse, capitalised at 16.67× plus deployable net cash out of ¥531,851百万円 D86 less the ¥20,460百万円 of prepaid tax at risk F1679, over 335,488,796 shares D1: about ¥9,520. Three jury selves re-ran it and returned ¥6,771, ¥6,855 and ¥10,555 — and the third made the sharpest structural point of the six items, that at this multiple roughly 61% of the downside is de-rating rather than earnings, because 78% equity D91, 43% returns D84 and five years without a revenue decline make an earnings collapse hard to write. Divergence in this lens may only subtract, so I publish ¥8,500, with an honest band of ¥6,800 to ¥10,600. A second jury divergence, on stamp-price anchoring, upheld a dissent in part: one self enumerated ten free parameters spanning a ¥4,684–¥16,600 output envelope, so I argued the 8% explicitly against the ledger's own 6.3–15.5% band F1537F1538 rather than leaving it as convention, sourced every component of the cash haircut, and published an interval rather than a point.

Why watch and not pass. My pass overrides required two of four and got none; growth still creates value here at roughly 43% incremental returns D84, the short case dents rather than wins — the business is not impaired, the price is the problem — and the capital-allocation record is the strongest item in my file: cumulative operating cash flow ¥1,128,245百万円 D100 against cumulative investing outflow of ¥153,380百万円 D101, with ¥253,873百万円 returned in the latest year D111. Verdict: watch, implied buy-below ¥8,500. Review by 2027-06-30, when the 第89期 有価証券報告書 settles the segment margin that is the whole bet.

What a student should take from this: a "one-off gain" is a claim about the future, and its sign can be the opposite of its accounting sign. HOYA booked ¥23,538百万円 of income because the price it had agreed to pay for a Chinese joint venture fell F2044E91. Read that as economics rather than accounting and it is an impairment of an option HOYA held, disclosed as profit. Nothing was hidden; nothing was cross-referenced either E335. Second: where a company's reporting boundary sits determines what you can and cannot know. This group knows the answer at a finer grain — cash-generating units are defined by product line — and reports it at a coarser one E247E238. The correct response is a confidence cap, not a haircut: missing disclosure is not evidence about the level of returns, and treating it as bad news is a lazy substitute for saying you do not know. Third: excellent businesses at excellent prices are the interesting case; excellent businesses at fair prices are the common one. When a business is this legible, its quality is in the price by the time you can see it F180F185. The work that pays is not confirming the quality; it is finding the number at which the quality is worth owning, publishing it, and waiting.

Synthesis

Where the five lenses agree

First, the business is genuinely excellent, and the excellence survives normalisation. Ex-one-off return on equity 23.00% D27 against a four-year mean of 21.00% D33; return on invested capital ex-cash 43.44% D84; pre-tax margin ex-one-off 31.31% D47, back inside the prior four-year band D43D42; five-year cash conversion of 116% — cumulative operating cash flow of ¥1,128,245百万円 D100 against cumulative 当期利益 of ¥969,877百万円 (165,322 F11 + 168,788 F12 + 182,566 F13 + 201,750 F14 + 251,451 F15); net cash ¥531,851百万円 D86 with no covenanted borrowing E266 on a 78.44% equity ratio D91. Every lens, including the one that declined and the one that passed, wrote that paragraph.

Second, the capital allocation is exemplary and nobody disputes it. ¥253,873百万円 returned D111, 110.70% of ex-one-off profit D114, net cash still growing D86D87, the share count down 9.43% with every movement a cancellation D122E121, cumulative investing outflow only 13.59% of cumulative operating cash flow D103. This is the mirror image of the hoarding failure mode.

Third, the FY2026 headline is one-off-inflated, and all five normalise to the same figures. ¥30,940百万円 pre-tax D14, ¥23,743百万円 after tax D19, earnings per share ¥674.14 rather than ¥743.93 D6F40, the multiple 35.36× rather than 32.05× D7D5.

Fourth, the moat cannot be measured where it lives. No segment assets in either year D79, so no segment return on invested capital exists for the business carrying 58.96% of segment profit D68. Five lenses found this; four made it load-bearing; one made it decisive.

Fifth, the disclosure has a consistent tilt, and all five name the same four instances. The three "none" declarations E331E332E333 — and the absence that binds them together: no ROE target, no PBR target, no cost-of-capital figure anywhere in the document, the only stated objective indicator never quantified, and the one printed 20.0% ROE existing solely as a share-plan performance condition E334E46E183. The only capital-cost number in the whole filing is the 6.3%–15.5% band disclosed involuntarily inside an impairment note F1537F1538 — and two lenses used exactly that band as their discount rate, because there was nothing else.

Where they diverge — a staged exchange

---

On the price: five named prices between ¥6,000 and ¥14,000 read off one ledger against a ¥23,840 stamp F2428 — four of them issued as thresholds, the Claude lens's ¥8,500 implied rather than declared, and Li Lu's ¥10,500 published and expressly refused as one — and the spread is a fact about the operations applied, not about the facts.

