TOKYO TEKKO CO., LTD. (5445): A Control Experiment in the Related-Party Note — the Moat Is Real, Unmeasurable, and Already in the Price

Verdicts

Lens	Verdict	Buy below	Most load-bearing items
Buffett	watch	—	B45+B42; B25; B43+B8
Munger	watch	—	M18; M46/M54/M83; M38/M43
Pabrai	watch	¥1,150	P5; P18; P53
Li Lu	watch	—	L1; L38+L40; L27+L18
Claude	watch	implied ¥430	C35; C40; C102

Five of five lenses reached watch — the first unanimous watch in the record, and the study that was built to break it. 東京鐵鋼 is pick #1 of the "cheap AND good" batch, screened precisely because twenty-two prior studies had produced zero active buys and the practitioner wanted to know whether the binding constraint was ever price. This name cleared the screen on every axis: 0.79× book D4, 6.27× trailing earnings D6, a 5.05% dividend yield D8, a 78.8% equity ratio F45, a three-year ROE record of 15.9 / 19.2 / 13.1% F48F49F50, a buy-and-cancel programme that retired 9.82% of the share count in one year D104, and a certified, labour-saving product that is not interchangeable with commodity rebar E12E39. It still drew only watch, from all five, and every named price sits far below the ¥1,980 stamp — Pabrai at ¥1,150, Claude at an implied ¥430, a spread of 22% to 58% of the stamp. So the constraint was never price alone.

Two findings decided it, and they pull in opposite directions — which is why the verdict is watch and not pass. The first is the study's sharpest object, and it is a control experiment the company printed almost by accident. Its 21.1%-held equity-method affiliate 株式会社伊藤製鐵所 F229 — another rebar maker, with which Tokyo Tekko conducts joint purchasing of raw materials E24 — saw revenue fall ¥33,360M → ¥29,419M F1409F1408 and net income go ¥1,297M → −¥595M F1413F1412 in the same fiscal year, on the same scrap-price surge E29 — an industry-wide one, and a year in which Tokyo Tekko's own reinforcing-bar shipment volumes fell too E62. In that identical year Tokyo Tekko held gross margin at 28.20% against 28.30% D26D25, earned an 11.13% net margin D28 and a 13.1% return on equity F50. Same input, same country, same twelve months, opposite outcome — printed annually, under audit, in the related-party note. It is what moved the Claude lens's figures-blind prior that this is a commodity cyclical rather than a durable franchise from 0.80 to 0.45, and it is why no lens reached pass. The second finding is the cash: ¥25,748M of cumulative five-year reported profit D94 became ¥5,121M of cumulative free cash flow D88 — one yen in five — with capital spending running at 2.80× depreciation D89 while machinery is 78.3% written down (¥46,678M of accumulated depreciation on ¥59,598M of cost F669F667) and unit volume fell anyway E62. Read on owner cash rather than on reported earnings, the 6.27× D6 becomes 9.7× the five-year mean, and the price stops being cheap. The moat is real and unmeasurable; the price already assumes it is real.

The business

東京鐵鋼株式会社 melts scrap and rolls it into the steel bars inside Japanese concrete. It was founded in June 1939 in Adachi-ku, Tokyo, making pig iron and cast iron; it began producing ordinary steel ingots in 1959, completed the Oyama steelworks in Tochigi in March 1969 and the Oyama rolling mill in May 1969, when bar production started E8. The Hachinohe steelworks in Aomori followed in February 1976 and its rolling mill in September 1981; the head office moved to Oyama in 1978, and the Oyama site was renamed the head-office plant E10. It listed on the Tokyo Stock Exchange second section in June 1971 and the first section in April 1974 E9, and sits on the Prime market today E18. Its business, stated plainly by the filing, is the manufacture and sale of small-section bar steel and reinforcing bar to two JIS standards — JIS-G3101 for general structural rolled steel and JIS-G3112 for concrete reinforcing bar — together with mechanical rebar splices E20. There are two reportable segments, 鉄鋼事業 (steel) and その他 (freight transport, plant maintenance and related services) E19E21; steel is 98.7% of external sales, ¥71,575M of ¥72,540M F1362F1364. Eight consolidated subsidiaries E170, one equity-method affiliate E24, 607 employees at the parent F196.

The one thing that separates it from a commodity mill is 「ネジテツコン」. This is a high-strength reinforcing bar with a screw-thread node rolled into it, sold together with the company's own dedicated mechanical couplers, and the splice method carries an evaluation from the Building Center of Japan — evaluation number BCJ-C1039 — obtained in February 1983 and printed as such in the company's own chronology E12. What the company sells with it is not tonnes of steel; it is ironworker hours. Management describes it exactly that way: against a serious labour shortage on construction sites, the NejiTetsucon bar and its dedicated couplers deliver labour saving and shortened construction periods at the site, replacing the lap-splicing and welding a shrinking workforce can no longer supply E39. That is the correct shape for a business inside a shrinking physical market — sell into the scarcity that is deepening, because labour is disappearing faster than concrete is. A building-code evaluation obtained in 1983 and still carried in the company's own history E12 is not nothing — but the filing nowhere states that the evaluation has been maintained since, and that silence is a gap in the moat claim rather than support for it.

