DENSAN CO., LTD. (3640): Cheap on Every Basis, and the Base Is the Whole Question

Verdicts

Lens	Verdict	Buy below	Most load-bearing items
Buffett	watch	¥1,600	B72; B42/B43; B71/B84
Munger	watch	¥1,900	M1/M61; M54/M56; M82
Pabrai	watch	¥1,450	P1; P56+P13; P49+P40
Li Lu	too-hard	—	L1; L35; L46
Claude	watch	implied ¥900	C33; C57; C98

Four watch and one too-hard, and three lenses name a price — ¥1,600, ¥1,900 and ¥1,450 — which is a spread of ¥450 between the highest and the lowest. 株式会社電算 is the closing pick of the "cheap AND good" batch, and the only one of the batch's three names on which the five lenses did not reach the same verdict, and the only one on which three of them wrote down a number. Pick #1 (東京鐵鋼) doubted the quality; pick #2 (フジシール) doubted the price; this one doubts the durability, and the doubt did not resolve. On the raw arithmetic the name is cheap on every basis anyone tried: 0.91× book D4, 3.35× reported earnings D6, 9.89× on normalised earnings D8, 6.17× enterprise value to normalised ordinary profit D90, with net cash of ¥1,899,819千円 D82 and a 63.3% equity ratio F50. It still drew no buy. The reason four lenses stopped at watch and the fifth stopped short of an opinion altogether is the same reason: nobody can size the base.

FY2026 is a bulge, and the ledger sizes it exactly. Consolidated revenue of 27,987,254千円 F5 is 1.5986× the mean of the three preceding years D16; ordinary profit is 3.0142× that mean D21; net income 3.0866× D25. The incremental revenue against the pre-bulge mean is +10,480,416千円 D17 and the incremental net income +3,027,382千円 D26, a 28.9% drop-through on the net line, measured against the 第58期–第60期 mean D29 and 40.87% on the operating line, measured against 第60期 alone — Δoperating profit of 3,778,884千円 (6,296,461 F852 less 2,517,577 F851) over Δrevenue of 9,246,324千円 (27,987,254 F5 less 18,740,930 F4). The two are not on the same base, and are not a matched pair: no 営業利益 is printed before 第60期, so the mean-based operating figure cannot be computed and the prior year stands in for it. The cause is not in dispute, because management states it: completing every customer's migration to standard-compliant systems by the original March 2026 national deadline 「が、売上、利益に大きく影響しました」 E123, with all 168 planned local bodies delivered on schedule F382E129. Every profitability figure in this document therefore says whether it is bulge-year or normalised, and the same price is quoted both ways throughout.

And the valuation only makes sense stated both ways. At ¥2,664 F1880 the market asks ¥14,784,308千円 for the whole company net of treasury D2. That is 0.91× book D4 and 3.35× the bulge year's ¥794.09 of earnings per share D6F40 — and 9.89× the ¥269.39 normalised EPS that is the mean of 第58期–第60期 D7D8. The dividend is the same trick in miniature: ¥140 a share F135 yields 5.26% D9, but ¥60 of it is a sixtieth-anniversary commemorative component F141F567E163, and the recurring ¥80 F566 yields 3.00% D10. Cheap on the bulge; still cheap, and merely ordinary in quality, on the run rate.

The business

株式会社電算 (DENSAN) is a regional systems integrator in Nagano, founded in 1966 by the local broadcaster 信越放送 to do the region's data processing E8, renamed to its present name later E9, listed since 2010 with the head office relocated in 2013 E16. It writes and runs the software that roughly 450 Japanese local public bodies F204E36 use to perform the unglamorous statutory arithmetic of local government: who lives here, what they owe in resident and light-vehicle tax, who receives the child allowance, what the water bill comes to, who is on the electoral roll, what the care premium is E39.

The product is called Reams, and it is the whole case. Reams is an in-house-developed integrated municipal package covering more than 30 packaged statutory business areas F210, with about 40 years of development and operating record behind it F211, and it is 73.7% of what the public segment sells F207E39. The public segment is 80.5% of group revenue F244D50, the industrial segment 19.5% D51, and 74.9% of all revenue comes from two prefectures, Nagano and Niigata — a concentration the company names as an issue in its own strategy section F228E65. The group has two reportable segments, which are simply the public/private cut of one market E20E282, four service lines E23, one wholly owned subsidiary acquired in 2016 E17F202F221, and 585 employees F85.

What is sold is not the software; it is build-plus-operate. The company sells system construction and then charges to run the result — operation, outsourcing, maintenance, data-centre hosting and cloud services, delivered as a one-stop service across the customer's IT lifecycle E24. The revenue-recognition note is the cleanest evidence of the mix: information-processing services are recognised at the recoverable portion of costs incurred and communication services rateably over the contract term E255; contracted software development runs on a cost-to-cost progress measure E256; and equipment is recognised at customer acceptance, supplies at deemed delivery E257. In FY2026, 19,937,913千円 of revenue was transferred over time and 7,890,954千円 at a point in time F1370F1367.

The moat mechanism is a switching cost, and the company states its own version of it. A municipality cannot stop computing its tax rolls, cannot go back to paper, and cannot swap the system over a weekend; in-house development is kept in house expressly so the package can be changed quickly when the law changes, which it does constantly E31E100. There is almost nothing else in the way of a barrier: the certifications are accreditations rather than concessions — systems-integrator accreditation in 1990, 特定システムオペレーション企業 certification in the 1990s, ISMS and cloud-security certification later E12E59E29 — and there is no territorial right, no exclusivity and no licence with a term.

There is essentially no plant, and one building carries the business. Tangible plus intangible fixed assets are 7,075,120千円 D77, 27.6% of assets F790D77, turning 3.96× in revenue D80. The single property is simultaneously head office, central operations facility and data centre — base-isolated, staffed 24/365, cloud-security certified E148E29 — carried at land of 1,716,469千円 F762 and buildings and structures of 3,680,117千円 F754. There is no goodwill anywhere and no impairment in either year E287, no overseas sales and no overseas property E286, and no customer at 10% or more of consolidated revenue, so the by-customer table is omitted E122E286.

Control sits with two corporate holders, and the filing never adds them up. 信越放送 holds 39.3% of the voting rights including 2.4% held indirectly F223F224E49; TOPPANエッジ holds 15.9% F226E50. Both are classified その他の関係会社; the filing presents them only ever separately, and states elsewhere that it has no parent company under Article 24-7(1) E52E315. 39.3% F223 + 15.9% F226 = 55.2% of the votes, a figure that appears nowhere in the document. Under the articles, ordinary resolutions carry with a one-third quorum F588 and special resolutions require two thirds of the votes cast F589 — so the two together elect every director and the larger one alone can defeat any special resolution.

The numbers

The five-year record is a sawtooth with a spike on the end, not a climb. Consolidated, 千円: revenue 17,306,483 → 17,804,937 → 15,974,648 → 18,740,930 → 27,987,254 F1F2F3F4F5; ordinary profit 1,242,594 → 2,507,109 → 1,233,744 → 2,523,882 → 6,294,427 F6F7F8F9F10; profit attributable to owners 859,281 → 1,611,829 → 898,610 → 1,842,221 → 4,478,269 F11F12F13F14F15. Ordinary margin runs 7.18 / 14.08 / 7.72 / 13.47 / 22.49% D34D35D36D37D38 — up, halved, up, halved, tripled. Revenue fell outright in 第59期, to 0.897× the prior year D45. Return on equity ran 12.2 / 19.8 / 9.1 / 15.7 / 30.9% F51F52F53F54F55. Earnings per share ran ¥171.75 / ¥321.46 / ¥169.29 / ¥317.42 / ¥794.09 F36F37F38F39F40.

Underneath that, book more than doubled and the balance sheet turned into a fortress. Equity went 7,412,628 → 16,256,331千円 F21F25; book value per share ¥1,476.85 → ¥2,927.30 F31F35; the equity ratio 35.6% → 63.3% F46F50. At 2026-03-31 there are no bonds D69 — the 社債明細表 reads 該当事項はありません — and no long-term borrowings left E291; interest-bearing debt is 2,213,251千円 D69 against 3,847,152千円 a year earlier D70, and cash and deposits of 4,113,070千円 F730 leave net cash of 1,899,819千円 D82 where twelve months earlier there was net debt of −1,295,980千円 D83. Net cash is ¥342.33 a share D84, 12.85% of the net market capitalisation D85. Interest cover is 323.7× D103, and there are 6,898,000千円 of undrawn committed lines across four banks D102F1016. The only guarantee obligation in the document is 20,000千円 for the employee mutual-aid association F1013E263.

Cash conversion over the window is good; the reporting year's is not, for a reason that reverses. Cumulative five-year operating cash flow is 12,728,180千円 D110 against cumulative investing of −4,180,490千円 D111, leaving 8,547,690千円 of cumulative free cash flow D112 — a cumulative conversion of 1.3135× reported profit D122. The bulge year itself converted at 0.9625× D121, because operating cash flow of 4,310,247千円 F65 was struck after absorbing a 3,971,671千円 increase in trade receivables F957 (the MD&A prints the same movement as ¥3,971百万円 F318E115). Receivables plus contract assets stand at 10,854,701千円 D96, 141.6 days of sales D99, having grown 1.577× against revenue growth of 1.493× D100. If revenue reverts, that build reverses into cash.

The asset base was harvested, not rebuilt, and the rebuild is already budgeted. Cash capex was 481,007千円 D113 against depreciation of 1,609,203千円 F939 — 0.30× D115, after 0.77× the prior year D116. Software carrying value fell from 1,937,861千円 to 943,146千円 F771F772, with 1,074,883千円 of amortisation against 80,119千円 of additions on the parent's (提出会社) 附属明細表 F1840F1839 — the only place the roll-forward is printed, there being no consolidated equivalent. Consolidated R&D was cut 26.0%, from 733,491千円 to 543,054千円 F1029F1030, taking R&D intensity to 1.94% D59 from 3.91% (733,491千円 F1029 ÷ 18,740,930千円 F4) — in the peak-profit year. Against that, the next-generation Reams programme is disclosed at a total development cost of 2,984百万円 running 2024年11月 to 2029年3月, of which 488百万円 was spent in the period F411F412E145; and, separately and on an expressly non-additive basis, the capitalised side carries planned intangibles of 1,662,344千円 with only 140,888千円 paid F469F470E149. The company's own statement of what that outlay buys is 増加能力の測定不能 — the increase in capacity cannot be measured E149.

The forward book is the year's most important disclosure, and it is flat rather than falling. Group order backlog is 16,159,613千円, 102.5% of the prior year F349F350, in a year revenue grew 49.3% F377. Public-segment backlog is 12,891,858千円 at 102.0% F341F342 against public-segment revenue up 58.7% F388. Within it the composition moved sharply: software-development-and-system-provision backlog rose to 7,606,170千円, 137.9% of prior F357F358; information-processing-and-communication backlog rose to 110.8% on orders of 140.0% F354F352; equipment-sales backlog collapsed to 651,545千円, 32.6% of prior F361F362; other-related-services backlog fell to 84.4% F366. Total remaining performance obligations rose to 15,989,271千円 from 15,565,828千円 F1300F1299; the within-one-year tranche is essentially flat at 13,233,495千円 against 13,254,575千円 F1292F1291; the tail beyond three years rose from 457,782千円 to 832,753千円 F1297F1298. Coverage of a year's revenue fell to 0.5713× D58.