BUFFETT (B91, the private-owner appraisal; B93, the required discount; B101, the goodwill premium). My object is an appraisal and then a discount from the bottom of it, and I will name every input. Operating owner earnings — scrubbed profit of ¥229,342百万円 D22 plus ¥58,218百万円 of depreciation F1064 less maintenance triangulated between ¥56,581百万円 of cash spend F1226 and ¥65,711百万円 on the accrual basis F1336, then less after-tax interest income of ¥13,153百万円 F2069 at the group's own 23.26% rate D18 — capitalised with no growth credited at the 6.3% floor of the company's own disclosed pre-tax cost-of-capital band F1537, with the cash haircut because ¥475,474百万円 of undistributed subsidiary earnings carry no recognised deferred tax E256. That is ¥11,600 to ¥18,000 a share. A third off the bottom is ¥7,700. Note what I did not do: I did not import a market interest rate, because the archive contains none, and I said so rather than inventing one.

BUFFETT, continued. And here is the line that actually decides my verdict, because it is multiple-independent. Net tangible assets are ¥1,020,460百万円 of equity F1012 less ¥54,405百万円 of goodwill F970 less ¥21,578百万円 of intangibles F972 — about ¥2,815 a share on 335,488,796 shares D1. At ¥23,840 F2428, 88% of the price is economic goodwill on a roughly thirty-one-year undiscounted payback against ¥229,342百万円 D22, and the whole-business owner-earnings yield is about 2.8% D2. I then took the friendliest reading I can defend — the most generous maintenance estimate at twenty-five times operating earnings — and it reaches only about ¥20,650, still under the stamp F2428. If the conclusion survives its own best case, the multiple is not what is doing the work.

MUNGER (M35 unanswerable; M36/M94, the gravity fallback; M88, the margin of safety). And I am the high number at ¥14,000, so let me say plainly what it is and is not. It is not an appraisal, and it is not an opportunity cost. My own item asks what else I could do with the money, and under lens isolation I had no comparator at all — no peer, no alternative, no rate. So I answered the item data-insufficient and substituted the napkin: ¥674.14 of ex-one-off earnings D6 at a fair-not-great 20× to 22× is ¥13,500 to ¥14,800, and I publish ¥14,000. My frontmatter records that substitution as the basis, because a reader who mistakes a gravity multiple for an opportunity-cost solve will misread everything downstream of it. What makes my number the highest and my verdict no looser than Warren's is that I am not pricing a business at all — I am pricing a re-rating to a fair multiple on a company whose filing misstated its own risk register three times E331E332E333.

PABRAI (P1, the floor; P53, the fifty-cent dollar; P60, not overpaying for the best). We are all reading the same ledger and I arrive at ¥6,000, and the difference is not arithmetic — it is that my first item fails outright and routes the verdict regardless of everything above it. Mark this balance sheet the way a liquidator would: goodwill and intangibles of ¥75,983百万円 D92 to zero, and what is left that I can actually mark is the cash net of every yen of debt, ¥531,851百万円 D86 — ¥1,585.30 a share, 6.65% of the price D88D89. Ninety percent of what I pay has no asset behind it. An unbreakable balance sheet at a premium multiple protects the enterprise, not the purchase price — and on my checklist a P1 failure is a pass, not a discount. My ¥6,000 is then half of a conservative value built by fading the return rather than the multiple: 21.00% mean return on equity D33 on ¥997,241.5百万円 of average equity D25, ex-interest-income, at fifteen to twenty times, plus the net cash D88 — ¥10,374 to ¥13,303, halved.

PABRAI, continued. And note where Charlie and I actually touch. My conservative value range and his napkin multiple are the same object. The top of my range — twenty times ¥585.91 of faded operating earnings D33D25 plus ¥1,585.30 of net cash D88 — sits at the bottom of his; we differ by rounding, not by reasoning. What separates our verdicts is not the number — it is that he stops at the number and I have a floor test in front of it.

LI LU (L1, the ten-year knowledge bar; L18, the fifty-cent dollar; L15, the closed-exchange owner test). Then I will state the fourth construction, and why it is published and refused in the same sentence. Mine is a required-return solve: normalised owner earnings of ¥209,421百万円 — the same 21.00% mean D33 on the same ¥997,241.5百万円 D25 Mohnish used — but demanding a 12% ten-year owner return with no re-rating, at 6% durable growth against a delivered 9.41% D59 and payout near 100% D114. That fixes the entry yield near 6% and the multiple near 16.7×, giving about ¥10,400 over 335,488,796 shares D1; half the stamp is ¥11,920 F2428; I take the lower and publish ¥10,500. And I refuse to issue it as a threshold. A buy-below is a promise that at that price I would act, and I would not, because the number rests on a multiple I cannot defend from any disclosure about why the margin exists.

BUFFETT. Then the spread is not disagreement about the ledger at all. Five of us used the same normalised earnings D22D6 and the same net cash D88 and produced ¥6,000, ¥7,700, ¥8,500, ¥10,500 and ¥14,000 because we applied five different operations to them: half of a faded value with a floor test in front of it, a third off a private-owner appraisal discounted at the company's own disclosed cost-of-capital floor F1537, a jury-marked bear case, a required-return solve, and a fair multiple on normalised earnings. That is worth writing down as a finding rather than smoothing into an average. 14,000 ÷ 6,000 is 2.33×, and the honest object is the interval and its five constructions, not any point inside it.

---

On what settles it for the two lenses that declined outright — and they decline for opposite reasons.