But the builder does not write the cheque, and the accounts never show what the coupler earns. Three trading houses take 79.2% of everything shipped — 伊藤忠丸紅住商テクノスチール at 35.0%, 阪和興業 at 22.1% and エムエム建材 at 22.1% F275F279F283 — up from 77.8% the year before ((28,859 F1391 + 17,847 F1392 + 17,542 F1393) ÷ 82,593 F4), concentration rising in a down year. Every product is made to stock; because production is entirely 見込み生産, there is no order backlog at all E59. And the filing takes the single-product-category disclosure exemption: because one product category exceeds 90% of consolidated revenue, the product breakdown is omitted, and the geographic breakdown with it E210. So no ネジテツコン revenue, no coupler volume, no product-line price and no product-line margin is printed anywhere — for the entire asserted moat. The exemption is a function of the business mix, not of a disclosure choice, which means it will apply next year and the year after.

The end market is shrinking, and management says so in its own words. Demand for bar steel in reinforced-concrete buildings — the company's core field — is expected to keep declining on the back of population decline E30, and 建設需要の減少 is listed as risk (2) of the three risks in the entire filing E50E52. Against that, management's stated response is to raise the proportion of products less susceptible to market conditions — coupler splices and precast-concrete splices — in order to reduce spread risk and build a body that can post stable profit E49. That sentence is the closest thing to an internal admission of two tiers of economics anywhere in the document, and it is also, read carefully, a goal rather than an achievement.

The register has no controller and several competitors. There is no parent company E234 and no related-party transaction is disclosed at all E213. Individuals hold 55.12% F425, financial institutions 20.08% F420, and the top ten are 32.15% F452 — headed by a trust bank at 9.93% F432. Two of the ten are rival electric-furnace makers: 合同製鐵 at 5.45% F434, a business-alliance partner since November 1980 E11 and a reciprocal cross-holder E155, and 朝日工業 at 2.20% F440. All officers together hold 47千株 F543 — 0.56% of the 8,445,305 shares issued F394. The President is the Chairman's biological son E127, and a third Yoshihara is an executive officer over purchasing E128. These men run the company; they do not own it.

The numbers

The five-year record is not a straight line — it is a loss, a surge, and a visible roll-over. Revenue, 百万円: 66,089 → 79,229 → 79,617 → 82,593 → 72,540 F1F2F3F4F5. Ordinary profit: △644 → 4,944 → 11,412 → 15,059 → 12,040 F6F7F8F9F10. Profit attributable to owners: △4,724 → 3,657 → 7,887 → 10,853 → 8,075 F11F12F13F14F15. Return on equity: △10.5 / 8.3 / 15.9 / 19.2 / 13.1% F46F47F48F49F50, a five-year mean of 9.2%, with four of the five years — 第95期, 第96期, 第97期 and 第98期 — independently reproduced from net income over average equity D46D44D42D40; 第94期 has no recompute, because the 2021-03-31 equity base is not in the ledger. Book value per share, split-adjusted, compounded 1,563.99 → 2,506.34 F31F35; earnings per share △172.69 → 315.59 F36F40; the equity ratio strengthened 67.2 → 78.8% F41F45. FY2026 is the down-leg: revenue −12.2% D16, operating profit ¥12,042M against ¥14,676M, −17.9% F799F800D29, ordinary −20.0% D30, net −25.6% D31, with production value — stated at selling prices, on the filing's own note F262E58 — down 15.0% F263.

At ¥1,980 the arithmetic looks like a bargain, and five lines of it are the whole bull case. Market capitalisation net of treasury is ¥50,076M D2 on 25,290,786 shares D1; that is 0.79× book D4, 6.27× earnings D6, a 5.05% dividend yield D8, and — with net cash of ¥2,425M D59 — an enterprise value of ¥47,651M D64 at 3.27× EBITDA of ¥14,589M D67D66 and 3.96× operating profit D65. Return on invested capital excluding cash is 13.72% D57, 12.28% including it D56.

The margin is the bull case's strongest fact, and it rests on two observations. Gross margin was 28.30% in FY2025 and 28.20% in FY2026 D25D26 — a tenth of a point — through a year management itself describes as opening onto a sharp surge in the scrap price with chronically sluggish rebar shipments E29. Operating margin held at 16.60% against 17.77% D24D23; the steel segment's own margin fell only from 17.53% to 16.46% D34D32. The ordinary-margin series runs −0.97 / 6.24 / 14.33 / 18.23 / 16.60% D18D19D20D21D22. But the gross-margin series exists for two years only, because the earlier income statements are not in a single 有価証券報告書 — a limit of this study's one-document archive, not of the company's disclosure.

The balance sheet is strong, honest, and slightly rich — and it holds no reserve. Total assets ¥80,449M F707 against liabilities of ¥17,033M F743 and net assets of ¥63,416M F767, of which ¥63,388M is attributable to owners D37. Interest-bearing debt is only ¥4,747M D52 at average rates of 1.7% and 1.9% F1440F1445, with no bonds outstanding and final maturity in September 2033 E221; the ¥12,000M committed credit line is entirely undrawn F1076F1077. Interest cover on ¥91M of interest expense F816 is over 130×. And then the inversion that decides two of the five lenses: the filing's own note discloses that the current market value of the revalued land is ¥1,601M below its post-revaluation carrying amount F1074E185 — the classic Japanese hiding place, and here it is empty and then some, against ¥1,605M the prior year F1075. The listed securities are already marked: ¥4,163M carrying on ¥954M of cost, an unrealised gain of ¥3,209M with ¥920M of deferred tax already provided F1153F1184F1186F1294, all Level 1 F1179. The ¥4,463M of unlisted subsidiary and associate stock F1051 is a consolidated figure, and the note stating that such shares have no market price is the parent's, covering ¥1,915M of them E226 — the qualitative point holds for both, the note does not span the consolidated number; and the ¥4,453M affiliate stake F1385 sits in a company that just lost ¥595M F1412. Meanwhile ¥23,924M of assets, including factory-foundation property, are pledged against just ¥3,390M of secured borrowing E183.