The segments, on the bulge basis. Public-segment external revenue is 22,529,129千円 F1374 against 14,194,868千円 the year before F1325, and segment profit 5,764,624千円 F1381 against 2,112,947千円 F1332; the industrial segment earned 536,435千円 F1382. Segment margins are 25.59% public and 9.83% industrial D48D49. 45% of consolidated assets sit outside both segments as corporate assets — 11,611,268千円 F1390 against total assets of 25,645,634千円 F790 — because the head office and data centre are excluded from segment assets entirely E285.

Capital return arrived in one year, and one act of it was struck above today's price. Declared shareholder return for FY2026 is 1,585,614千円 D133, 35.41% of bulge net income D134 but 109.29% of normalised net income D135. The buyback was a single resolution — 270,000 shares for ¥808,650,000 F547F548, executed in one off-auction ToSTNeT-3 session, the resolution exhausted E157 — at an average of ¥2,995 D142, 4.63% of the issued count D141, and 12.4% above the ¥2,664 stamp (2,995 D142 ÷ 2,664 F1880). The dividend policy was rewritten by board resolution on 15 May 2026 onto an explicitly progressive basis with a 3.0% DOE guide F575E160E164, which on the 2026-03-31 common-share equity of 16,245,525千円 F1404 implies roughly 487,366千円 of annual dividend, about ¥88 a share on the 5,549,665 net-of-treasury count D1 — above the ¥80 recurring base F566.

And the market has been scoring this company for five years. The filing prints shareholder total return of 91.3 / 72.7 / 62.7 / 68.8 / 131.7 F182F183F184F185F186 beside a dividend-inclusive TOPIX of 102.0 / 107.9 / 152.5 / 150.2 / 202.2 F187F188F189F190F191. In four of five years the owner was behind where he started while the index nearly doubled — and the book value more than doubled over the same window F21F25. The bulge year moved the quote hard: the stamp sits at 0.676× the fiscal year's own high of ¥3,940 D12F196 and 1.985× its low of ¥1,342 D13F201. One stamp day's turnover — 21.4千株 on 55,898.200千円 F1890F1891 — is 0.38% of the net market capitalisation D14.

Returns, normalised, are ordinary. On the pre-bulge three-year mean of 1,450,887千円 of net income D24 and 2,088,245千円 of ordinary profit D20, normalised ROE on the 2026-03-31 equity base is 8.93% D65 — below the company's own published 10% ROE target F231E72, which it claims as met at 30.9% on the bulge year F233E73 — and normalised ROIC excluding cash is 10.13% D76 against 30.55% reported D75. The operating business earns well on the little iron it needs; the shareholder does not get it, because the equity base doubled while earning power did not.

The five lenses

Buffett — watch, buy below ¥1,600

Start with what the company sells, who pays, and why they keep paying. In one office building in Nagano — which is also its own data centre E148 — about six hundred people write and run the software that roughly 450 Japanese city and town halls use to do the unglamorous arithmetic of local government F204E36. Thirty-odd separate municipal jobs sit inside one package called Reams, which this outfit has been building and running for about forty years F210F211E39. Reams is 73.7% of what the public side sells F207; the public side is 80.5% of the whole company F244D50; and 74.9% of every yen comes from two prefectures F228E65. I like the shape of that. It is a toll bridge with a very particular kind of traffic — the town cannot stop computing its tax rolls, cannot go back to paper, and cannot swap the software over a weekend. There is essentially no plant: buildings, machines and software together come to 7,075,120千円 D77, only 27.6% of assets, turning nearly four times in revenue D80. A shopkeeper can hold that ledger in his head.

Now the first thing a student must learn: find out which year you are looking at. This one is a bulge, and the company tells you so — management says plainly that finishing the national migration by the deadline had a major impact on sales and profit E123, and they migrated all 168 bodies they had planned F382E129. Revenue reached 1.60× the average of the three years before it D16, ordinary profit tripled D21, and return on equity went from 15.7% to 30.9% F54F55. Even the dividend is dressed for a party — ¥140 a share, of which ¥60 is a sixtieth-anniversary present F135F141E163. So what is normal? Line up the three ordinary years just before: earnings per share of ¥321.46, ¥169.29 and ¥317.42 F37F38F39. Average them and normal is about ¥269 D7; average the profit and normal is 1,450,887千円 D24. Note the sawtooth — up, halved, up again. That is not a business with gently rising earning power; it is a business that breathes on a procurement cycle. Everything I say from here runs on that normalised basis.

On that basis the owner's return is ordinary. Normalised return on equity is 8.93% D65 — under the company's own published 10% ROE target F231E72, which it hit only in the bulge F233E73. Normalised return on invested capital net of cash is 10.13% D76; the published target governs return on equity and the operating margin, not that measure E72, and ten per cent on capital is a wage paid to the capital, not a franchise. The operating business earns handsomely on the little iron it needs, but the shareholder does not get it, because the balance sheet keeps filling up with things that earn nothing: equity went from 7,412,628千円 to 16,256,331千円 in five years F21F25 and the receivable-and-contract-asset pile is now 10,854,701千円, 141.6 days of sales D96D99, growing faster than sales did D100.

Which brings me to the item I care most about — for every dollar retained, has a dollar of market value been created? The filing answers it for you. Over five years the shareholder total return index runs 91.3, 72.7, 62.7, 68.8, 131.7 against a dividend-inclusive TOPIX of 102.0, 107.9, 152.5, 150.2, 202.2 F182F183F184F185F186F187F188F189F190F191. In four of five years the owner was behind where he started while the market nearly doubled — and the book value nearly doubled too. That is the Sanborn pattern in plain sight: earnings retained, book piling up, quote going nowhere, until one extraordinary year finally moved it. Time was not the friend of this shareholder.

The moat is real, though, and I want to be fair about it. The standardisation programme was the biggest re-procurement event this market has ever had — a common national specification is exactly the moment a competitor takes your account. DENSAN kept every one of them and won additional work besides: 62 family-register systems, 25 financial-accounting replacements, 36 network-equipment replacements F383F384F385E129, with the operating margin rising from 13.43% to 22.50% through it D32D33. A moat attacked and held. But understand what the state was doing: a common specification exists to make vendors interchangeable, the deadline is now extended to end-FY2030 for the laggards E58, and the government cloud is to sit alongside the company's own data centre E71. The next Reams generation is budgeted at 2,984百万円 through March 2029 F411E145, and the filing states the capacity effect of the associated 1,662,344千円 of planned intangibles as "not measurable" F469E149. Management put the biggest number in its plan next to the words "we cannot measure the return."

Two more things a partner should know. First, the forward book is flat: the one-year remaining performance obligation is 13,233,495千円 against 13,254,575千円 a year earlier F1292F1291, and total backlog grew 2.5% F350 in a year revenue grew 49.3% F377. Whatever next year is, it is not this year. Second, the reporting. The filing tells you it is under "no constraints" on capital policy E87E90 ten pages before it discloses that a 15.9% shareholder holds a contractual veto over any dilutive issuance and the right to name a director E138E143; it reports no related-party transactions E288 while tabulating 836,489千円 of them with its two big holders — 126,818千円 F247 plus 15,501千円 F249 plus 586,806千円 F253 plus 107,364千円 F255 E86E89E289; it reports no significant subsequent events E311 while three board decisions of 15 May 2026 sit elsewhere in the same document E312. None of that is fraud — the opinions are clean on the statements and on internal control E302E308. But it is not confession-time candor, and internal audit has no line to report directly to the board E222.

Value it as a private owner would. Eight to twelve times normalised earnings of 1,450,887千円 D24, plus the net cash of 1,899,819千円 D82, over 5,549,665 shares D1, gives ¥2,434 to ¥3,480 a share. Book is ¥2,927.30 F35; tangible book is about ¥2,728 (equity 16,256,331千円 F840 less intangibles 1,116,092千円 F776, over 5,549,665 shares D1). Only now the price: ¥2,664 F1880. That sits inside my range, not below it — 9.89 times normalised earnings D8, 0.91 times book D4, essentially at tangible book. The reported 3.35 times D6 is a mirage cast by the bulge; do not swing at mirages. And note that management itself bought 270,000 shares at an average of ¥2,995 last August D142F547, 12% above where the stock sits today.

Verdict: watch, buy below ¥1,600. There are no called strikes. This is a fair business at a fair price, and I want an adequate business at a wide discount. Take a third off the bottom of my range and I am a buyer under ¥1,600 — 0.55 times book F35, about 5.9 times normalised earnings (1,600 ÷ ¥269.39 D7), with net cash covering a fifth of the purchase price D84 and a 5.0% recurring dividend on the ¥80 base F566. That is a price this stock traded through in each of the last three fiscal years, whose lows were ¥1,482, ¥1,301 and ¥1,342 F199F200F201. Until then, watch — and watch the first post-bulge year, not the quote.

What a student should take from this: when a company's own filing tells you a year was driven by a one-time programme, that sentence is worth more than any multiple on the page — normalise first, price second, or the cheapness will be an optical illusion built out of a peak. And when you want to know whether retained earnings have worked for the owner, do not argue about it: many Japanese filings print five years of shareholder total return beside a dividend-inclusive index, which settles the one-dollar test in two rows F186F191. A moat that survives an attack is real and rare; a moat plus a fortress balance sheet plus a fair price is still only fair, and fair is a pass.

Munger — watch, buy below ¥1,900

Not another manufacturer, then. No plant to depreciate, no buggy-whip substitution risk. Good — that is exactly why we must be more careful, not less. A new kind of business tempts you to relax the checklist. Don't.

Invert it first, always. How does this company die, or at least disappoint? Five ways, all in the filing, none requiring imagination. One: this year is a bulge, not a run rate — a national statutory deadline forced every municipality to migrate by March 2026, the company finished all 168 of its planned migrations on schedule E129, revenue rose 49.3% F377 and operating profit 150.1% E125. That is a one-time toll booth closing, not a moat widening — and the public-segment order backlog, the actual forward pipeline, grew all of 2.0% F342 while billed revenue in that segment jumped 58.7% F388. And here is the tell: of the seventeen risk factors this filing enumerates, not one is written against the completion of the very programme that made the numbers E105E123. Inversion is not decoration — you write the obituary before you write the thesis, and this filing's own author skipped that step. Two: 80.5% of revenue in one segment, 74.9% in two prefectures E83E65. Three: two "other affiliated companies" — 39.3% F223 and 15.9% F226 — who together hold a majority never once printed as a combined figure anywhere in the document E52. Four: the president is eighty, came up through the 39.3% holder's own management, and the filing names succession as an open risk with no candidate on the page E102E193. Five: in-house headcount has fallen every year for five years, 694 to 585 F81F85, even as the company insists in-house development is how it protects its know-how E100.

Now the part that actually decides the verdict, and it isn't the arithmetic — the arithmetic is fine. Net cash of 1,899,819千円 D82 against a 14,784,308千円 market capitalisation D2, no bonds D69, interest cover over 300 times D103, a price below book at 0.91× D4 and under ten times normalised earning power at 9.89× D8. If cheapness were the hurdle this would clear it standing up. Cheapness is not the hurdle. The hurdle is: would I want to be in business with these people, on these terms, indefinitely? And here the filing answers its own question three separate times, each time the wrong way. It says the two large shareholders impose no constraint on strategy, personnel or capital policy E87E90 — on the same pages that disclose TOPPAN's contractual right to name a director and its veto over any dilutive issuance, an arrangement the company then calls 「軽微」, a minor governance effect E138E142. It says related-party transactions are 該当事項はありません E288 while disclosing, in the risk section rather than the related-party note, 126,818千円 F247 and 586,806千円 F253 of actual annual trade with those same two shareholders E86E89. It says there were no material subsequent events E311 while three board resolutions after year-end rewrote the dividend policy and abolished the shareholder-benefit programme E312E164E314. One such gap is a judgment call. Three, in one document, is a pattern — the report's promotional layer quietly disagreeing with its own audited layer.