PABRAI. Mine is a price decline, and it is falsifiable at a number. Enterprise value ¥7,466,202百万円 D90 against five years of actual free cash flow of ¥974,865百万円 D102 is thirteen percent recovered — a payback near thirty-eight years on that five-year average, and 27.6 years even on the best single year of the five D99. The business is not the problem. At ¥6,000 every one of my unanswerables becomes tolerable; at ¥23,840 F2428 none of them is.

LI LU. And mine is a knowledge decline, which no price cures — that is the whole distinction and a student should not let it blur. Segment assets are not disclosed, in either year, in any form D79, so no segment return on invested capital exists for the business carrying 58.96% of segment profit D68 at a 54.14% margin D62. A margin like that is either a moat or a mirage, and the way to tell is to know what capital sits behind it. I would carry that gap if I could trust that what is withheld is small — and the same document declares a category immaterial three separate times while describing material things inside it E331E332E333. Having watched a firm do that three times, I can no longer read its silences as evidence of smallness, which is exactly what accepting the segment-asset gap would require of me. The two facts are survivable apart. Together they are the verdict.

MUNGER. And I held the same two facts and stopped at watch, which is the disagreement worth staging. My test is whether the decisive variable is nameable, not whether it is currently known. It is nameable — the 情報・通信 segment margin against 54.14% D62, the transfer-pricing appeal against the ¥20,460百万円 carried asset F1679, a third consecutive year of Life Care impairment E248E249 — and each resolves in a document with a date on it. Too-hard says the question cannot be posed. This one can. It cannot be answered today, and the correct response to that is a demanding price, not silence.

LI LU. Then we differ on what the shortfall costs, and my own item settles it against you. Priced at worst case, the unknown is not covered. Net cash is 6.65% of the price D89; the other 93% is paid for earning power whose capital base is undisclosed. And note the honesty test I applied to myself: I record that in establishing this profile's frontmatter format I opened the first twenty lines of this lens's own profile from an earlier, unrelated study. It carried no HOYA content and none of it bears on this analysis — but the assignment said no other company, and an instrument that hides its own scars is not measuring anything.

---

On the one-off, which is the study's principal forensic finding, and which no lens would have produced from the ledger alone.

CLAUDE (C6, the pre-registered priors; C112, the case-level lesson). I registered this figures-blind at about 0.50, before a single magnitude was visible, and I registered the inference rule with it so §2 could not rationalise the answer afterwards: if the liability is a formula-linked buyout amount, then a gain on remeasuring it means the expected purchase price fell — a bad operating signal booked as a good financial one. Management states it verbatim: because the actual acquisition amount fell below the initial estimate owing to changes in the market environment, the difference was recorded as a one-off gain and profit increased substantially E91. HOYA's largest single income item of the year is the news that an asset it had contracted to buy got cheaper — ¥23,538百万円 F2044, with その他の長期金融負債 falling ¥23,793百万円 to ¥614百万円 F1019F1020.

CLAUDE, continued. And the second-order consequence is worse than the item itself, because of how this filer reports. There is no 営業利益 and no 経常利益 here, and セグメント利益 is 税引前当期利益 F1328 — so the gain sits above the line in the pre-tax margin D46, the net margin D52, the printed return on equity F55, the operating proxy D36 and the segment margin of whichever segment the item lands in. The joint venture is a cataract intraocular-lens business E91, which is Life Care. Locate it there and ライフケア's printed 21.93% D60 becomes about 17.94% — (¥129,531百万円 F398 less ¥23,538百万円 F2044) over ¥590,683百万円 F1310 — a gain of roughly 1.5 points on the year rather than 5.5, while 情報・通信's 54.14% D62 is untouched by it. The segment note quantifies none of this and the filing cross-references none of it E335. I hold the location at 0.90 and publish it as an inference, not as a disclosed fact.

MUNGER. Then say what that does to the pay, because it is the sharpest small fact in the file. The FY2026 bonus gate was struck on reported earnings per share of ¥743.93 against a target of ¥592.43 F838F837 — not on ¥674.14 ex-one-off D6. A remeasurement gain arising because a Chinese asset got cheaper moved the compensation. The design is otherwise sober; the basis is not, and no one had to intend it for it to be true.

BUFFETT. And it lands on the buyback the same way. The count record is excellent — down 9.43%, every movement a cancellation, never an issue D122E121. The price record is not. ¥21,239, ¥27,664 and ¥28,756 apiece F679F686F692 is thirty-one to forty-three times ex-one-off earnings D6 and near 7.8 times book D4, on a rationale stated as restraining the expansion of capital to improve capital efficiency E126 — a ratio rationale, not a value rationale — set by people whose pay turns on the very ratios a repurchase mechanically lifts E178E179. The buyback passes on count and fails on price, and those are two different tests that a shrinking share count is very good at blurring.

---

On the Claude lens's own scorecard: the two blind calls that landed, and the band whose top the truth cleared by sixteen points.