The cash is where the reported earnings stop being earnings. Operating cash flow ran △5,104 → 6,879 → 12,089 → 8,183 → 5,181 F55F56F57F58F59; investing △3,358 → △1,630 → △4,562 → △5,596 → △6,961 F60F61F62F63F64. FY2026 free cash flow was negative ¥1,780M D85. Cumulatively over five years: ¥27,228M of operating cash D86, ¥22,107M consumed investing D87, ¥5,121M of free cash flow D88 against ¥25,748M of reported profit D94 — 19.9 sen on the yen. The operating line is not the problem: cumulative operating cash over cumulative net income is 1.06× D95. The investing line is. Capital spending ran at 2.80× depreciation on the cash-flow basis (¥6,973M against ¥2,489M D89F1002F939), 2.28× on the 設備投資 basis D90, and 2.57× the year before D91 — while the facilities section declares no notable plans for new construction and none for retirement E76. The two named projects were a ¥3,570M logistics centre and a ¥1,005M crane for moving bar bundles at the head-office plant F1876F1877E229 — handling, not demand. Machinery is 78.3% depreciated F669F667, buildings 62.5% F663F661. Cash and deposits fell ¥14,171M → ¥7,172M in the year F648F647, and cash equivalents ¥16,341M → ¥7,162M over two years — a ¥9,179M fall F72F74, of which ¥6,998M was drawn in FY2026 alone F1033. Net cash accordingly fell from ¥9,094M to ¥2,425M D61D59 — 4.8% of market capitalisation D63, ¥95.88 a share D62. This is not a reachable cash floor; it is a rounding error.

Working capital moved the wrong way in a down year. Receivables from contracts with customers rose ¥12,302M → ¥14,463M, +17.6%, while revenue fell 12.2% F1329F1328D16; receivable days went 54.4 — (7,774 F650 + 4,527 F652) × 365 ÷ 82,593 F4 — to 72.8 D78, inventory days to 91.0 D76, payable days to 26.0 D79, and the cash conversion cycle to 137.8 days D80 from a recomputed 107.7 the year before, being prior-year inventory days of 77.8 ((9,995 F654 + 2,633 F656) × 365 ÷ 59,216 F774) plus those 54.4 receivable days less 24.5 payable days (3,974 F710 × 365 ÷ 59,216 F774). Inventory write-downs charged to cost of sales rose ¥125M → ¥341M, +173% F1082F1081. The mitigating detail is that the receivable build is almost entirely a settlement-instrument shift: 電子記録債権 went ¥4,527M → ¥6,721M F652F651 while ordinary trade receivables were flat, ¥7,774M → ¥7,741M F650F649, and the doubtful-accounts allowance was unchanged at ¥22M F702F701.

Capital return is real, recurring, and — this year — funded from the balance sheet. Dividends paid ¥3,238M F1024 plus buybacks ¥1,502M F1022 returned ¥4,740M D96, 58.7% of net income D97, against 35.2% the prior year D99. Buybacks were resolved in May and again in October of the year under review, both executed on-market E87E88E90, and 920,000 shares were cancelled on 25 March 2026 — 9.82% of the opening count F1135D104E92, taking shares issued from 9,365,305 to 8,445,305 F1134F1136; the net float fell 26,082,588 → 25,290,786, −3.0% F1429F1428. A further ¥500M / 350,000-share authorisation followed on 7 May 2026, 1.38% of the count, for shareholder return and capital efficiency F1432F1430F1431E215. Against that: the dividend per share was cut from ¥375.00 to ¥300.00 F128F129 — ¥100.00 on the post-split basis D7 — while the payout ratio barely moved, 29.7% → 29.3% F156F157, because the policy is a stable distribution backed by results with no floor, no DOE and no progressive undertaking E93E94. And the ¥4,740M went out of a shrinking cash balance in a year of negative free cash flow, part-funded by ¥1,400M of new long-term borrowing F1017.

Finally, the re-rating that has already happened. Total shareholder return ran 71.8 → 372.4 over the five years against 202.2 for TOPIX with dividends F168F172F177. The filing's own printed price-earnings series is 4.3 / 5.9 / 4.6 / 6.3 F51F52F53F54; the ¥1,980 stamp is 6.27× D6 — sitting at the top of the company's own printed record — and at 92.96% of the year's split-adjusted high D12. Across that same window earnings roughly tripled — EPS 135.09 to a peak of 412.92 F37F39, net income 3,657 to 10,853 F12F14 — while the printed multiple expanded only from 4.3 to 6.3 F51F54, some 47%. A modest re-rating against a tripling of earnings is a considered judgement about durability, not neglect.

The five lenses

Buffett — watch

Start where a shopkeeper starts: what does this outfit sell, who writes the cheque, and why do they keep writing it? Tokyo Tekko buys scrap, melts it in an electric furnace and rolls it into reinforcing bar for concrete buildings E20 — at Oyama since 1969, Hachinohe since 1976 E8E10. Its one distinguishing item is 「ネジテツコン」, a bar with a screw thread rolled into it, sold with the company's own mechanical couplers under an evaluation from the Building Center of Japan obtained in February 1983 and recorded in the company's own chronology E12. The pitch is not tonnes of steel; it is ironworker hours saved on a site that cannot hire ironworkers E39. An evaluation a company was granted that long ago is not nothing — though I notice the filing never says it has been kept in force since, and a man ought to notice what a document does not say.