What is left underneath is real: forty years with Reams F211, roughly 450 municipal customers F204, a genuine switching-cost moat — but a moat that has stayed exactly where it was born. Nationwide expansion has been the stated, unmet ambition for years running E65E98; four branches opened in Kansai, Tohoku, Kyushu and Hokkaido were all subsequently closed E11. One tendency, not several in confluence — no network effect, no brand conditioning, no habit loop beyond the customer relationship itself. A single-force business, riding a single-force year: pre-bulge ordinary margin bounced 7.18 / 14.08 / 7.72 / 13.47% D34D35D36D37, project-cycle volatility, not compounding. And the one large discretionary act of the year was a buyback struck at ¥2,995 D142 inside a fiscal-year range of ¥1,342 to ¥3,940 F201F196 — the upper half, and above today's quote.

Verdict: watch, buy below ¥1,900 — and I want it understood that this number is a character discount, not a valuation output. My value is a range, not a point: roughly ten times normalised earning power D8 on a business earning about ten per cent on invested capital ex-cash D76. ¥1,900 is 0.65 times book (1,900 ÷ ¥2,927.30 F35) and about 7.05 times normalised earnings (1,900 ÷ ¥269.39 D7), and the gap between that and the ¥2,664 quote F1880 is what I require to be paid for an eighty-year-old chief executive with no named successor E193E102, two related parties holding the votes F223F226, and a filing that says "not applicable" or "no constraints" three times where something exists. Buy it cheap enough to be paid for all of that, or don't own it at all.

What a student should take from this: a government deadline can manufacture a year's worth of numbers that look exactly like a widening moat if you only look at the trailing figures — always ask what the backlog, not the revenue, says about next year F350F377. And read the actual contract behind a related-party alliance before you accept the company's own adjective for it; 「軽微」 is a claim, not a fact E142, and a filing that says "no constraints" on one page and prints a veto right on the next is telling you more about its authors than about its shareholders.

Pabrai — watch, buy below ¥1,450

Let me start where I always start, which is not with the story. It is with the question: if I am wrong about everything, what do I own?

Cash and deposits of 4,113,070千円 F730. Receivables and contract assets of 10,854,701千円 D96, owed overwhelmingly by Japanese municipalities — the bad-debt reserve against that book is 60千円 F746, which tells you the payers do not default. A piece of Nagano land at 1,716,469千円 F762. Sixteen unlisted stakes F679 at 202,927千円 of cost F1135 and three listed ones F681 marked at 439,286千円 on the parent's 政策保有株式 table F682. Against all of it, total liabilities of 9,389,303千円 F820, of which interest-bearing debt is 2,213,251千円 D69, no bonds D69 — the 社債明細表 reads 該当事項はありません — and no long-term borrowings left at all E291, and net cash of +1,899,819千円 D82 where a year earlier there was net debt D83.

Now mark it the way you would if you actually had to sell. Cash at par. Receivables down a fifth even though they are municipal. Inventory and other current assets down half. Nagano land down a fifth. And the buildings, 3,680,117千円 F754, down sixty per cent — because that "building" is a purpose-built, base-isolated, 24/365-staffed data centre in Nagano E29, and there is no secondary market for a Nagano data centre. Ships trade. Motels trade. This does not. Software goes to zero: 1,116,092千円 of intangibles F776 that are worth exactly what the franchise is worth, and the franchise is the thing we are trying to value. Deferred tax assets of 658,302千円 F780 go to zero too. Add it up and take the liabilities off and I get roughly 8,579,622千円 — ¥8.58bn — about ¥1,546 a share on 5,549,665 shares D1, a floor at 58% of the ¥2,664 purchase price F1880. And the truly hard part, needing no buyer at all — net cash D82 plus the listed cross-holdings F682 — is about ¥421 a share, sixteen per cent of the price.

Stop right there. That is a soft floor. Heads I win; tails I lose forty per cent. This is not the Japanese setup my checklist was built for, where net cash and listed securities cover most of the market capitalisation. This is asset-light, and asset-light means there is nothing to liquidate. Write it on the wall: an equity cushion is not an asset floor. Book is ¥2,927.30 F35 and the stock trades at 0.91× it D4, and none of that book protects you, because most of it is receivables that unwind only if the business keeps running and a data centre nobody else wants.

So the whole case has to come from earnings, which means looking at the denominator before the multiple. The headline is a 3.35× P/E D6. Beautiful, and a mirage. Revenue was 1.60× the pre-bulge three-year mean D16, net income 3.09× D25, operating margin went from 13.43% to 22.50% D32D33 — because a statutory deadline pulled several years of demand into one year E123E129. On the pre-bulge mean, EPS is ¥269.39 D7 and the same ¥2,664 is 9.89× D8; normalised ROE is 8.93% D65 and normalised ROIC ex-cash 10.13% D76. Roughly ten times earnings for a business earning roughly its cost of capital. That is a fair price. It is not a fifty-cent dollar.

Do not take my word for the normalisation — take the order book's. Revenue grew 49.3% F377 while total backlog grew 2.5% to 16,159,613千円 F349F350; public backlog grew 2.0% F342 against 58.7% F388; equipment backlog collapsed to 651,545千円, 32.6% of prior F361F362; one-year remaining performance obligations went flat, 13,254,575千円 to 13,233,495千円 F1291F1292. And when I went looking for management's own warning about it, I found seventeen enumerated risk factors and not one addressed to the completion or expiry of the migration work E105. That is my whole problem with this name at this price. The Dhandho hunting ground is where the market prices uncertainty as if it were risk. Here the market has priced it correctly. Nobody is panicking: the shares nearly doubled off the year's low of ¥1,342 F201, total shareholder return was 131.7 F186, and the company itself bought 4.63% of its stock D141 at ¥2,995 D142, twelve per cent above where it trades today. Munger's cannibals eat themselves cheaply; this was a nice, expensive meal.

There is real quality here. Reams has forty years of record F211, covers thirty-plus packaged municipal functions F210, sits inside roughly 450 bodies F204 and is 73.7% of public-segment revenue F207. Ripping out the system that runs a city's tax roll is not something a mayor does to save money. The balance sheet is genuinely clean D102D103D104, five-year cumulative operating cash flow of 12,728,180千円 converts at 1.31× reported profit D110D122, and the new progressive-dividend policy with a 3.0% DOE guide E160F575 pays you about three per cent to wait D10. But look at what the moat's own numbers say: ordinary margin 7.18, 14.08, 7.72, 13.47, 22.49 D34D35D36D37D38 is not a durable spread, it is a business whose profits arrive in policy-driven pulses. And the acid is visible — national standard specifications are by design a commoditising force, the company's own risk factor names state-led standardisation and joint use among municipalities as things that may hurt it E84, and management describes the government cloud being used alongside its own data centre E71.

So I ran the tree. Pulse drains, no second act, ten times normalised earnings plus three years of the ¥80 recurring dividend F566D7: about ¥2,930, roughly flat. The FY2030 extension for laggard bodies E58 gives a smaller second bite: about ¥4,475. Commoditisation bites and margin reverts to the 第59期 trough of 7.72% D36 on pre-bulge revenue D15: about ¥1,703. Something breaks — a leak of entrusted municipal personal data E94, a data-centre failure E93 — and you are at the asset floor, about ¥1,546. Weight those 40/30/20/10 and I get about ¥3,010, thirteen per cent over three years. Four per cent a year, with two branches out of four losing you thirty-six to forty-two per cent. That is a coin flip dressed as a bargain, and few-bets-big-bets means I do not take coin flips.

Verdict: watch, buy below ¥1,450. Run the identical tree at ¥1,450 and the shape inverts: the worst branch, the asset floor near ¥1,546, sits above the purchase price; the recurring dividend yields 5.5% on the ¥80 base F566; net cash of ¥342.33 a share D84 is nearly a quarter of what you paid; and you are at 0.50× book (1,450 ÷ ¥2,927.30 F35) and 5.4× normalised earnings (1,450 ÷ ¥269.39 D7). That is Papa Patel arithmetic. And one last thing, which is why my number is where it is and not a little higher: two strategic holders own 39.3% and 15.9% of the votes F223F226; the filing says twice, verbatim, that the group is under no constraint as to capital policy E90E87 and then discloses a prior-written-consent veto over dilutive issuance it calls 「軽微」 E138E142; the related-party note reports nothing against 836,489千円 of tabulated trade E288E289; internal audit has no line to the Board E222; independent officers fall from five to four with no reason given F613F616E190 while the incoming outside director is a serving TOPPAN executive E209 and the incoming full-time audit-committee member is the company's own General Manager of Corporate Planning E210. None of that is fraud. All of it is a company where the governance apparatus is being staffed inward while two counterparties hold the votes. I do not need a discount for that if I am buying at fifty cents. I very much need one at ninety.

What a student should take from this: asset-light is a compliment to a business and an insult to a margin of safety — when there is nothing to liquidate, book value is not a floor, so mark everything else to zero before you argue about multiples. And when a single year's earnings were manufactured by a statutory deadline, the honest denominator is not in the income statement, it is in the order book: revenue grew 49.3% F377 while backlog grew 2.5% F350, and that one divergence told me more than the 3.35× P/E D6 ever could.

Li Lu — too-hard

Twenty-five studies in, this is a business with no factory, and I want to begin where that absence usually helps a buyer and here does not. When you own a plant you can walk it and count what it makes. When you own a regional systems integrator you own contracts you cannot see, renewed on terms you are not told, by customers whose budgets are set by statute. Everything therefore turns on one question: does the public record let a part-owner know what the next ten years of those contracts look like? The answer is no, and I want to show you exactly which sentence defeated the analysis, because a student who says "too hard" without naming the sentence has learned nothing but a posture.

Begin with what is genuinely admirable. A sixty-year-old company founded by a local broadcaster to do the region's data processing E8, serving roughly 450 local public bodies F204 on an in-house package covering more than thirty statutory functions F210 with about forty years of record F211. In the year under review it migrated every one of the 168 bodies it had planned onto the national standard by the original March 2026 deadline E129 — a deadline the state itself could not meet nationally, and has now extended to end-FY2030 for the laggards E58. That is executed work, not narrated work. The balance sheet is unimpeachable: net cash of 1,899,819千円 D82 where a year earlier there was net debt of −1,295,980千円 D83, interest cover of 323.7× D103, a 63.3% equity ratio F50, and 6,898,000千円 of undrawn committed lines D102. And the figures reconcile with a discipline I rarely see: the balance-sheet identity is exact in both years D147D148, segment profit foots to operating profit exactly D146, and even the 12,129千円 gap between the two buyback figures resolves precisely to the off-auction commission D130.

Now the sentence. Management states the cause of the year plainly: completing the standard-compliance migration by the original March 2026 deadline 「が、売上、利益に大きく影響しました」 E123. The owner's question follows immediately — what does the business earn when that work is finished? I read the whole of 事業等のリスク, all seventeen headings E105, and not one is directed at the completion or expiry of the migration. The programme appears as opportunity on page after page and never once as an expiring revenue event. What forward disclosure the company does give points the other way: public-segment order backlog grew 2.0% F342 against 58.7% revenue growth F388, and total contracted forward work of 15,989,271千円 F1300 is 0.57× a single year's sales D58.