CLAUDE (C1–C6, the outside view; C109, the surprise ledger). Report the hits and the misses at the same volume. The inversion landed and I under-weighted it. Figures-blind I put P(Life Care, not Information Technology, proves the weaker pillar) at 0.45, against a consensus framing that would have said the opposite — and I say now it should have been about 0.75, because the evidence I already had all pointed one way: every impairment in both disclosed years sits in Life Care E248E249, goodwill sits in eyeglass lenses and medical devices E254, Life Care's growth is credited to promotion and bolt-on M&A E104, and Life Care is the pillar whose own risk section admits the speed of price decline may outrun its ability to absorb it E82. The buyout hypothesis landed too, at 0.50, and management confirmed it in its own words E91. And the band that broke was the Information Technology margin: I registered 22–38%, central 29%, and the truth is 54.14% D62 — 25.14 points above my central and 16.14 above the top of my band. It invalidates the reasoning I built the band on. Two smaller misses run the same direction: the equity ratio at 60–75% against 78.44% D91, and return on equity at 14–22% against 20.3–25.4% in every year F53F55 — internally inconsistent, because I predicted a lower equity ratio and a lower return at the same time.

CLAUDE, on the jury. Two of six items diverged, and every divergence subtracted. On the bear case, three selves returned ¥6,771, ¥6,855 and ¥10,555 against my own ¥9,520, so the published threshold fell to ¥8,500 and became a band of ¥6,800 to ¥10,600. One of them supplied the structural insight I adopted: at this multiple roughly 61% of the downside is de-rating rather than earnings, because 78% equity D91, 43% returns D84 and five years without a revenue decline D54D55D56D57 make an earnings collapse hard to write. On anchoring, one self dissented and enumerated ten free parameters spanning a ¥4,684–¥16,600 envelope; the dissent was upheld in part, so the 8% required return is now argued explicitly against the ledger's own 6.3%–15.5% band F1537F1538 rather than left as convention. The four convergent items are recorded as "no divergence detected" and cited as support for nothing. Convergence among copies of one reasoner is not evidence.

LI LU. Then say what the convergence that did occur is evidence of. Three selves, handed the same rows, all concluded that too-hard was an over-application and watch the correct ceiling — and all three independently required that the one-off's segment location be stated aloud rather than left with the printed 21.93% D60 quoted as an operating result. That is not a finding about the jury. It is a finding about the disclosure, and it is the same finding I reached by a different route.

---

On the outside view's verdict, which is the one number none of us can argue with.

CLAUDE (C110, the base-rate candidate). The company prints its own scorecard: five-year total shareholder return 210.2 against a dividend-inclusive TOPIX at 202.2 F180F185E128 — 16.02% a year against 15.12%, 0.90 percentage points a year of excess, over five years containing no revenue decline D54D55D56D57, 20–25% returns on equity F51F55, 43% return on capital ex-cash D84, and a share count down 9.43% D122. That is not a forecast to be argued with. It is a completed experiment with a published result: near-flawless execution, bought at a franchise multiple, delivered index performance.

BUFFETT. Which is the whole of it, and it is why my verdict is watch and not pass. The business did everything asked of it. The owner got the index F180F185. When a business is this legible, its quality is in the price by the time you can see it — and the only work that pays is naming the number at which the quality is worth owning, publishing it, and waiting. Mine is a third off the bottom of my appraisal, and it has not been offered there. There are no called strikes.

The bet, located

Strip the agreement away and one falsifiable proposition is left. Is 情報・通信's 54.14% segment margin D62 a structural rent, or the utilisation peak of a fixed-cost base riding extreme-ultraviolet development spend, a Chinese flat-panel plant ramp and nearline hard-disk demand E94? Expressed as price: the bull needs roughly seventeen years of ~9% growth at these returns before fading; the bear needs eight to ten. That gap in excess-return duration is the entire disagreement, and it is what the 35.36× multiple buys D7. The bull mechanism is on the record — a margin flat at 54.14% against 54.68% D62D63 through 14.0% segment revenue growth, a decades-old alliance E18, a named technical roadmap E113, and 43.44% returns on capital ex-cash D84. The bear mechanism is equally on the record — 58.96% of segment profit on 37.43% of revenue D68D66, one customer at 11.75% of group sales inside that segment F436E241, production to forecast with no backlog E101, the 2.5-inch line already falling substantially E94, and segment depreciation up 50.7% on capacity additions F1292F1320 that will now work against the margin. It resolves in the 第89期 有価証券報告書 segment note, on a date, in a document anyone can read — and the filing's own comparative-disclosure record says the segment asset line still will not be there D79.

Prediction-vs-actual: VOID

This was an autonomous headless cycle. predictions.md carries void: no-human-prediction, and every prediction verdict is null by design — no practitioner was present at run time, and a blind call is never forged to fill the slot. No prediction-vs-actual scoring applies to this study, and it renders as void in docs/calibration.md, visibly distinct from a practitioner who declined. The five profile verdicts above still count in full for verdict accounting; only the human calibration half is skipped. The Claude lens's own figures-blind priors are a separate instrument and are scored, above and in its profile run — including the inversion call and the buyout hypothesis, both registered before the ledger and both confirmed and under-weighted, and the Information Technology margin band, whose top the actual 54.14% cleared by 16.14 points and whose central it cleared by 25.14 D62.

Self-distance note. The Claude lens holds one of the five verdicts compared above (watch, implied buy-below ¥8,500) and wrote this synthesis; it also built the reconciled figure table and the evidence ledger all five lenses consumed. That is an unusual concentration of authorship — the answerer, the ledger-builder and one of the five voters are the same system. Read the synthesis with that in mind. Four partial mitigations are on the record and should be weighed for what they are worth: the Claude lens ran figures-blind through its outside-view stage, so its priors were registered and pushed before any magnitude was visible and are scored against the ledger above — with the Information Technology margin band, the equity ratio, the return band, the intangibles share and the payout band all scored as failures; its jury of selves diverged on two of six items and both divergences moved the published threshold down and widened its band; the Munger run was executed on a different model, which is the only genuine cross-model check in this study; and no red team ran, because the verdict distribution did not reach the consensus threshold — so the consensus mitigation available in some studies is explicitly absent here.