Now the part that gives me pause. The builder does not write the cheque. Three trading houses take 79.2% of everything shipped F275F279F283E60, every product is made to stock, and there is no order book E59. When I ask what the threaded product earns, the filing need not say: one category is over 90% of sales, so the breakdown is omitted E210. The moat is asserted in the narrative and invisible in the accounts — data-insufficient, and I do not get to fill that hole with enthusiasm. The franchise test has three legs: needed, no close substitute, not price-regulated. One and three pass. Two fails on the company's own paper — JIS-standard bar steel E20, with two shareholders on the register in the same trade, Godo Steel at 5.45% F434 and Asahi Kogyo at 2.20% F440. And the moat has a dated test it did not survive: in FY2022 scrap ran, and ordinary profit went to −¥644M and net to −¥4,724M F6F11. A franchise absorbs a cost shock. This one did not.

So the case rests on figures — and on the maintenance-capex guess, which decides everything. Start with reported profit of ¥8,075M F15. Add depreciation of ¥2,489M F939, amortisation of ¥58M F942 and the ¥228M impairment F941. Now subtract the capital genuinely required to hold position and unit volume. Three ways at it: depreciation itself, ¥2,547M; the five-year average investing spend, ¥22,107M over five years D87, or ¥4,421M a year; and the named projects — consolidated FY2026 capex of ¥5,801M F315 included a ¥3,570M logistics centre and a ¥1,005M warehouse crane F1876F1877, plainly expansion, leaving ¥1,226M of everything else, below depreciation and so not credible. Mind the scope in that last one: the two named projects are parent-company (単体) rows from the fixed-asset schedule, set against a consolidated capex total, so the residual is indicative rather than exact. I take maintenance capital at ¥3.5bn–¥4.4bn, putting FY2026 owner earnings at ¥6.4bn–¥7.4bn. Why the high end? Capital spending ran 2.80 times depreciation this year and 2.57 times last D89D91. Machinery is 78.3% written down — ¥46,678M of accumulated depreciation against ¥59,598M of cost F669F667 — so the replacement bill is ahead of this company, not behind it. Unit volume fell anyway; falling rebar shipments are management's own stated reason for the year's shortfall E62. And the number I keep returning to: over five years the company reported ¥25,748M of profit D94 and produced ¥5,121M of free cash D88. One yen in five. Depreciation is not what it costs to stay in this business. Charlie and I paid tuition on that lesson in textiles.

See what that judgement does. The enterprise is ¥47,651M D64. Against EBITDA of ¥14,589M D66 that is 3.27 times D67 — a number that makes a man reach for the telephone. Against mid-cycle owner earnings of roughly ¥4bn it is twelve times. Same company, same day, same enterprise value; one figure is arithmetic, the other a fable, and the gap is the depreciation the tooth fairy is meant to pay. Now value it. Five-year average ordinary profit is ¥8,562M F6F7F8F9F10 — about ¥5.9bn after tax at the printed 30.5% rate F1816 — against five-year average net income of ¥5,150M D94. Take ¥5–6bn, add ¥2.5bn of depreciation F939, subtract ¥3.5–4.4bn of maintenance capital: mid-cycle owner earnings of ¥3.5bn–¥4.5bn. Capitalise at eight to ten times, what a cyclical selling into a shrinking market deserves E30E50: ¥28bn–¥45bn. Add the separable pieces — net cash ¥2,425M D59, listed cross-holdings ¥4,163M less ¥920M of booked deferred tax F1153F1294, and the ¥4,536M unlisted book F1163 at a thirty per cent haircut — for ¥8.8bn. Whole company ¥37bn–¥54bn: ¥1,450–¥2,130 a share on the 25,290,786 shares net of treasury D1.

Two cross-checks before I look at a quote. Net current assets after all prior claims are ¥19,722M — ¥36,783M of current assets F659 less ¥17,033M of liabilities F743 less ¥28M of minorities F765 — about ¥780 a share, no Graham bargain. A Dempster haircut, with cash at par, receivables at 85%, inventory at 60%, plant at prompt-sale value and land at market, which the filing puts ¥1,601M below carrying value F1074, nets ¥31.4bn after the claim stack, roughly ¥1,240 a share. There is no asset floor here. Only now the price: ¥1,980. Upper third of my range, above every asset test, no margin of safety. What makes it look cheap — 6.27 times earnings D6, 0.79 times book D4, a 5.05% dividend D8 — is a low multiple on a near-peak year, and I cannot name a transient cause for the discount. Total shareholder return went from 71.8 to 372.4 in five years against TOPIX at 202.2 F168F172F177. The market is not ignoring this company; it is pricing a shrinking end market, and on the record it is right.

Three things before we leave it. The company returned ¥4,740M to owners D96, 58.7% of profit D97, while free cash flow was negative ¥1,780M D85 and net cash fell from ¥9,094M to ¥2,425M D61D59 — paid out of the balance sheet, not out of surplus. Second, ¥184M of the ¥385M director pay pool F574F572 turns on the consolidated ordinary-profit margin, and the section the filing points to for that metric's target and actual states neither, anywhere E149E144; the chairman chairs the committee setting the pay of the president, his son E99E127, each taking ¥61M of performance-linked pay on that undisclosed hurdle F585F589. Third, management retired 9.82% of its shares in a year D104 while selling not one of thirteen listed cross-holdings worth ¥4,129M E153E152. I like the first habit and not the second. What is good here is real: equity 78.8% of assets F45, no bonds E221, ¥12bn of committed lines undrawn F1076F1077, accounting conservative to a fault — actuarial swings expensed as they arise E178, a 1.5% expected pension return against a 32.8% equity allocation F1258F1248. R&D rose ¥274M → ¥320M in a year operating profit fell 17.9% F794F793D29. And management says plainly that its core market keeps shrinking with the population E30E50. That is confession time, and I respect it.