So I must supply the number the filing withholds. The ledger's normalisation — the mean of the three pre-bulge years, 1,450,887千円 D24 — is the natural anchor, and at ¥2,664 F1880 that is 9.89× D8. But those three years are not a clean base either; they contain the migration's own run-up. Strip further, to the worst printed year in the window, 第59期: net income 898,610千円 F13 and EPS ¥169.29 F38, at which the stamp is 15.7× (2,664 F1880 ÷ ¥169.29 F38). Note what that year was — not a recession, not a crisis, an ordinary year in which revenue fell 10.3% D45 and the ordinary margin halved to 7.72% D36. A business whose margin can halve in an ordinary year, priced at 15.7× the last time it did so, is not a bargain waiting to be recognised.

Then my second question: is intrinsic value compounding or quietly melting? Across the four pre-bulge years revenue went from 17,306,483千円 to 18,740,930千円 F1F4 — three annual intervals, about 2.7 per cent a year — while headcount fell from 694 to 585 F81F85. Trace the retained yen: book value per share rose ¥1,476.85 to ¥2,927.30 F31F35 and cumulative dividends over 第58期–第61期 were ¥259 (¥45 F132 + ¥37 F133 + ¥37 F134 + ¥140 F135) against cumulative EPS of ¥1,602.26 (¥321.46 F37 + ¥169.29 F38 + ¥317.42 F39 + ¥794.09 F40) — a ratio of 1.08×, or about 0.99× once the ¥123 year-end dividend F570 still sitting inside the closing book is removed. Four years, including the best year in the company's history, converted one yen retained into roughly one yen of book. Nothing more. That is because there is nowhere to put it: capex ran at 0.30× depreciation D115, the software asset fell from 1,937,861千円 to 943,146千円 F771F772, and the one place capital must go is the Reams rebuild — 1,662,344千円 of planned intangibles F469, more than a full year of normalised earnings D24, whose stated capacity effect is 増加能力の測定不能 E149. When the operator will not say what the largest committed outlay produces, the outside owner cannot say either.

Meanwhile the return itself converges downward. Normalised ROE on today's enlarged equity is 8.93% D65 — below the company's own 10% target F231 — while the filing celebrates the bulge-year 30.9% E73. The new policy is a 3.0% DOE guide F575 on a progressive basis E160. Three per cent paid out of nine per cent earned, in a business that cannot reinvest, is the mechanism by which cheap Asian shares stay cheap. And the one large discretionary act of the year — 270,000 shares repurchased at ¥2,995 D142 — was executed at 1.37× the opening book value (2,995 D142 ÷ ¥2,185.77, the 第60期 BPS F34) and 12.4% above the price I am invited to pay today F1880.

Finally, read the structure plainly. Two corporates hold 39.3% F223 and 15.9% F226 — 55.2% between them — and the filing never once combines them, states there is no parent company, and presents them as two separate risk factors E52E315. One supplied the president E193 and an outside director who is its own president and is the single outside director not designated independent E200E214. The other holds a contractual right to nominate a director and a prior written consent right over any dilutive issuance E138, which the company characterises as 「軽微」 E142 while asserting on two earlier pages that it is under no constraint whatever as to capital policy E143E87E90. The formal related-party note reports 該当事項はありません E288 while the risk section tabulates 836,489千円 of transactions with those same two holders E289. The subsequent-events note reports none E311 against three board decisions dated 15 May 2026 disclosed elsewhere in the same document E312. None of these is a misstatement of a number. All of them are the same disclosure instinct, and my rule is that accounting and disclosure you cannot fully trust caps the verdict before valuation begins.

Verdict: too-hard, on L1 first and on L35, L5 and L46 behind it. Not because the business is bad — it is a sound, solvent, well-run regional franchise run by people who deliver what they promise. Because the one variable that decides its next decade, the post-standardisation revenue base of the public segment, is absent from the record, is not even listed as a risk, and at ¥2,664 F1880 the price absorbs none of it. No conservative construction I can build reaches twice the price: book is 16,245,525千円 F1404 against a 14,784,308千円 market capitalisation D2, 1.10×; capitalising the normalised NOPAT-equivalent of 1,453,419千円 D76 at ten times and adding net cash D82 gives about 1.11× the price. The margin of safety is a tenth, not a half. Intellectual honesty about that boundary is the whole of the discipline. I would rather be plainly ignorant here than cleverly wrong.

What a student should take from this: when a company's best year is created by a statutory deadline, the deadline is not a tailwind — it is a maturity date, and your job is to find the sentence that tells you what falls due after it. If the filing never names that event as a risk while naming seventeen lesser ones E105, treat the silence as the disclosure: the company either does not know or will not say, and either way you do not know. And measure the return on the equity you would actually be buying, not the equity the company earned it on — a business whose book value has doubled while earnings have not F21F25D65 is converging toward its cost of capital no matter how good last year looked.

Claude — watch, implied buy-below ¥900

I registered my priors figures-blind, and the ledger has embarrassed me in one specific, expensive way. I predicted the standard programme-cycle trap: cheap on trailing earnings, ordinary-to-full on normalised ones, with nothing changed except which number went in the denominator. I called it "the prediction most likely to embarrass me and the one I most want graded." It is wrong. The stock is 3.35× reported earnings D6 and 9.89× the three-year pre-bulge mean D8 and 0.91× book D4. My registered normalised band was 13–22×, with P(normalised P/E above 15×) = 0.60. The actual is below the bottom of it. That miss is the difference between a pass and a watch, and I want it on the record first.

What I got right. The backlog family, almost cleanly: total backlog within ±10% (102.5% F350), software-development backlog up more than 25% (137.9% F358), equipment backlog down more than 30% (32.6% F362), public backlog within ±10% against public revenue up far more (102.0% F342 against +58.7% F388), the remaining-performance-obligation near band and tail inside 65–85% / 15–35% F1291F1292F1299F1300. The working-capital caveat I registered — that a progress-recognising integrator with a March year-end strikes its cash high against a receivable peak — was the best call in the file: receivables plus contract assets are 141.6 days of sales D99, growing 1.577× against revenue's 1.493× D100. Terminal-year cash conversion below 1.0 (0.9625× D121) for exactly the reason I named, cumulative five-year conversion above 1.0 (1.3135× D122), free cash flow positive in all five years D105D106D107D108D109, the home prefectures at 74.9% F228, the flagship at 73.7% of public revenue F207, the equity ratio at 63.3% F50, salary growth +9.5% F715, treasury at 4.93% F544 and held rather than cancelled F553E156 — all inside their registered bands. And the drop-through landed dead centre: I said 30–55% with a ~40% central, and the operating measure — struck against 第60期 alone, which is the base my registration named — is 40.87% (Δ営業利益 3,778,884千円 from F851F852 ÷ Δ売上高 9,246,324千円 from F4F5).

What I got wrong. I understated the bulge throughout: revenue +49.3% against a predicted +10–30% F377; public revenue +58.7% against +15–45% F388; the public segment at 80.5% of revenue against 55–75% D50; operating margin 22.50% against 11–17% D33; ROE 30.9% against 13–20% F55. One root cause: I modelled a programme spread across a ramp, and this was 168 bodies migrated in the single terminal year F382. I also understated the pre-bulge level — ordinary margin averages 11.76% across 第58期–第60期 D35D36D37 against my 6–11% band. Say the basis out loud, because the metrics do not match: the band I registered was a through-cycle operating margin band, and 11.76% is an ordinary margin. No 営業利益 is printed before 第60期, so the substitution is forced rather than chosen — but a prediction scored against a different metric must be scored as such, and I score it as such. There is also a caveat the mean hides: the series alternates 7.18 / 14.08 / 7.72 / 13.47% D34D35D36D37, a biennial oscillation, not a stable premium. And I overstated the balance sheet: I predicted net cash at 25–50% of market capitalisation; it is 12.85% D85, and a year earlier the group was net debt −1,295,980千円 D83. There is no Japanese cash box here.

The prior I got right and under-weighted was the register. From the その他の関係会社 classification alone E49E50, with no figure, I put the two holders' combined stake at ≥30% with p 0.75, ≥40% with p 0.35 and either individually ≥33.4% with p 0.15. They hold 39.3% F223 and 15.9% F226 — 55.2% together, one alone above the special-resolution blocking threshold F589. Right on direction, far too timid on magnitude. I also pre-registered the consequence: if the two hold a combined blocking or controlling position, the verdict caps at watch regardless of price. I honour that.

Did the annuity step up? Modestly, yes, and I had the shape right while choosing one wrong measure. I nominated the over-time revenue share as the test; that was a construction error, because the over-time leg includes cost-to-cost development and is not the annuity — it fell from 73.9% to 71.6% while the annuity rose. Note the denominator, because I first computed it on the wrong one: the share is over-time revenue of 13,783,722千円 F1321 then 19,937,913千円 F1370 against revenue from contracts with customers of 18,644,146千円 F1315 then 27,828,867千円 F1364, not against total 売上高, which also carries sublease income outside the standard's scope — 96,783千円 F1324 and 158,387千円 F1373. The direction is unaffected either way. The real annuity is the information-processing-and-communication line: +16.6% to 4,068,640千円 from 3,490,117千円 F1352F1303, its backlog 110.8% F354, its orders 140.0% F352; and the obligation tail beyond three years rose from 457,782千円 to 832,753千円 F1297F1298. Against that, total remaining performance obligations rose only 2.7% F1299F1300, the within-one-year band fell F1291F1292, and coverage of revenue collapsed from 0.83× (15,565,828千円 F1299 ÷ 18,740,930千円 F4) to 0.5713× D58. The company delivered its forward book and did not refill it in proportion.

The owner arithmetic, downside first. Downside revenue = the pre-bulge three-year mean of 17,506,838千円 D15 plus the demonstrated annuity increment of 578,523千円 F1303F1352, rounded to 18,000,000千円; margin = the observed five-year low of 7.18% D34; tax at the forward 31.3% rate E279; less 27,563千円 of stock compensation F657F665; on 5,554,665 diluted shares D1F1414. That is ¥154.9 of downside owner earnings per share and, at a 10% required return, ¥1,549 — a −41.8% loss from the stamp F1880. My jury ran the same item three times and returned ¥1,419, ¥893 and ¥476. That is genuine divergence, it is one-directional, and it drags the published threshold down: implied buy-below ¥900, with the honest object the band ¥476–¥1,549 rather than the point.

Why watch and not buy, when the arithmetic looks close. At ¥2,664 F1880 the normalised owner yield is roughly 10% on my own working — and all three selves, running the private-owner-yield item independently, came back DECLINE at 6.2%, 7.6% and 7.7% against hurdles of 8.3% to 9.0%, each applying a bigger maintenance-capex correction and crediting zero or negative deployable cash. Three independent runs of my own method disagreeing with my draft is not confirmation of anything, but it is a correction, and I take it. Add the register cap I pre-registered F223F226. Add the epistemic inventory: three of my five load-bearing unknowns are resolvable now from free public sources this two-document archive does not contain — management's own FY2027 forecast, the predecessor auditor behind a four-year-old change F647, and whether the blockholders file as joint holders. A buy leaving those unread is conviction manufactured from time spent rather than facts established.

Why not pass. Growth here does not destroy value: normalised return on invested capital ex-cash is 10.13% D76 on an asset base that is 27.6% of the balance sheet and turns 3.96× in revenue D77D80. The register makes every unlock unreachable — but the thesis needs no unlock, because the price clears a hurdle without one. The reverse-DCF is the cleanest statement of what is on offer: at a 10% discount rate and zero growth, the enterprise value of 12,884,489千円 D86 implies perpetual after-tax operating earnings of 1,288,449千円 against the normalised 1,453,419千円 D76. The price embeds through-cycle earnings settling at about 89% of the pre-bulge three-year mean, forever — the boundary of the least likely state in my own registered distribution. I think that is too pessimistic. I am not confident enough in that opinion, on a two-document archive with a controlled register and a forward book that did not grow, to convert it into a purchase.