Verdict accounting (fixed ex-ante)

• A buy-below-¥X verdict is price-falsifiable against the unadjusted stamp. Four lenses issued explicit thresholds — Pabrai ¥6,000, Buffett ¥7,700, Munger ¥14,000 (published as a range of ¥13,500–14,800), and the Claude lens an implied threshold of ¥8,500 with a published band of ¥6,800–¥10,600. Munger's figure is not an opportunity-cost solve: his own required-alternative item was answered data-insufficient under lens isolation and the number is the checklist's gravity-math fallback, roughly 20–22× ex-one-off earnings per share of ¥674.14 D6. Future scoring must read it as such, and his frontmatter records the substitution.
• Li Lu publishes ¥10,500 and explicitly refuses to issue it as a threshold. He computed it from a 12% required ten-year owner return on normalised owner earnings of ¥209,421百万円 D33D25 and cross-checked it against half the stamp F2428, then declined to make it a trigger, because too-hard is a statement that the decisive variable is absent from the record D79D84, not that the price is too high. It becomes live only on the disclosure condition in his falsifier — never on price alone.
• pass / watch / too-hard are recorded but unscored in any future review. Three verdicts here are watch, one is pass and one is too-hard. Every named price sits far below the ¥23,840 stamp F2428; the highest, ¥14,000, is 58.7% of it.
• The original verdict counts at its original stamp regardless of later corrections.
• On a stock split, reverse split, or consolidation, the buy-below threshold restates mechanically by the announced ratio (corporate-action disclosure cited); the stamp itself never restates. There is no split inside the window — every change in 発行済株式総数 across the five years is a cancellation of treasury shares and none is an issue E121. The denominator used throughout is 335,488,796 shares D1, being 338,414,320 issued F110 less 2,925,524 treasury F704, which reproduces the filing's own printed book value per share of ¥3,041.71 F35 to the sen. A further 3,576,300 shares were cancelled on 15 May 2026 E301, after the balance-sheet date and before the stamp; the study carries both bases, and the market capitalisation on the post-cancellation count is ¥7,912,794百万円 D123 against the ¥7,998,053百万円 used here D2.
• Every profitability figure in this thesis states whether it is reported or ex-one-off, and this study's headline numbers are the ex-one-off ones. As reported: ROE 25.38% D26, pre-tax margin 34.57% D46, net margin 26.53% D52, P/E 32.05× D5, ROIC ex-cash 48.16% D80. Ex-one-off: ROE 23.00% D27, pre-tax margin 31.31% D47, net margin 24.03% D53, P/E 35.36× D7, ROIC ex-cash 43.44% D84, EPS ¥674.14 D6. The two bases are never mixed silently.
• This group publishes no 経常利益 and no consolidated 営業利益. Every operating figure above is the ledger's constructed proxy D36D37D38 and is named a proxy at each direct use; the ex-one-off NOPAT built from that proxy D76 is labelled proxy-derived where it first appears and is not re-labelled at every later mention. The parent-only Japanese-GAAP statements are not comparable on any measure E205E216E307 and no figure in this thesis is drawn from them.
• Two years is not a segment history. The segment note exists for FY2025 and FY2026 only, and it carries no segment assets in either year D79. Every segment claim in this thesis — the 54.14% D62, the 21.93% D60, the ~17.94% relocation, the 58.96% profit mix D68 — rests on two observations and on one inference about where the one-off sits that the filing does not make E335.
• This is a VOID study for prediction-vs-actual purposes only; the verdict accounting above is unaffected.

Red team

No red team ran, and the reason is the rule rather than an oversight. The consensus red-team stage fires only when four or more lenses agree on a verdict class; a fresh adversary is then dispatched with the ledger alone to build the strongest opposing case. This study's distribution is three watch, one pass, one too-hard — the largest class reaches three of five, one short of the threshold — so no adversary was dispatched and there is no red-team.md in this study directory. The section is present rather than omitted, and says so plainly, because a reader comparing this thesis with one that carries a red team is entitled to know which mitigation was available and which was not.

What that costs the reader, stated rather than glossed. The strongest structural criticism this record can make of an AI-generated consensus is that it is one model family agreeing with itself; the red team exists to attack that. Here the mitigation is absent, and the substitutes are weaker and named in the self-distance note: the Munger run on a different model, the Claude lens's figures-blind pre-registration, and its within-lens jury, which diverged on two of six items and moved the published threshold down each time.

And the record should note what the dispersion itself supplies. The three declining-to-buy positions are not one argument in three voices. Pabrai passed on a floor test that fails at 6.65% of price D89; Li Lu declined on knowability, not value D79; Munger stopped at a candour gate over three "none" declarations E331E332E333. Each is, in effect, an adversary to the other two: Li Lu's knowledge objection would have Pabrai's ¥6,000 be unbuyable at any price, and Munger's answer to Li Lu — that a nameable variable resolving in a dated document is not a too-hard — is on the record above in the synthesis. That is not a substitute for a dispatched red team and it is not offered as one.