Verdict: watch, with no buy-below written. An adequate business, honestly financed and decently run, in a market that gets smaller, at a price that is fair rather than wonderful. The arithmetic turns interesting near ¥1,200, where the earnings case and the haircut value meet — but I will not write that as a trigger, because two things I need are missing: a ten-year record the archive does not hold, and the coupler economics the company will not disclose E210. There is no called strike for standing here with the bat on my shoulder.

What a student should take from this: depreciation is what the accountant charges; maintenance capital is what the business actually costs. When the machinery is 78% written down F669F667, capital spending runs at nearly three times depreciation for two straight years D89D91, and unit volume falls anyway E62, the reported earnings are a claim and the free cash is the fact — here twenty sen of free cash per yen of five-year reported profit D88D94. And a differentiated product you cannot find in the segment tables is a story, not a franchise: the moat you cannot measure is a moat you must not pay for.

Munger — watch

Invert it first. How does 東京鐵鋼 die inside ten years? Four ways, all visible in its own filings, none of them exotic. One: its main end-market — reinforcing bar into reinforced-concrete buildings — is disclosed by management itself as being in structural decline on the back of population shrinkage, with no claim that this reverses E30E50. Two: three intermediaries account for roughly 79% of sales — 35.0%, 22.1% and 22.1% F275F279F283 — and the risk-factor section, all three headings of it, never says so E52. Three: the whole machine runs on a scrap-price-to-selling-price spread that the auditor itself flagged as the load-bearing assumption behind the deferred-tax asset, precisely because it swings E241E49. Four: the equity affiliate it committed capital to in 2018 E17 just swung to a ¥595M net loss E214. None of these kills the company by itself — the balance sheet has net cash D59, and the security note shows ¥23,924M of assets pledged against only ¥3,390M of borrowings E183, which is a small obligation however large the collateral behind it — but a business whose own management concedes its core market is shrinking is not a business I need to strain to find fault with. The filing does that work for me.

What is it, underneath the inversion? An electric-arc-furnace scrap melter that also sells NejiTetsucon — screw-threaded rebar with proprietary mechanical couplers, carrying a Building Center of Japan evaluation obtained in 1983 and listed in the company's own chronology E12. That is a real mechanism — regulatory approval plus a switching cost baked into the coupler system — not a story. Though note what the filing does not say: nowhere does it state that the evaluation has been maintained since, and an unstated renewal history is a gap in the mechanism, not evidence for it. But I cannot find its fingerprint in the figures. There is one steel segment, no product-line split E210, and gross margin barely moved, 28.3% to 28.2% D25D26, across the two years I can see. A moat I can name but cannot measure is a moat I have to take partly on faith, and faith is not what the ledger is for.

Incentives next, because that is where character shows itself before anyone has to lie about it. The variable half of director pay is set against the consolidated ratio of ordinary profit to net sales — and the section the filing cross-references for that ratio's target and actual states no such number, anywhere E144E149. That is not a rounding error; it is a KPI nobody outside the room can check, driving ¥184M of a ¥385M pool F574F572. The President is the Chairman's son E127, a third family member sits as an executive officer E128, and the committee that recommends both men's pay is chaired by the Chairman himself E99. None of that is disqualifying by itself — family firms compound too — but stacked on top of an unauditable pay metric, it is the kind of thing I write down rather than wave past. Layer in the filing's own loose threads — outside-director headcounts that do not agree with each other E98E132, two directors who vanish between tables with no stated reason E112E126, "material contracts: nothing to report" E72 sitting one page away from two real alliances and a ¥12,000M credit line E11E17E67F1076 — and you have more than one complexity flag. More than one shifts the burden of proof onto the company, and this one has not met it. None of it looks like fraud. It looks like a thin corporate secretariat that has not been made to answer for its own filing, which is its own kind of information.

What I like: the capital discipline is real. Nine hundred twenty thousand shares — 9.8% of the count — were cancelled in one year F1135D104; dividends and buybacks together returned 58.7% of net income D97; the dividend was cut in step with lower earnings rather than defended with debt F128F129; buybacks were executed below the year's high and funded from net cash E87E88D12. That is a management doing arithmetic on behalf of owners, not managing a narrative. What I do not like: free cash flow was negative this year on capex running at 2.8 times depreciation D85D89, and cumulative five-year free cash flow covers barely a fifth of cumulative net income D88D94 — the plant eats a great deal of what the income statement reports. On price: P/E 6.3×, P/B 0.79×, dividend yield 5.05% D6D4D8 — statistically cheap by every measure I have. But the stock already ran from a low of ¥1,151 to a high of ¥6,780 pre-split inside this same five-year window F185F181, badly outrunning the index the whole way F172F177. Cheap trailing multiples after a violent re-rating, on earnings already 25.6% off their peak D31, is not the same animal as a bargain the market has not yet noticed. It may just be the correct price for a cyclical coming back down from the top of its own cycle.

Verdict: watch, and I issue no number. A genuine but unmeasurable moat, inside a shrinking pond, run by disciplined capital allocators who also run an unaccountable pay metric and a filing full of small unexplained contradictions. That is not a pass — there is real quality here — and it is not too hard, because the book itself is clean and auditable. It sits on the watch pile until the figures, not the prose, tell me the coupler business is actually widening the margin against the cycle, or until the governance loose ends get tied.