Verdict: watch, implied buy-below ¥900, band ¥476–¥1,549. Two independent caps hold it there and would at any price on this run: the resolvable-now unknowns left unresolved, and the jury's dispersion on the item that carries the whole distance between watch and buy.

What a student should take from this: three things. One: a bulge year is not one distortion, it is four, and they point in different directions. Everyone sees that revenue and margin are inflated. Fewer see that the balance sheet is inflated too — net cash of 1,899,819千円 D82 having been net debt of −1,295,980千円 twelve months earlier D83, with income taxes payable of 1,644,147千円 against 665,414千円 the year before F799F798 sitting on top of it. Fewer still see that the asset base is deflated — software fell from 1,937,861千円 to 943,146千円 F771F772 and capex ran at 0.30× depreciation D115, so the year's margin was earned partly by consuming an asset that must now be rebuilt for a named 1,662,344千円 with only 140,888千円 paid F469F470. And fewer again see the one that runs the other way: that same amortisation is about to roll off, which is a margin tailwind into the reversion — 1,074,883千円 of it against 80,119千円 of additions, on the parent's (提出会社) 附属明細表 F1840F1839, no consolidated roll-forward being printed. Count all four before you normalise. Two: name the measure before you look, and be willing to say the measure was wrong — a pre-registered test that turns out to measure the wrong thing must be retired out loud, not quietly replaced with the one that agrees with you F1321F1370F1303F1352. Three: what the jury is for. Three independent runs of my own bear case returned ¥1,419, ¥893 and ¥476 — a 3× spread on the same rows and the same instruction. That spread is not noise to be averaged away; it is the finding.

Synthesis

Where the five lenses agree

Four reached watch and one too-hard, and beneath that split the panel agrees on almost everything factual.

First, the year is not the run rate, and nobody disputes the size of the distortion. Revenue 1.5986× the three-year pre-bulge mean D16, ordinary profit 3.0142× D21, net income 3.0866× D25, on management's own attribution to a statutory deadline E123 and 168 migrations delivered F382E129. Every lens re-ran its arithmetic on the 第58期–第60期 mean — 1,450,887千円 of net income D24, ¥269.39 of EPS D7 — and none used the reported multiple for anything except to reject it.

Second, the balance sheet is not the question. Net cash of 1,899,819千円 D82, a 63.3% equity ratio F50, interest cover of 323.7× D103, no bonds D69 and no long-term borrowings E291, 6,898,000千円 undrawn D102. Survival is not at issue on any lens.

Third, the normalised economics are ordinary, not franchise-grade. Normalised ROE 8.93% D65 — below the company's own 10% target F231 — normalised ROIC ex-cash 10.13% D76, and a five-year record whose ordinary margin alternates 7.18 / 14.08 / 7.72 / 13.47% before the spike D34D35D36D37.

Fourth, the disclosure has a consistent tilt, and all five name the same four instances: 「何ら制約等は受けておりません」 asserted twice against a disclosed prior-written-consent veto over dilutive issuance called 「軽微」 E87E90E138E142E143; 関連当事者情報 該当事項はありません against 836,489千円 of tabulated blockholder trade E288E289; no significant subsequent events against three dated 15 May 2026 board decisions E311E312; and seventeen enumerated risk headings with none addressed to the end of the programme that made the year E105.

Fifth, the forward book did not grow with the revenue. Total backlog 102.5% F350, public backlog 102.0% F342, one-year remaining performance obligations flat F1291F1292, coverage of revenue down to 0.5713× D58.

Where they diverge — a staged exchange

---

On the price: three buy-below numbers, ¥450 apart, on the same ledger — and the spread is itself the finding.

BUFFETT (B91, the private-owner appraisal; B93, the required discount from the conservative figure). My object is an appraisal, and then a discount from the bottom of it. Eight to twelve times normalised earnings of 1,450,887千円 D24, plus net cash of 1,899,819千円 D82, over 5,549,665 shares D1, is ¥2,434 to ¥3,480. The quotation of ¥2,664 F1880 sits inside that range, not below it. A third off the conservative end is ¥1,622, so I write ¥1,600 — 0.55 times book F35, and a level this stock traded through in each of the last three fiscal years, whose lows were ¥1,482, ¥1,301 and ¥1,342 F199F200F201. Note what I did not do: I did not build a scenario tree and I did not go to the trough. The three-year mean is my base, and the discount is my protection against the mean being wrong.

PABRAI (P53, the fifty-cent dollar; P3, the probability-weighted tree; P1, the floor). We normalise to the same base and I arrive ¥150 lower, and the difference is entirely in the object. Yours is a discount from an appraisal; mine is half of a conservative intrinsic value, and I computed that value at ¥2,900 — ten to eleven times normalised EPS of ¥269.39 D7, cross-checked against book of ¥2,927.30 F35. Half is ¥1,450. But the number that actually did the work is the floor, and it is the reason I would not go higher. Mark this balance sheet the way a bank would: buildings of 3,680,117千円 F754 down sixty per cent, because that building is a purpose-built base-isolated Nagano data centre E29 with no secondary market; intangibles of 1,116,092千円 F776 to zero, because they are Reams itself; deferred tax of 658,302千円 F780 to zero. That leaves about ¥1,546 a share, 58% of the ¥2,664 stamp F1880. An equity cushion is not an asset floor. At ¥1,450 the worst branch of my tree sits above what I paid. That is the entire reason for the number.

MUNGER (M88, the gap; M54, the character gate). And I am the outlier at ¥1,900, so let me be plain about what my number is, because a reader would otherwise assume it is a valuation and it is not. My value is a range, not a point — about ten times normalised earning power D8 on a business earning about ten per cent on capital D76 — and I decline to pretend to more precision than that. ¥1,900 is 0.65 times book (1,900 ÷ ¥2,927.30 F35) and about 7.05 times normalised earnings (1,900 ÷ ¥269.39 D7), and the distance from ¥2,664 F1880 is what I require to be paid for an eighty-year-old chief executive with no named successor E193E102, two counterparties holding 55.2% of the votes F223F226, and a filing that writes 該当事項はありません twice where something exists — the related-party note E288 and the subsequent-events note E311 — while asserting 「何ら制約等は受けておりません」 over a disclosed veto it calls 「軽微」 E143. Warren discounts a valuation. Mohnish halves one. I am pricing a character defect, and that is why my number is the highest of the three.

BUFFETT. Then the spread is not disagreement about the ledger at all. We three used the same normalised base — 1,450,887千円 D24, ¥269.39 a share D7 — and produced ¥1,450, ¥1,600 and ¥1,900 because we applied three different operations to it: half of a value, a third off an appraisal's floor, and a governance haircut. That is worth writing down as a finding rather than smoothing into an average. When a company's earning power depends on where a one-time programme leaves the base, reasonable methods disagree by roughly a third — 1,900 ÷ 1,450 is 1.31× — and the honest object is the interval, not any point inside it.

---

On the anchor: the mean or the floor? This is the disagreement underneath the prices.

CLAUDE (C33, the bear case valued first; C104, the verdict-flipping parameter). My threshold is a different animal again, and I should say so before it is compared to theirs. ¥900 is not a value; it is a downside case. I built it from the observed five-year low ordinary margin of 7.18% D34 on a revenue base of 18,000,000千円 — the pre-bulge mean of 17,506,838千円 D15 plus the demonstrated annuity increment of 578,523千円 F1303F1352 — taxed forward E279, with zero deployable cash, because the net-cash position is one year old D82D83 and 1,644,147千円 of taxes payable sits against it F799. The whole distance between watch and buy is one parameter inside that item: re-run it at the pre-bulge mean margin of 11.76% D35D36D37 instead of the observed low, holding every other link fixed — the same revenue base, the same forward tax rate E279, the same 27,563千円 stock-compensation deduction F657F665, the same 5,554,665 diluted shares D1F1414 — and the identical chain gives ¥2,568, which is 3.6% below the stamp F1880, moving my threshold from ¥900 to roughly ¥2,570. One parameter. And a discrepancy I will state rather than reproduce: my own profile run prints ¥2,620 for this substitution at C104, a figure that drops the stock-compensation deduction the ¥1,549 chain applies and so is not the identical chain; ¥2,568 is what the chain as described here actually produces. That is why the item was put to a jury, and why its dispersion is the most informative thing in my file.

PABRAI. We are closer than the numbers look, then, because my 0.20-weighted branch is the same idea — margin reverting to the 第59期 trough of 7.72% D36 on pre-bulge revenue D15, giving about ¥1,703. Where we part is that I weight the trough at a fifth and you anchor on it. Both are defensible. Neither is the mean, and the mean is what the ledger's normalisation hands you.

LI LU (L2, the ten-year worst case; L13, projection honesty). And this is precisely where I get off. A defensible trough exists in the filed rows — 第59期 revenue of 15,974,648千円 F3, ordinary margin 7.72% D36, net income 898,610千円 F13, EPS ¥169.29 F38 — but that trough was set while the migration programme was already running. It is a floor for the old business, not for the post-programme business, and the deeper floor cannot be built from this archive at all. So when Claude anchors on the observed low and Warren anchors on the mean, both are choosing between two numbers that describe a company that no longer exists in the same form. At ¥2,664 F1880 the stamp is 15.7× the worst printed EPS (2,664 F1880 ÷ ¥169.29 F38). You are all pricing a base. I am telling you the base is unobservable.

---

On why Li Lu alone says too-hard — and it is a knowledge claim, not a value claim.

LI LU (L1, the ten-year knowledge bar). Let me name exactly what I could not know, because "too hard" without the sentence is a posture. Three variables decide the outcome and the record answers none of them. (i) The public segment's post-migration revenue base: answerable only from backlog, and backlog says +2.0% F342 against +58.7% revenue F388 with total forward work at 0.57× a year's sales D58. (ii) Reams customer retention and contract economics across the ~450 bodies F204: there is no disclosure at all — no contract term, no renewal rate, no bodies gained or lost, anywhere in the document. (iii) Whether the government cloud displaces or complements the company's own data centre: answered only by the aspiration of 併用, running alongside E71. And the enumerated risk section, seventeen headings deep, never names the completion of the programme that produced the year E105. The decisive variable cannot be answered from the public record. That is not a low value; it is an absent one.

MUNGER. And I looked at the same three gaps and did not call it too-hard, which is the disagreement worth staging. My test is whether the decisive variables are nameable, not whether they are currently known. They are nameable — post-bulge public-segment revenue reverting toward rather than below the pre-bulge band; TOPPAN's nomination and veto rights not tightening E138; a credible successor for the president E102E193 — and each resolves in a document with a date on it. A verdict of too-hard says the question cannot be posed. This one can. It just 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 the arithmetic settles it in my favour on my own item. Priced at worst case, the unknowns are not covered. The bulge contributed 3,027,382千円 of incremental net income D26; losing all of it returns you to 9.89× D8; a base below the pre-bulge mean puts you at 15.7× the worst printed EPS F38. At ¥2,664 F1880 the price requires the unknowns to resolve at least neutrally — and I cannot say they will. Charlie's watch and my too-hard are the same reading of the facts with a different rule about what to do when the decisive disclosure does not exist.