What would change our minds

Pre-registered falsifiers, per lens, taken from each profile run's falsifier: line. Future review notes score against these, not hindsight. Four of five converge on the same document — the 第89期 有価証券報告書 — and on the same two observables: the 情報・通信 segment margin and the segment asset line.

• Buffett (watch, buy below ¥7,700). The quality premise dies — and this becomes a pass rather than a watch — if the 情報・通信 segment margin D62D63 falls below 40% in two consecutive years, or if group return on equity normalised for one-offs falls below the 21.00% four-year mean D33 in two consecutive years. Either would show that the mask-blank and substrate position is being re-earned on worsening terms.
• Munger (watch, buy below ¥14,000). A third consecutive year of Life Care goodwill or cash-generating-unit impairment — after FY2025 E248 and FY2026 E249E250 — or an adverse Tokyo High Court ruling forcing write-off of a material share of the ¥20,460百万円 prepaid-tax asset F1679E258E259E322 would falsify the clean-quality read and move this from watch to pass. His stated kill conditions also include the 情報・通信 margin below about 45% for two straight years D62 and rising Seagate concentration alongside continued 2.5-inch decline F436E94.
• Pabrai (pass, buy below ¥6,000). The pass was the error if the 第91期 filing shows ex-one-off return on equity at or above the 23.00% of FY2026 D27 in each of FY2027–FY2029, the 情報・通信 segment margin still above 50% against the 54.14% of FY2026 D62, and shares issued down a further 5% or more from 334,838,020 F595 — and the shares have compounded above the TOPIX from the ¥23,840 stamp F2428 across that window. In that case P53's fifty-cent rule was the binding mistake rather than the protection.
• Li Lu (too-hard; ¥10,500 published, not a trigger). Flips to watch — and the ¥10,500 becomes live — if HOYA discloses segment assets or segment invested capital (IFRS 18 lands FY2028 E209) such that 情報・通信 return on invested capital is computable for two consecutive years and exceeds 30% while its segment margin holds above 45% D62. The decline is instead confirmed on business rather than knowledge grounds if the 情報・通信 segment margin falls below 45% D62 for two consecutive years, or ex-one-off return on equity falls below the 21.00% four-year mean D33 for two consecutive years. Note the asymmetry: a lower price alone changes nothing here.
• Claude (watch, implied buy-below ¥8,500). Three, each with a threshold and a named resolving document. (a) 情報・通信 segment margin below 45% in the 第89期 or 第90期 segment note flips watch to pass. (b) ライフケア segment margin below the 16.40% of FY2025 D61, ex any one-off, in the 第89期 filing flips watch to pass. (c) 情報・通信 segment margin at or above 50% with segment profit at or above the ¥192,325百万円 of FY2026 F404 in a year in which Seagate's disclosed share falls below 11.75% F436 converts the margin from cycle to rent and makes ¥8,500 a live buy-below. Review by 2027-06-30.

The single observable all five converge on is where the FY2027 情報・通信 segment margin settles against the 54.14% just printed D62 — together with whether the segment note finally carries an asset line D79. Both resolve in one document, on a date, and neither requires anyone's judgement.

What this taught the checklists

Queued for the next study (F2 revision proposals; see docs/process/evolution.md), attributed per lens. Checklist versions froze at this study's stamp commit; these take effect next time.