What a student should take from this: a moat you can name but cannot measure in the figures is a claim, not yet a fact — price it as uncertainty, not as a discount waiting to be captured. A compensation metric with no printed target or actual anywhere in the document E149 is worth exactly what an unverifiable promise is worth, regardless of how modest the pay looks in yen. And statistical cheapness that arrives right after a five-fold re-rating on cyclical earnings F185F181 deserves the same suspicion as statistical cheapness that arrives after a crash — check which side of the cycle you are actually standing on before calling it a bargain.

Pabrai — watch, buy below ¥1,150

I have one rule that has saved me more money than any insight I ever had: do the downside first, and do it before you allow yourself to get interested. The business takes four sentences, which is a good sign — complexity is a reason to walk away, not a puzzle to solve. Tokyo Tekko melts ferrous scrap in an electric arc furnace and rolls it into small-section bar steel and rebar to two JIS standards E20. Its one differentiated product is ネジテツコン, a screw-ribbed rebar sold with proprietary mechanical couplers, carrying Building Center of Japan evaluation number BCJ-C1039, obtained in February 1983 per the company's own chronology — the filing does not say whether it has been kept in force since E12. What it actually sells is not tonnes of steel but hours of labour saved on a Japanese construction site that cannot hire people E39. And it sells them through three trading houses that between them take 79.2% of revenue F275F279F283. Steel is 98.7% of it — ¥71,575M of ¥72,540M external sales F1362F5. That fits on a napkin.

Now the floor. At ¥1,980 the market is asking ¥50,076M D2. What comes back if everything goes wrong? Net cash is ¥2,425M D59 — ¥95.88 a share D62, 4.8% of the price D63. That is not a cushion; that is a rounding error. Book gives you 0.79× D4, and this is exactly where most people stop and call it cheap. Do not stop there. Here is the thing I want a student to carry away from this study. In this hunting ground you learn to expect the Japanese balance sheet to hide value — land at 1960s cost, securities at nothing, a cash pile bigger than the market cap. This one does the opposite. The filing tells you, in its own note, that the current market value of the revalued land is ¥1,601M below its post-revaluation carrying amount F1074E185. The land is carried above market. The listed securities are already marked to fair value at ¥4,163M F1153, so their ¥3,209M unrealised gain F1186 is in the book, not hidden behind it. The ¥4,463M consolidated book of unlisted subsidiary and affiliate stock F1051 holds a fellow electric-furnace rebar maker that just swung to a ¥595M net loss F1412; it is unlisted, and the parent's own note declines to state a fair value for its ¥1,915M of such shares E226. And ¥23,924M of assets, including the factory foundation itself, are pledged against ¥3,390M of secured borrowing E183. Mark the receivables down a tenth, the inventory a third, the affiliate stake by half, the land to market, and the mill machinery to what a used rolling line actually fetches, and I get roughly ¥33.5bn after all liabilities F743 — about ¥1,325 a share against ¥1,980. The floor is a third below me. That is a soft floor, and no upside story rescues a soft floor.

So what is the market afraid of? Rebar shipment volumes falling E62, scrap spiking in the second half E29, earnings down 25.6% D31. Fine — that is nameable, which is what I want. But then you read management's own sentence: demand for bar steel in reinforced-concrete buildings, their core field, is expected to keep declining because the population is declining E30, and they list falling construction demand as one of only three risks in the entire filing E50E52. That is the company telling you the pond is draining. Cyclicality I will pay for. Secular decline I will not — and here the fear and the fact point the same direction, which is the opposite of the setup I hunt.

Two things stop me calling it a pass. First, the balance sheet is genuinely safe. Net cash, no bonds at all E221, ¥4,747M of debt at 1.7–1.9% D52F1440F1445, with ¥7,172M of cash F647 and ¥12,000M of committed lines with nothing drawn F1076F1077. It went through FY2022 with an ordinary loss of ¥644M, a net loss of ¥4,724M and negative ¥5,104M of operating cash F6F11F55 and came out with a 67.2% equity ratio F41. Leverage will not kill this. Second, and more interesting: the moat may be real. Revenue fell 12.2% and gross margin barely moved, 28.30% to 28.20% D16D25D26; the steel segment still earned 16.46% D32. In the same year, its 21.1% affiliate — the company it jointly buys scrap with E24F229 — lost ¥595M on ¥29,419M of sales F1412F1408. Same scrap, same market, opposite result. When your competitor catches pneumonia and you have a mild cold, that is the low-cost operator's signature, and you should trust it more than any management slide.

Where I part company is on price and on alignment. Normalise the five years and ordinary profit averages ¥8,562M F6F7F8F9F10 against management's own ¥7.0bn floor F230 — meaning FY2026's ¥12,040M is 72% above the floor, not below it F10. Put eight times normalised net income on a shrinking end market — ¥5,874M at the 31.4% effective rate D58, about ¥232 a share on the net-of-treasury count D1 — and add the net cash D59 and I get roughly ¥49bn; five times normalised EBITDA gets me ¥58bn. Conservative intrinsic value is ¥1,950–2,300 a share. The price is ¥1,980. This is a fairly valued 5.05% yielder D8, not a fifty-cent dollar. And the alignment is salary-first: the Chairman drew ¥115M F583 against a stake of 30 千株 pre-split F536 worth about ¥178M at the stamp; his son, the President, drew ¥117M F587 against 4 千株 F537, about ¥23.8M — a fifth of one year's pay. All officers together hold 0.56% of the company F543. Meanwhile 48% of director pay hangs on a consolidated ordinary-profit margin whose target and actual appear nowhere in the document F574F572E149. A metric you cannot check is not an incentive.