BUFFETT. For my part the reason I proceeded is narrower than either of you: the price is below tangible book of about ¥2,728 (equity 16,256,331千円 F840 less intangibles 1,116,092千円 F776 over 5,549,665 shares D1), the business earned positive ordinary profit in every year of the record F6F7F8F9F10, and five-year cumulative free cash flow is 8,547,690千円 D112. That is enough to write an appraisal and a buy price on. It was not enough to make the appraisal narrow, and my range spans 43% top to bottom for exactly the reason Li Lu names.

---

On the Claude lens's own scorecard: the biggest registered prediction was falsified, and the miss is the verdict.

CLAUDE (C1–C6, the outside view; C109, the surprise ledger). Report the misses as prominently as the hits, so here is the biggest one first. My figures-blind §1 predicted the standard programme-cycle trap — cheap on trailing, ordinary-to-full on normalised — with a normalised band of 13–22× and P(normalised P/E above 15×) = 0.60. The ledger printed 9.89× normalised D8, 3.35× reported D6 and 0.91× book D4. It is cheap on every basis, and the actual is below the bottom of my band. That is not a near miss on a decoration; it is the load-bearing registration of the file, and it is the difference between the pass I would have written on my own prior and the watch I wrote on the ledger. If the normalised multiple had come in at 16×, this would be a pass, and I would have been wrong.

CLAUDE, continued. A second miss, in the other direction. My register prediction was directionally right and far too timid. From the その他の関係会社 classification alone E49E50, with no figure in front of me, I put the two holders' combined stake at ≥30% with p 0.75, ≥40% with p 0.35, and either individually at or above the 33.4% blocking threshold with p 0.15. The actual is 55.2% — 39.3% F223 plus 15.9% F226 — with the larger holder alone comfortably above the threshold at which no special resolution can pass without it F589. I assigned 0.15 to a state that is simply true. And I understated the bulge on every axis: revenue +49.3% against +10–30% F377, public revenue +58.7% against +15–45% F388, the public share 80.5% against 55–75% D50, operating margin 22.50% against 11–17% D33, ROE 30.9% against 13–20% F55, net cash 12.85% of market capitalisation against 25–50% D85.

BUFFETT. The hits are worth naming too, because a scorecard that only reports failures is as useless as one that only reports successes. You registered, before seeing a figure, that the working capital would be struck at a receivable peak — and it is 141.6 days D99, growing 1.577× against revenue's 1.493× D100. You registered a drop-through of 30–55% with a ~40% central, and the operating measure against 第60期 — the base you named — is 40.87% F851F852F4F5. You registered the backlog family almost exactly F350F358F362F342. That is the pre-registration doing its job: the same instrument that scored those calls is what makes the normalised-multiple miss impossible to bury.

CLAUDE. And the honest reading of the miss is that I priced the class and ignored the register. I predicted a P/B of 1.0–2.2× with P(above 1.0) = 0.75 on the reasoning that this class does not usually trade below book, and wrote that a sub-book print "would itself be information — most likely about the float and the register rather than about the business." It printed 0.91× D4, and my own conditional was right. One day's turnover is 0.38% of the net market capitalisation D14, the top ten hold 69.01% F535, and other corporations alone are 58.34% of the shares F505. The discount is partly a liquidity and control discount, and I registered the mechanism while assigning it the wrong probability.

---

On the jury: three selves, one item, a 3× spread — and what that says about the company rather than the method.

CLAUDE (C33 and C102, jury-eligible; C108, one-directional consumption). Eighteen selves were dispatched across six items. Three came back divergent, and the largest divergence is the one that decides my threshold. On the downside valuation, three independent selves given the same rows and the same instruction returned ¥1,419, ¥893 and ¥476 — a 3.0× spread. Self A used the observed trough revenue of 15,974,648千円 F3 at the observed-low margin D34. Self B adopted the observed decremental margin instead — 69.6%, computed as (1,233,744 F8 − 2,507,109 F7) ÷ (15,974,648 F3 − 17,804,937 F2) — and deducted the unfunded pension of 1,024,760千円 F814. Self C marked the fixed-cost base up by the bulge-year depreciation step of 743,474千円 in the parent's (提出会社) cost-of-sales expense breakdown (1,458,143 F1661 less 714,669 F1660; the consolidated depreciation series runs 918,532千円 F938 to 1,609,203千円 F939), pushing breakeven revenue to within a few per cent of the observed trough. My own pre-jury answer was ¥1,549 — above all three. Under the one-directional rule the published threshold moves to ¥900 and the honest object becomes the band ¥476–¥1,549. Two other items diverged: three selves put the probability of the bull side of the central bet at 0.58, 0.62 and 0.35, so I publish 0.42 rather than the midpoint; and the count of resolvable-now unknowns came back 3 / 3 / 2, so I carry the largest, which is the most damaging to my own verdict.

MUNGER. Then say what that dispersion is evidence of, because the temptation is to treat it as noise in the instrument. It is not. Three copies of the same reasoner, given identical rows, sized this company's intrinsic value across a three-times range because the rows do not determine the answer — the answer turns on whether you treat 743,474千円 of bulge-year depreciation in the parent's (提出会社) cost-of-sales breakdown F1660F1661 as a run-rate or a run-off, and the filing does not tell you. That is a statement about the company's disclosure, not about the jury. A business whose base three independent readings cannot locate within a factor of three is one whose base is genuinely unresolved, and my verdict says exactly that in different words.

CLAUDE. I depart from the most cautious self at exactly one named assumption, and I will say where: Self C's depreciation step-up is contradicted by the intangible roll-forward, because software carrying value has fallen to 943,146千円 F772 with 1,074,883千円 of amortisation against 80,119千円 of additions on the parent's (提出会社) 附属明細表 F1840F1839 — that charge largely runs off within the forecast window rather than persisting. Because I depart from the most cautious self, I discharge the divergence through the verdict instead: no buy-below is issued at any price inside the band, and the verdict is held at watch. Convergence is cited as support for nothing.

---

On the moat: was it paid to dismantle itself?

LI LU (L14, the value bottleneck; L43, switching costs and the winner-take-all line). Here is the argument in its strongest form, and I want it stated without hedging because it is the most consequential idea in the study. A standardisation specification is, by construction, a portability specification. The moat here is customer-specific customisation accumulated across roughly 450 bodies F204 over about forty years F211 in a package covering thirty-plus statutory functions F210; the company's own stated mechanism is that in-house development lets it change the code fast when the law changes E31. Under the national standard, the state specifies the revision and every vendor receives the same specification. The differential responsiveness that was the moat is exactly what the programme abolished. Then add the second half: a government cloud supplies a common hosting substrate, which the company answers by running 併用 — alongside — rather than against E71; the company's own risk (1) response names joint use of systems among local public bodies as the residual threat E84; and risk (10) names intensifying price competition including overseas entrants, with the explicit risk that order unit prices fall beyond expectation E97. The company has been paid, once, to make its customers' data portable, and it has priced its own rebuild at 2,984百万円 F411 with 488百万円 spent F412 — expressly to strengthen competitiveness after the migration completes E145. That is what dissolving a switching cost looks like on a cash-flow statement.

BUFFETT. And here is the rebuttal, which I hold and which is not weak. A common specification cuts both ways, and the filing shows it cutting this company's way first. Standardisation is what let a Nagano regional vendor complete 168 migrations in a single year F382E129 and what makes the 特定移行支援 population — bodies whose incumbent missed the March 2026 deadline, now extended to end-FY2030 — addressable at all E58. The switching cost that actually binds municipal software is not code portability; it is data, workflow and staff training, and the migration year has just refreshed that binding for 168 customers on a replacement cycle measured in years. The forward composition supports it: software-development-and-system-provision backlog rose to 7,606,170千円, 137.9% of prior F357F358; the recurring information-processing line rose 16.6% to 4,068,640千円 F1303F1352 with its own backlog at 110.8% F354; the obligation tail beyond three years rose from 457,782千円 to 832,753千円 F1297F1298. And a vendor that did not commit the rebuild would be the one to worry about — 2,984百万円 spread across 2024年11月 to 2029年3月 F411 is the ordinary cost of a product generation. State the conversion rather than leaving it to the reader, since the two disclosures print in different units: 2,984百万円 is 2,984,000千円 on the statements' own basis, which is under half of one bulge year's operating profit of 6,296,461千円 F852 and about 1.4× a single normalised year's ordinary profit of 2,088,245千円 D20 — spread across four and a half years.

PABRAI. The tiebreak is not in either argument; it is in the price of being wrong. If Li Lu is right, this is a cigar butt and the mean is a ceiling; if Warren is right, the mean is a floor and ¥2,664 is a bargain F1880. I cannot distinguish them from this filing, so I do what the checklist tells me to: I put the commoditisation branch in the tree at a fifth of the weight, value it at the 第59期 trough margin D36, and demand a price at which that branch does not hurt me. That price is ¥1,450. Nobody in this room has resolved the moat question. Three of us have priced it and one has declined to.

MUNGER. One thing all four of us should concede plainly. The mechanism's existence is established; its size is not, and the filing will not print it. There is no retention rate, no contract term, no bodies gained or lost, no per-body fee — and the company's own risk section names state-led standardisation and joint use as channels of harm while never listing the programme's completion as a risk E84E105. A moat I can name but cannot measure is one I must take partly on faith, and faith is not what the ledger is for.

The bet, located

Strip the agreement away and one falsifiable proposition is left, and all five lenses land on it: does post-programme earning power settle nearer the pre-bulge three-year mean of 2,088,245千円 of ordinary profit D20, or nearer the alternating floor of roughly 1,233,744千円 F8? At the mean the stamp is 6.17× enterprise value to normalised ordinary profit D90 and 9.89× normalised earnings D8 — cheap. At the floor, on 第59期's printed EPS, the same stamp is 15.7× (2,664 F1880 ÷ ¥169.29 F38) — dear. The bull mechanism is on the record: the annuity line up 16.6% F1303F1352 with backlog at 110.8% F354, the obligation tail up F1297F1298, roughly a billion yen of software amortisation rolling off on the parent's (提出会社) 附属明細表 F1840F1839, and a fresh operating base of 168 newly migrated bodies F382 plus the FY2030 laggard runway E58. The bear mechanism is equally on the record: forward-work coverage down to 0.5713× D58, a fixed-cost step-up of 566,473千円 in the two least volume-sensitive parent cost lines — salaries 2,124,433千円 → 2,436,517千円 F1650F1651 and outsourced maintenance 1,914,849千円 → 2,169,238千円 F1656F1657 — equal to 27.1% of normalised ordinary profit D20, a 26.0% R&D cut in the peak year F1029F1030, and about 1,835,280千円 of named forward investment still to be paid — 1,991,376千円 of total planned outlay F469F471F472F473 less the 140,888千円 F470 and 15,208千円 F474 already paid against it. It resolves in the FY2027/3 yūhō, on a date, in a document anyone can read.

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 falsified normalised-multiple registration, which is reported here as prominently as the hits.

Self-distance note. The Claude lens holds one of the five verdicts compared above (watch, implied buy-below ¥900) and wrote this synthesis; it also built the reconciled figure table and the evidence ledger all five lenses consumed, and the red team ran on the same model family. That is an unusual concentration of authorship — the answerer, the ledger-builder, one of the five voters, and the adversary are the same system. Read the synthesis with that in mind. Three 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 largest registered claim scored as a failure; its jury of selves produced divergences on three items, every one of which moved a published number down; and the Munger run was executed on a different model, which is the only genuine cross-model check in this study.