• Buffett — three. B42 needs an explicit rule for IFRS filers with no operating line. The item assumes a reported operating figure exists; this filer presents expenses by nature with no gross profit, no operating profit and no ordinary profit at all, and the ledger had to build a proxy D36. Propose adding to Good: "where the filing presents expenses by nature and reports no operating subtotal, state the operating measure used, its construction, and that it is a proxy — never import an operating margin from another accounting basis," because the parent-company statements in such filings carry an 営業利益 that is not comparable and an agent reaching for it produces a badly wrong margin E205. B92 should require a repatriation haircut on offshore cash. The Sanborn test takes excess cash at face value; for a group earning roughly 80% of revenue abroad E81 with ¥475,474百万円 of undistributed subsidiary earnings carrying no recognised deferred tax E256, excess cash is not fully available to the owner at book. B73 should name ratio-management as a distinct failing rationale. Buybacks justified expressly as restraining the expansion of capital to improve capital efficiency E126, executed at about 7.8 times book D4, by a board whose pay turns on earnings per share and return on equity E179, pass both existing tests. Propose: "a rationale stated in terms of capital efficiency, balance-sheet optimisation or hitting a return-on-equity level is a ratio rationale, not a value rationale, and fails unless a separate price-versus-value statement is also made."
• Munger — three. M13 and M57 need a formal not-applicable status distinct from data-insufficient, since for a non-financial, non-founder company both produce boilerplate that adds no information. M20, and parts of M7 and M19, assume an SG&A or advertising line that IFRS by-nature filers do not disclose — worth a standing note that such filers route these items to data-insufficient by construction, not by analyst shortfall, so a thin-looking answer is not misread as thin work. M27–M31, the moat-as-psychology items, require churn, repeat-purchase or referral data that neutral extraction from a Japanese yūhō is structurally unlikely to carry. Flag them as usually data-insufficient for Japanese-filer studies so a profile does not force an answer from thin material.
• Pabrai — five, all failure-derived. P1 should distinguish a soft floor from a hard floor priced away: the item's fail condition currently reads the same for decaying assets and for intact assets at 7.84× book D4. Propose "state whether the floor fails because the assets are weak or because the price has outrun them, and if the latter, carry the price at which the floor would hold," so the verdict guidance can route the second case to a stated buy-below. P39 should make the execution-price test mandatory, not implicit — HOYA passes on count D122 and fails on price at 31× (¥99,999,084,000 F679 over 4,708,300 shares F678) to 43× (¥28,050,977,500 F692 over 975,500 F691) against ¥674.14 D6; require the average execution price of each authorisation, stated as a multiple of normalised earnings and of book. P57 should add a segment-disclosure sub-clause: where segment assets are undisclosed D79, answer at group level only, mark the segment answer data-insufficient, and cap position size, because a moat that cannot be measured where competition occurs is not evidenced. P40 should ask whether share-based awards are equity-settled or cash-settled — half of the PSU E283 and half of the RSU E282 here are cash-settled, and the employee phantom-option scheme is wholly cash-settled E284, which is a price-linked bonus rather than ownership. P56 should check whether incentive actuals were struck on reported or normalised figures: the FY2026 bonus gate used reported earnings per share of ¥743.93 F838 against ¥674.14 ex-one-off D6.
• Li Lu — five. A new item: the issuer's own materiality judgement as evidence. L12 governs the analyst's duty to obtain complete information and has no provision for the case met here, where the issuer declares a category immaterial while the same document describes material items inside it E331E332E333. Propose: where two or more such contradictions appear in one filing, the analyst must widen the unknowns column to cover every undisclosed category and may not treat any omission as evidence of smallness. L21 needs a carve-out for repurchase above book — the retained-to-value test scores HOYA a failure purely because ¥171,970百万円 F1254 was repurchased at 7.84× book D4; run the test on per-share earning power instead and judge the repurchase separately against incremental returns. L26 should separate "no target with strong action" from "target with boilerplate", and check what a printed ROE actually is — HOYA has no ROE, PBR or cost-of-capital target anywhere, and the only ROE number in the document is a share-plan performance condition E334, while the actions are among the strongest available E126E189. L46 needs an explicit segment-opacity trigger for IFRS filers, since IFRS 8 permits segment revenue and profit with no segment assets D79: state that where profit is concentrated in a segment whose invested capital is undisclosed, segment return on capital does not exist and the L1 knowledge bar is failed for that company regardless of group returns. L25 needs a sub-test on the basis of incentive metrics F838D6.
• Claude — five, plus two library entries. (1) C33 needs a fallback for archives with no observed drawdown — its inputs are a ten-year margin low and an observed revenue decline, and a five-year monotonically-rising archive has neither D54D55D56D57; require the substitution to be named and graded pass, not data-insufficient. (2) C35's pass/fail semantics are ambiguous — two jury selves read fail as "the company does not clear the hurdle" and one read pass as "the item executed cleanly," producing a spurious divergence on an item whose substantive answers agreed within 38bp. (3) C34 and C40 assume an operating line exists and had to be re-expressed through a constructed proxy D36; require IFRS by-nature filers to name the proxy, its construction, and how many years it is computable for. (4) A new forensic item on remeasurement gains read for their economic sign — for every gain arising from remeasuring a liability, an obligation to acquire, contingent consideration or a written put, state what the remeasurement implies about the underlying asset and locate the matching movement in equity, non-controlling interests or the liability F2044F1019E270E335. C30 counts such items and C37 sizes them; neither asks what their direction means, and this study's principal finding came from asking. (5) C91's recency threshold needs a two-year-archive carve-out, since segment and note-level disclosures exist for two years only and a five-year citation is impossible for most forensic claims. Library, staged for the approval contract: Class-level, N=1, no rate claimed — "Japanese IFRS-reporting two-pillar precision-component group, no controlling shareholder, high-margin materials segment plus an ordinary consumer/medical segment, trading above 30× earnings," with HOYA (7741) the first member and its outcome on the class's key axis measurable: five-year total shareholder return 210.2 against a dividend-inclusive index at 202.2 F180F185E128. Case-level, so it never anchors a future outside view — "when a company books a gain on remeasuring an obligation to acquire something, ask first what the gain says about the thing being acquired; a downward revision to a purchase price is news about the asset, and its accounting sign is the opposite of its economic sign" E91E335.
• A process note that belongs in the record, and it is a correction. The standing rule — restated in full in companies/6861-keyence/thesis.md, that no ordinal or comparative claim about the record enters a stamp, thesis or deck without being checked against docs/calibration.md and the repository first — was broken by this study's own stamp, by both halves of its ledger, and by four of its five profile runs — stamp.md, figures.md, evidence.md, and the Pabrai, Munger, Li Lu and Claude runs all assert that HOYA is the record's first IFRS filer. It is not. companies/5202-nippon-sheet-glass/figures.md records an IFRS consolidated basis, stamped 2026-07-10, and companies/6464-tsubaki-nakashima/ likewise, stamped 2026-07-14 — both before this study. The claim was not checkable in docs/calibration.md, which records verdicts and predictions rather than accounting bases, and nobody asked which file would settle this before asking whether that file had been consulted. And because it entered the ledger, it was not one artefact's error but the shared basis every lens read — which is what makes this worse than a stamp-only slip. The claim appears nowhere in this thesis, and every observation the profiles hung on it — that IFRS by-nature presentation disables SG&A-dependent checklist items, that segment profit is pre-tax profit here F1328, that no 営業利益 or 経常利益 exists D36 — stands on its own without it. The lesson is the rule's, not the lenses': the check must fire on any "first/only/most" about the record, and its first step must be naming the file that would settle it. And one structural reason it did not fire is worth recording: the rule exists only in prior theses and in the study prompt — it is written into neither CLAUDE.md nor docs/process/, and no gate enforces it. Propose adding it to the integrity rules and to the citation-audit stage's agent checklist. The append-only rule governs the stamp and the profiles, so this thesis records the correction rather than editing them.