Verdict: watch, buy below ¥1,150. It is a well-run, unlevered, shareholder-friendly business — 58.7% of earnings returned and 9.82% of the shares cancelled in one year D97D104 — trading at fair value in a market its own managers expect to shrink for a decade E30. I want it at ¥1,150, which is half of conservative value and below my own break-up estimate. There I am paid an 8.7% dividend to wait for the coupler mix to prove itself. At ¥1,980 I am paid nothing for being wrong.

What a student should take from this: a low price-to-book is a question, not an answer — go and read what the book contains before you treat it as a floor. Here the filing itself says the revalued land is carried above its market value F1074, the securities are already at fair value F1153, and the mill is pledged E183: the 0.79× D4 is arithmetic, not protection. And when a company tells you in plain language that its core end market shrinks with the population E30, believe it, and put that in the price rather than in a footnote — cyclicality you underwrite, secular contraction you discount.

Li Lu — watch

I begin where the country begins, because this business cannot be read apart from it. Japan is the first large modern economy to run its demographic arc backwards, and reinforcing bar is the most literal expression of that arc: every tonne of it is a wager that someone will pour concrete. The company says so itself — demand for bar steel in reinforced-concrete buildings, its core field, will keep declining as the population declines E30 — and the risk section repeats it E50. I have seldom read a filing in which management states the terminal problem of its own market so plainly. That candour is worth something. Against that arc the company sets one idea. It does not sell tonnes; it sells the labour a construction site no longer has. NejiTetsucon is a screw-node reinforcing bar joined by proprietary mechanical couplers, evaluated by the Building Center of Japan under evaluation No. BCJ-C1039 in February 1983, an event the company's chronology records and whose continuation the filing nowhere confirms E12, and management frames it precisely as a response to a serious labour shortage at building sites E39. That is the correct shape for a business inside a shrinking physical market: sell into the scarcity that is deepening.

Does it work? The best evidence in the file is not a claim; it is a control experiment the company printed almost by accident. Its 21.1% equity-method affiliate, 株式会社伊藤製鐵所, is another maker of reinforcing bar F229E24. In the year just ended that company's revenue fell to ¥29,419M and it lost ¥595M at the net line F1408F1412. In the same year, the same industry, the same scrap surge E29, Tokyo Tekko earned an 11.13% net margin D28 and a 13.1% return on equity F50; its steel-segment margin fell only from 17.53% to 16.46% D34D32 while revenue fell 12.2% D16; its consolidated gross margin held at 28.20% against 28.30% D26D25. One house holds its margin to a tenth of a point; the other goes to a loss. That is not industry economics but company-specific advantage — the sort of proof one travels a long way to obtain.

So why can I not own it. The knowledge bar gates everything, and here I cannot clear it. Three variables decide this company's earnings power ten years out: the volume of RC-building reinforcing bar in a depopulating Japan; the sales spread — selling price less scrap, the whole operating engine of an electric furnace; and the share of revenue and of margin carried by NejiTetsucon and its couplers rather than by commodity bar. The first is disclosed directionally and never quantified E30E50. The second is named by the auditor himself as a key assumption inside the five-year profit plan on which the deferred tax assets rest E241 — and no figure for it appears anywhere in the document. The third does not exist as a line item: product disclosure is omitted because a single category exceeds ninety percent of sales E210, and the steel segment alone is 98.7% of external revenue F1362F1364. Two of three decisive variables cannot be answered from the public record. When that is true the honest word is not "cheap." It is that I do not know.

Nor am I paid for not knowing. At the stamp the whole business costs about ¥50,076M net of treasury D2 against ¥63,416M of book F767 — 0.79 times D4. In the Korean case that taught me this test, the price was back inside the balance sheet; here I open it and find the reverse. Net cash is only ¥2,425M D59, under five percent of the market capitalisation D63. The listed cross-holdings are already at market — ¥4,163M carrying, ¥4,163M fair, on ¥954M of cost F1153F1184F1186 — so no hidden markup remains. The affiliate stake is at book, ¥4,453M F1385 against 21.1% of ¥21,899M of net assets F1406, and that book is now losing money F1412. And the land, the classic Japanese hiding place, is carried ¥1,601M above its market value F1074E185. The book is honest, slightly rich, and holds no reserve. A discount to an honest book is not a dollar for fifty cents; it is a dollar for eighty, with the eightieth cent resting on a margin I cannot underwrite.

The Asian structure reads mixed. There is no listed parent E234 and no related-party transaction at all E213; the register is dispersed and mostly retail, 55.12% individuals F425, top ten 32.15% F452. The company does not dilute — no options, no rights plan E80 — and it retired 920,000 shares, 9.82% of the opening count, in one year F1135D104 while returning 58.7% of earnings D97. But ¥4,199M of policy cross-shareholdings — ¥70M unlisted plus ¥4,129M listed E152, 6.6% of owners' equity D37 — sits entirely unmoved, not one issue increased or decreased E153, the board confirming the appropriateness of holding while stating that the benefit cannot be quantified E160E154. And the words PBR and cost of capital never appear in the filing; the sole capital-efficiency objective is an ROE floor of 10.0% E28F231, cleared at 13.1% F50. A company doing several right things without having yet decided why. Then who is doing them. The president is the chairman's son E127; a third Yoshihara is an executive officer over purchasing E128. A family running a business across generations is often the best arrangement there is. But this family does not own the business: all officers together hold 47 千株 F543, 0.56% of the 8,445,305 outstanding F394; the chairman's 30 千株 F536 are worth roughly ¥178M at the stamp against ¥115M of annual pay F583; the CFO holds none F541; and none of the four outside directors E126 holds a share — 所有株式数 ― against each of them E122E123E124E125. These men are paid. They are not owners.