Verdict accounting (fixed ex-ante)

• A buy-below-¥X verdict is price-falsifiable against the unadjusted stamp. Three lenses issued explicit thresholds — Buffett ¥1,600, Munger ¥1,900, Pabrai ¥1,450 — and the Claude lens publishes an implied threshold of ¥900 with a published band of ¥476–¥1,549. Li Lu issues no number, because too-hard is a statement that the decisive variable is absent from the record, not that the price is too high E105F204.
• pass / watch / too-hard are recorded but unscored in any future review. Four verdicts here are watch and one is too-hard; all named prices sit far below the ¥2,664 stamp F1880 and are the level at which each lens would revisit toward buy.
• 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 five-year window — 発行済株式総数 has been 5,837,200株 at every one of the five year-ends F111F115, with exactly one movement ever recorded on the capital-history table, 27,200 shares on option exercise in FY2013 F479E153 — so every per-share figure in this thesis is on the ordinary, unadjusted basis. The denominator used throughout is 5,549,665 shares net of treasury D1, being 5,837,200 issued F115 less 287,535 treasury F553, which is the filing's own BPS denominator of 5,549千株 F1406.
• Every profitability figure in this thesis states its basis, and the same price is quoted both ways. Reported (bulge): net income 4,478,269千円 F15, EPS ¥794.09 F40, ROE 30.9% F55, operating margin 22.50% D33, ordinary margin 22.49% D38, P/E 3.35× D6, EV/EBITDA 1.63× D89. Normalised (mean of 第58期–第60期): net income 1,450,887千円 D24, ordinary profit 2,088,245千円 D20, EPS ¥269.39 D7, ROE 8.93% D65, ROIC ex-cash 10.13% D76, P/E 9.89× D8, EV/normalised ordinary profit 6.17× D90. Basis-independent: P/B 0.91× D4. The dividend is stated both ways too — 5.26% on the full ¥140 D9F135 and 3.00% on the recurring ¥80 D10F566, the difference being the ¥60 sixtieth-anniversary commemorative component F141F567E163.
• Units are never mixed. The financial statements and most ledger rows are in 千円; the MD&A, the R&D programme table and parts of the capital-plan disclosure are in 百万円. Where both appear for the same fact they are stated separately, not converted — for example the receivable build of 3,971,671千円 in the cash-flow statement F957 and ¥3,971百万円 in the MD&A commentary F318E115. The two Reams-programme figures are on a non-additive basis and must not be summed: 2,984百万円 of total development cost F411 is the expensed R&D programme; 1,662,344千円 F469 is the planned intangible investment, of which 140,888千円 has been paid F470.
• This is a VOID study for prediction-vs-actual purposes only; the verdict accounting above is unaffected.

Red team

A consensus red team was dispatched — four of five lenses had converged on watch, a non-decline — to argue pass: that a watchlist slot spent on 3640 is a slot wasted. It concluded plainly that its own case does not defeat the consensus, and it did something better than that first.

Two of the leads it was handed were contradicted by the record, and it said so. Both corrections are recorded here before its case, because both narrow the bear argument this study started with.

Correction 1 — the order book does not show a bigger cliff than the normalisation admits. Group order backlog rose 2.5% to 16,159,613千円 F349F350; remaining performance obligations rose to 15,989,271千円 from 15,565,828千円 F1300F1299; and the within-one-year tranche is essentially flat at 13,233,495千円 against 13,254,575千円 F1292F1291. The 32.6% collapse F362 is confined to a single line — system-equipment sales backlog of 651,545千円 F361 — which is a bought-in resale line standing behind equipment revenue of 7,407,735千円 F374, while the software-development-and-system-provision book rose to 137.9% F358. This study's own stamp note leaned on the equipment-backlog figure as evidence of a visible cliff, and the record shows the fact was narrower than it looked. Recorded on the record, unedited.

Correction 2 — the bulge was not a terminal labour push. Consolidated headcount fell from 589 to 585 through a year of 49.3% revenue growth F84F85F377, and has fallen every year from 694 F81. The red team's own words: "a flat cost base delivering a 49.3% revenue year F377 is exactly the structure under which profit reverts to the pre-bulge mean rather than undershooting it — which is the assumption the consensus is relying on." That inversion cuts for the consensus, and the adversary wrote it down anyway.

Its three strongest surviving points, engaged by name.

1. "The cycle floor, not the cycle mean, is the right anchor for the year after a peak." The printed ordinary-profit series alternates — 1,242,594 / 2,507,109 / 1,233,744 / 2,523,882 / 6,294,427千円 F6F7F8F9F10 — and the two troughs sit within 0.7% of each other two fiscal years apart, 第57期 and 第59期 F6F8. FY2027 does not open in the middle of a cycle; it opens the day after the largest wave the company has ever ridden broke. On 第59期's printed EPS of ¥169.29 F38 the stamp is 15.74× (2,664 F1880 ÷ ¥169.29 F38); on 第57期's ¥171.75 F36 it is 15.51×. And a trough is worse now than it was then, because the equity it must be earned on has grown: 898,610千円 F13 against the 2026-03-31 common-share equity of 16,245,525千円 F1404 is a 5.53% return, against a company target of 10% F231. Partly conceded, and it is the point the synthesis adopts. The counterweight the red team itself supplies is real: the two prior troughs both occurred with the standardisation programme still ahead of the company, FY2027 opens with 168 newly migrated bodies to operate F382 and 62 standard-compliant family-register systems F383, the government has extended the deadline to FY2030 for bodies whose incumbent missed March 2026 E58, and the forward-work disclosures do not show the drain the trough anchor assumes F350F1300F1292. The honest resolution is the one the red team wrote: the floor is not the base case, but it must govern entry price, because the difference between 9.89× D8 and 15.74× is the difference between a bargain and a full price.

2. "The normalised base is overstated at both ends." On one side, a permanent cost step-up measured after the normalisation window closed: parent salaries and allowances rose 2,124,433千円 → 2,436,517千円 F1650F1651 and outsourced maintenance 1,914,849千円 → 2,169,238千円 F1656F1657, a combined 566,473千円 in the two least volume-sensitive lines, equal to 27.1% of normalised ordinary profit D20, with average parent annual pay up 9.5% to ¥8,054,485 F714F715 on a workforce with 19.6 years' average tenure and no union F713E246. On the other, a bulge margin flattered by reinvestment that did not happen: R&D cut 26.0% from 733,491千円 to 543,054千円 F1029F1030, intensity down to 1.94% D59 from 3.91% (733,491千円 F1029 ÷ 18,740,930千円 F4), and cash capex at 0.30× depreciation D115 against 0.77× the year before D116. "Normalising to 第58期–第60期 therefore compares the old cost base's profits against the new cost base's obligations." Conceded on the salary and maintenance lines — this is the sharpest genuine haircut to the normalised base in the study, and no lens applied it explicitly. Two qualifications the red team also supplies: most of the rest of the cost increase is variable and self-reversing (bonuses 347,246 → 496,172千円 F1654F1655, outsourcing fees 899,722 → 1,546,102千円 F1658F1659, subcontracting 583,335 → 936,186千円 F1662F1663); and the doubling of depreciation inside the parent's (提出会社) cost of sales to 1,458,143千円 F1661 from 714,669千円 F1660 — the consolidated series runs 918,532千円 F938 to 1,609,203千円 F939 — is the standardisation software amortising, not a permanent charge: software fell from 1,937,861千円 to 943,146千円 F771F772 with 1,074,883千円 of amortisation against 80,119千円 of additions on the parent's 附属明細表 F1840F1839, so roughly a billion yen of annual charge largely runs off. Those two forces point in opposite directions and are individually large against normalised profit D20. No checklist item made anyone compute them together; that gap is queued below.

3. "Two holders own the votes, the minority owns the residual, and the filing contradicts itself about it." 信越放送 holds 39.3% of the votes including 2.4% indirect F223F224 and TOPPANエッジ 15.9% F226 — 55.2% combined, against 36.92% and 15.86% on a shareholding basis F515F517 and 69.01% for the top ten F535. Against that: both shareholder risk paragraphs assert 「当社グループの事業戦略、人事政策及び資本政策等について、何ら制約等は受けておりません」 E87E90, while the 重要な契約等 section discloses that TOPPAN Edge holds the right to nominate one director candidate and a prior-written-consent right over any issuance that would dilute it below a set level — calling the governance effect 「軽微」 E138E142E143. And 【関連当事者情報】 reads 該当事項はありません for both years E288 against 836,489千円 of transactions with those same two holders — 126,818千円 F247 + 15,501千円 F249 + 586,806千円 F253 + 107,364千円 F255 — 2.99% of revenue F5, disclosed only under the 主要株主 format in the risk section E86E89E289. Fully conceded, and it is where four of the five lenses independently landed. The defences the red team wrote against itself are honest and partial: 836,489千円 against 27,987,254千円 of revenue F5 is small and is disclosed with counterparty, nature, amount and period-end balance for both years, which is more granular than a related-party note; 関連当事者情報 該当事項はありません is a scope judgement under the standard rather than a suppression; and the year's capital actions ran with the minority — a 4.63% buyback D141 and a board-resolved switch to a progressive dividend with a 3.0% DOE guide F575E160 that by construction cannot be cut. Good defences individually. Together they still describe a filer whose default under discretion is the legal minimum.

The red team's own verdict, in its words: "The bear case is real, and it is stronger than the consensus's own framing of the risk. But it does not defeat watch, and I will not pretend otherwise." Its reasoning: pass requires either that the price already reflects fair value on trough earnings or that a defect disqualifies the business outright, and neither holds at 0.91× book D4, 6.17× EV to normalised ordinary profit D90, with ¥342.33 a share of net cash D84, a dividend policy that cannot be cut F575, a working-capital unwind of roughly ¥4bn coming if revenue falls F318F957, and five falsifiable propositions the next two filings will settle. And its closing recommendation is adopted here on the record: "The consensus survives. What it should not survive unchanged is its confidence: the four watch verdicts should carry the cycle floor, not the cycle mean, as the number that governs entry." Read watch on 3640 accordingly — keep the name on the list, but let the cycle floor govern entry price and position size, and note that the floor and the mean differ by roughly 60% (15.74× against 9.89× D8) on the same stamp. The liquidity constraint belongs beside it: one stamp day's turnover is 0.38% of the net market capitalisation D14, and the strategic block against 21.4千株 of daily volume F1890 means a minority has neither voice nor a quick exit.

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 the five converge on the same document — the FY2027/3 yūhō — and on the same two observables: the post-programme ordinary-profit line and the forward book.