---

Corrections

Correction — 2026-08-03

• What was wrong: the claim that HOYA is "the record's first IFRS filer" entered this study in seven places, not one. The stamp: stamp.md:28 — "Pick #2 of the 'great businesses, at any price' batch — and the record's first IFRS filer." Both halves of the ledger: figures.md:6 — the header "BASIS NOTE — THE FIRST IFRS FILER IN THIS RECORD"; figures.md:192 — the source-substrate line, "the first IFRS filer in this record"; figures.md:197 — the basis note, "Basis note — the record's first IFRS filer."; and evidence.md:25 — "This is the record's first IFRS filer, so figures are not comparable line-for-line with the JGAAP studies." And four of the five profile runs: profiles/pabrai.md:23 ("this is the first IFRS filer I have read here"), profiles/munger.md:133 ("IFRS 'nature of expense' presentation filers — this being the record's first"), profiles/li-lu.md:139 ("This is the record's first IFRS filer"), and profiles/claude.md:263 ("this is the record's first IFRS filer and it will not be the last"). Only the Buffett run is clean. The orchestrator also carried the claim through the extraction and profile briefs for this study, which is how it reached all seven.
• What is correct: HOYA is not the first IFRS filer in this record. At least two precede it — 5202 Nippon Sheet Glass, whose figures.md header reads "Consolidated (連結), IFRS, March FYE", and 6464 Tsubaki Nakashima, whose header reads "Consolidated (連結), IFRS, December 31 fiscal year-end". Neither was checked before the claim was made.
• Root cause: reasoning-error — an uncited ordinal claim about the repository, the same failure mode corrected on 2026-07-31 in companies/5445-tokyo-tekko/thesis.md. The rule adopted then requires verification against docs/calibration.md and the repo before any such claim enters a stamp, thesis or deck. It was not applied here, and the claim entered the stamp — which, being part of the pushed provenance anchor, cannot be edited.
• Why this is worse than a stamp-only error, and it is the point of the correction: the claim reached both ledger files, and the ledger is the one artefact every lens reads. Under the isolation rule a profile run receives its checklist, figures.md and evidence.md — so a false statement in a ledger header is not one artefact's mistake that a reader can quarantine; it is established basis presented to all five lenses simultaneously, in the very files the integrity rules designate as the authority a profile may not go behind. Four of the five then reproduced it in their own words, which is exactly what a contaminated ledger looks like from the outside: independent-seeming corroboration of a single unchecked input. A stamp error contaminates the record's framing; a ledger error contaminates the evidence base. This strengthens rather than merely illustrates the case that the verification obligation binds the orchestrator's briefs: the ledger is written from the extraction brief, before any lens exists to dissent, and there is no downstream stage at which an isolated profile could have caught it.
• Impact on the analysis: none on the verdicts, the valuation or any lens's reasoning. Nothing in the four affected profile runs turns on whether other IFRS filers preceded this one — in every case the claim is a rhetorical preface to a checklist-revision proposal or a basis caution, and each of those stands unchanged with the ordinal struck. The IFRS conventions the briefs and the ledger specified (売上収益, 税引前当期利益, 資本合計, no 経常利益, JGAAP parent statements) are correct and were applied correctly, and no figure row or evidence passage is affected — the false claim sits in the ledgers' prose headers, not in any [Fn] or [En]. The verdicts are not re-issued or withdrawn.
• A second, more useful consequence: because the claim was believed, the balance-sheet-identity problem IFRS creates — that check-figures.py matches net assets|純資産 while IFRS prints 資本合計 — was briefed as a novel wrinkle to be solved from scratch. It was not novel. 6464 Tsubaki Nakashima's ledger already carries the note "IFRS presents total equity", meaning that study had solved it. A false ordinal claim caused real duplicated work, which is the practical argument for the rule beyond mere accuracy.
• Follow-up: the thesis above never asserts the claim and records it as a process note in the checklist-learning section. The standing rule is restated there and extended in two directions. First, the verification obligation binds the orchestrator's briefs, not only the artefacts, since this claim entered the study through a brief and propagated into a stamp, both ledgers and four profile runs before any gate could see it. Second, it binds the ledger stage specifically: an ordinal or comparative claim about the record has no place in figures.md or evidence.md at all, because neutral extraction is meant to quote and locate rather than to situate the filer against the repository — a claim of that shape in a ledger header is a framing the isolation rule then guarantees every lens will read. Propose that scripts/audit-citations.sh flag "first/only/most/Nth … in this record" strings in ledger and stamp files for human adjudication, since no citation can resolve them.