• Buffett (watch, buy below ¥1,600). Normalised earning power of 1,450,887千円 D24 does not survive the end of the standardisation cycle if 第62期 consolidated ordinary profit lands below the 第59期 trough of 1,233,744千円 F8 — the worst of the three pre-bulge years. That would show the installed base is not the annuity this case assumes; the ¥2,434–¥3,480 appraisal and the ¥1,600 level both come down and the study is re-run from the ledger. Leading indicator: the one-year remaining performance obligation at 2027-03-31 falling below the 13,233,495千円 booked at 2026-03-31 F1292.
• Munger (watch, buy below ¥1,900). Watch becomes pass if, over the next two annual reports, public-segment order backlog and billed revenue revert toward the pre-bulge band without a corresponding blow-up in SG&A F849F850, TOPPAN's board-nomination and dilution-veto rights are neither exercised nor expanded E138, and a named successor with an internal operating record is designated for the eighty-year-old president E193E102. Watch becomes exit if backlog stays flat while revenue reverts hard, or if normalised ordinary margin falls below its pre-bulge trough of 7.18% D34 for two straight years.
• Pabrai (watch, buy below ¥1,450). Either of these breaks the thesis and shows the ¥1,450 level far too low: (a) 第62期 consolidated ordinary profit at or above ¥4.0bn, roughly double the 2,088,245千円 pre-bulge mean D20; or (b) total order backlog at 2027-03-31 above ¥18.0bn against 16,159,613千円 F349, with public-segment backlog above ¥14.0bn against 12,891,858千円 F341. Either would show the programme left a structurally higher earning base rather than a pulse.
• Li Lu (too-hard, no number). Two disclosures would move this off too-hard: (a) a filed, repeated statement of the Reams customer base with contract term and renewal or churn — bodies under contract, weighted-average remaining term, and bodies gained and lost in the year — sustained across two consecutive filings F204; and (b) a stated post-migration public-segment revenue base, either as guidance or as two consecutive years of actuals settling at or above the 14,194,868千円 of FY2025 F1325 on a margin at or above 14.89% (2,112,947千円 F1332 ÷ 14,194,868千円 F1325). Conversely, if the FY2027 filing shows public-segment revenue below 14,194,868千円 F1325 with backlog again flat, the too-hard becomes a permanent pass rather than a suspended judgement.
• Claude (watch, implied buy-below ¥900). Three, each with a threshold and a named resolving document, plus a symmetric upgrade. (1) If the FY2027/3 yūhō shows 第62期 consolidated ordinary profit below ¥1,300,000千円 — the observed-low 7.18% margin D34 on a reverted revenue base — the through-cycle level is in the BELOW state and the verdict goes to pass. (2) If total remaining performance obligations at 2027-03-31 are below ¥12,000,000千円 against 15,989,271千円 F1300, with public-segment backlog below 90% of prior year F342, the forward book is draining structurally and the verdict goes to pass regardless of the profit line. (3) If either その他の関係会社 raises its stake such that combined voting rights exceed 66.7% F223F226, or either announces a tender below the ¥2,927.30 book value F35, the minority is a price-taker and the watch ends — uninvestable at any price this lens would set. Upgrade trigger, stated for symmetry: 第62期 ordinary profit at or above 2,088,245千円 D20 with the information-processing-and-communication line at or above 4,068,640千円 F1352 would discharge the annuity hypothesis on two independent measures. Review by 2027-06-30.

The single observable all five converge on is where 第62期 ordinary profit settles between the pre-bulge mean of 2,088,245千円 D20 and the alternating floor near 1,233,744千円 F8 — together with whether the one-year remaining performance obligation holds its 13,233,495千円 F1292. 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 — eight, of which three are structural. B43 needs a five-year fallback for Japanese filings: the item demands ten fiscal years, a yūhō's 主要な経営指標等の推移 prints exactly five F1F15, and answering data-insufficient loses the finding that actually mattered — the sawtooth within the five D34D35D36D37D38. Propose: "where the filing carries only five years, answer on five and say so; a swing of more than ~1.5× in margin between adjacent years within the available window is itself a stability failure, not merely insufficient data." B42 needs a fourth maintenance-capex route for capitalised-software businesses: the three implied routes all mislead in a harvest year, where capex ran at 0.30× depreciation D115 because the previous product generation is fully built, while the filer discloses the forward programme directly — 2,984百万円 to March 2029 F411 and 1,662,344千円 of planned intangibles F469. Propose: "where the filer discloses a dated forward development or capital plan, use it as an independent triangulation route and state it." B23's condition (3) is under-specified for public-sector vendors: "not subject to price regulation" reads as a test for utilities, but a monopsonistic public buyer procuring by open tender is functionally a price constraint with no rate-setting regulator E97E104. Also: B72 should name the TSR-versus-benchmark route explicitly, since Japanese filings print 株主総利回り beside 配当込みTOPIX for five years F182F191 and answer the one-dollar test in two rows; B99 requires a long-term government bond rate the ledger does not carry and is not derivable from it, so the item needs a named fallback or the ledger's source index needs a Tier-B rate row; B25 and B95 are unanswerable in a single-company archive and need explicit archive-depth fallbacks naming what filed self-evidence is accepted; v0.4.0 has no [PRACTITIONER]-tagged items while carrying a front-matter note describing the tag; and B7 can be tripped by prose order rather than by reasoning — a price cited solely to be rejected as a bulge-year artefact is not anchoring if the value estimate was computed independently of it.
• Munger — three, all method-shaped. M54/M56 need an explicit "denial-language density" sub-check: count the times a filing states 「何ら制約等は受けておりません」, 該当事項はありません or 「軽微」 within pages of a disclosure that substantively qualifies it. This study found four such instances in one document E87E90E288E311E143 — exactly the pattern the item exists to catch — but the checklist asks the analyst to notice it unprompted rather than naming the count as evidence in its own right. M52 assumes a multi-year buyback history, and this company executed a single lump-sum ToSTNeT-3 repurchase in one session E157; propose a single-event variant comparing execution price against the fiscal year's own high/low range, which is what this run improvised (¥2,995 D142 inside ¥1,342–¥3,940 F201F196). M9/M19 assume several years of comparable operating-margin history, and a JGAAP highlights table is not required to print 営業利益 beyond two years D32D33; the guidance should say that "five-year highlights" cannot be assumed to include operating margin, and the item should ask which years are actually computable before scoring.
• Pabrai — four, all failure-derived. P1 needs a "name the buyer" sub-test. The item's Good text presumes this hunting ground offers a hard net-cash floor, which is true of manufacturers and false here; propose name the market in which each floor asset would actually be sold, and mark to zero any asset whose only buyer is a going-concern acquirer of the same business. Rationale: buildings of 3,680,117千円 F754 are a purpose-built Nagano data centre E29 with no secondary market, and intangibles of 1,116,092千円 F776 are Reams itself — both would have been silently counted as "floor" by the item as written, overstating it by roughly ¥860 a share. New item P82, the policy-pulse test: is the peak year driven by a dated regulatory or statutory event whose deadline has already passed, and does the forward order book confirm or contradict the reversion? Revenue grew 49.3% F377 while backlog grew 2.5% F350 — the single cleanest forward signal in this study, and no existing item asks for the divergence. P12 should widen from "largest customer" to "common cause": the item passes on its letter here (no customer at 10% E122) while economic concentration is total at 80.5% public sector F244 on one national calendar E123. P39 should price the buyback, not just count it: 4.63% of the issued count repurchased D141 — and held in treasury rather than cancelled F553 — at ¥2,995 D142 against a ¥2,664 stamp F1880 is capital destruction wearing a cannibal's costume.
• Li Lu — three. L39 should name statutory and programme-driven demand explicitly: the item enumerates "acquisitions, one-off gains, subsidies, or a cyclical spike", and this study's dominant driver is none of those — it is a government-mandated migration with a statutory deadline, which behaves like a maturity rather than a cycle E123E58. L1 needs a sub-test naming the customer-base artefact, not just the variables: for a recurring-contract business the decisive disclosure is a specific artefact — customer count, weighted-average contract term, gross retention — and its absence should be a stated failure mode rather than something the analyst must argue for; the ledger gives 450 bodies F204 and 73.7% of public revenue in one product F207 and nothing else. L26 should require the combined blockholder position to be computed when the filing does not: two その他の関係会社 hold 39.3% F223 and 15.9% F226, presented only separately alongside a statement that there is no parent company E52, and the item as written is silent on a coalition never named as one.
• Claude — six checklist sharpenings, plus two library entries. (1) C3 should require each discharging hypothesis to name its measure and what the measure would show if the mechanism were absent — the over-time revenue share was registered as the annuity test and moves with the project cycle, so it fell F1321F1370 while the recurring line rose 16.6% F1303F1352; forcing the null case out at registration would have caught the construction error before the ledger opened. (2) C35 should compute the private-owner yield at two maintenance-capex anchors, not one — three jury selves and the lens agreed on the earnings base and disagreed on the yield by four percentage points entirely through that correction, on a business whose depreciation is 3.3× its cash capex D115. (3) C47 should run the deployable-cash waterfall on the prior balance sheet too whenever the current year is flagged as a one-off — net cash of 1,899,819千円 D82 against net debt of −1,295,980千円 twelve months earlier D83 makes a one-year artefact look structural on a single date. (4) A new §2 item, "amortisation roll-off and the replacement charge": where a capitalised software asset has fallen materially, compute the implied run-off against the named forward development programme and state the net P&L effect by year. Here software fell 1,937,861千円 → 943,146千円 F771F772 against a named 1,662,344千円 plan with 140,888千円 paid F469F470 and a 2,984百万円 expensed programme on a separate non-additive basis F411; both forces are large against normalised profit D20, point in opposite directions, and the red team's second point turned on exactly this pairing — no existing item makes anyone compute them together. (5) C98 should require the line "of the resolvable-now unknowns, which were not resolved, and was that a property of the company or of this study's archive?" — three of five load-bearing unknowns here were resolvable from free public sources absent from a two-document archive, which is a gap in the work rather than uncertainty about the company, and the two have opposite implications. (6) C91's citation-period histogram should be measured against what the archive can support, not against an absolute standard, or it penalises a one-filing archive as though it were a recency preference.
• Claude library (v0.1.0 → next). Class-level, N=1, no rate claimed: a Japanese regional public-sector systems integrator reporting the terminal year of a mandated national standardisation programme — domestically concentrated regional SI, in-house core municipal package, own data centre, more than 70% public-sector revenue, reporting the fiscal year containing the programme's statutory migration deadline. Datapoint on the class's key axis, the size of the terminal-year bulge: revenue 1.5986× the three-year pre-bulge mean D16, ordinary profit 3.0142× D21, net income 3.0866× D25, drop-through 40.87% on the operating line measured against 第60期 alone F851F852F4F5 and 28.9% on the net line measured against the 第58期–第60期 mean D29 — two different bases, because no 営業利益 is printed before 第60期, and a future study reading this entry must not treat them as a matched pair — and the forward book did not follow, total remaining performance obligations up only 2.7% F1299F1300 with coverage of revenue falling from 0.83× (15,565,828千円 F1299 ÷ 18,740,930千円 F4) to 0.5713× D58. Note for the class definition: the resolving observation — where earning power settles — is not yet available, so this entry records the bulge magnitude and the forward-book behaviour only, not the outcome. Case-level, so it never anchors a future outside view: "When a pre-registered test measures the wrong thing, retire the measure out loud before reporting the result the right measure gives." Source: this study — the over-time revenue share was registered as the annuity test, is contaminated by cost-to-cost development work, and fell F1321F1370 while the recurring line rose F1303F1352; the hypothesis is recorded as partly discharged, not discharged.
• A process note that belongs in the record. This study's stamp note framed the pick as the record's first non-industrial. That is an ordinal claim about the repository, and it does not survive a check — it fails at least three times over. Study #1 itself, 7564 Workman, is an apparel retailer and franchisor; 8200 Ringer Hut is a restaurant chain; and 4676 Fuji Media Holdings is a certified broadcasting holding company and a Tokyo-and-Osaka landlord. Not one of the three is a manufacturer, a distributor or a contractor. The framing is not repeated anywhere above, and the substantive question the pick was built to ask — whether a persistent watch is an artefact of the sector the universe screen keeps surfacing — is answered without it: an asset-light, switching-cost, high-return, non-manufacturing business drew four watch verdicts and one too-hard, so the finding was never about the sector. Rule adopted going forward: no ordinal or comparative claim about the record — "the first", "the cheapest yet", "the Nth" — enters a stamp, thesis or deck without being checked against docs/calibration.md and the repository first.

---

Corrections