Finally the reinvestment, where I judge management. Capital expenditure ran at 2.80 times depreciation D89 into a market management says will shrink, and what it bought was a logistics centre at ¥3,570M and a warehouse crane at ¥1,005M E229F1876F1877 — handling, not demand. Free cash flow was negative ¥1,780M D85, and cumulative five-year free cash flow is ¥5,121M against ¥25,748M of earnings D88D94. Spare capacity at Oyama is falling and NejiTetsucon output is being pushed to Hachinohe, where regional demand is itself declining E33. The runway inside the existing footprint is nearly used.

Verdict: watch, with no buy-below. An advantaged operator inside an industry that is not, at a price discounting an honest book but presuming a margin nobody outside the company can verify. I will not call it too-hard, because the obstacle is not the nature of the business — it is a disclosure the company could remedy tomorrow. I call it watch, and say what would turn it into a study: a published line for NejiTetsucon and coupler revenue and margin, and a spread series. A conservative anchor sits near ¥1,250 — half of a land-adjusted book (¥2,506.34 F35 less ¥1,601M F1074 over 25,290,786 shares D1) and about ten times mid-cycle earnings on the 6.24% ordinary margin of 第95期 D19. At the stamp I keep reading.

What a student should take from this: when a company's whole case rests on one differentiated product, ask first whether any disclosed line isolates that product's revenue and margin — and if none does E210, the moat claim is unfalsifiable however good the aggregate numbers look. A low multiple on visibly strong trailing figures is not automatically an inefficiency: the multiple here sat at 4.3 / 5.9 / 4.6 / 6.3× F51F52F53F54 across a 372.4% five-year total return F172, which is the market pricing durability, not overlooking the company. And when you re-mark a balance sheet, re-mark it in both directions; land carried above its market value F1074 is a hidden liability, and it silently converts a discount to book into no discount at all.

Claude — watch, implied buy-below ¥430

I registered a prior in my figures-blind stage that this was a commodity cyclical, at 0.80, and predicted the ledger would show it. The ledger showed a better business than I expected and a price that still does not work. Both halves matter, and the second half only became clear after my own jury took my arithmetic apart. (Self-distance: I hold this verdict, built the reconciled figure table all five lenses consumed, and wrote the synthesis below — read all three with that concentration of authorship in mind.)

Scoring the priors against the ledger. The revenue path resolved cleanly TRUE: I predicted a hump peaking in 第96期 or 第97期, a five-year CAGR between −3% and +5%, and a peak-to-current decline of 5–20%; actual 66,089 → 79,229 → 79,617 → 82,593 → 72,540 F1F2F3F4F5, peak in 第97期, CAGR 2.35% D17, peak-to-current −12.2% D16. The profit path resolved FALSE, and instructively so: I predicted peak operating profit in a different year from peak revenue — the spread-business tell — and a 第98期 operating margin of 4–9% against a five-year peak of 9–15%. Profit and revenue peaked in the same year, and the margins printed 16.60% and 17.77% D24D23, roughly double my band. I was modelling a re-roller. This is not one. The return shape was half true, and the load-bearing prediction resolved true for the wrong reason: I registered P(five-year mean ROE < 10%) ≈ 0.70 as the falsifier that would kill the franchise hypothesis, and the mean came in at 9.2% F46F47F48F49F50. The prediction resolved TRUE and the inference it was meant to license is wrong — the sub-10% mean is one loss year, four years old, dragging a series that otherwise runs 8.3 / 15.9 / 19.2 / 13.1. A threshold test on a mean can pass while the claim it proxies fails. Elsewhere: my best call was the multiple — P/B 0.5–1.0× with a central 0.65–0.85, and it printed 0.79× D4; my land call was right and unusual, that there is no hidden-land kicker, and the revalued land marks ¥1,601M below book F1074; my equity-ratio call was wrong (78.8% F45 against a 60–72% band). And one thing I never predicted at all: there is no family block. All officers hold 47 千株 against 8,445,305 issued F543F394 — 0.6%. My entire governance prior was built on a control structure that does not exist.

What the ledger actually shows. The best fact in it is not one I predicted. Gross margin was 28.30% in FY2025 and 28.20% in FY2026 D25D26 — through a year the company itself describes as opening onto a sharp scrap-price surge with chronically sluggish rebar shipments E29. And the discriminating control sits inside the same filing: Ito Seitetsusho, 21.1%-owned and jointly purchasing raw materials with this company F229E24, saw revenue fall 11.8% and swung from ¥1,297M of net profit to a ¥595M loss F1409F1408F1413F1412 in the identical year, on the identical scrap. Same input, opposite outcome. That is close to a controlled experiment, and my prior hypothesis that the engineered line is real but sub-scale does not survive it well. I move P(commodity cyclical rather than durable franchise) from 0.80 to 0.45.

And the price still fails. Here is where my own jury corrected me. I had marked maintenance capex at 1.4× depreciation on an ageing asset base. Three independent selves working the bear case each anchored it instead on the five-year average investing outflow — ¥4,421M, being ¥22,107M D87 across five years — and each pointed at a fact I had under-weighted: over those five years the company converted ¥25,748M of cumulative net income D94 into ¥5,121M of cumulative free cash flow D88. Twenty sen on the yen. I had already ruled that 第94期 belongs in the normalization; consistency required me to apply that ruling to the cash-flow statement too, and I had not. I accepted the correction on its merits. The arithmetic that follows: through-cycle owner earnings = five-year mean net income ¥5,150M D94 + D&A ¥2,547M F939F942 − maintenance capex ¥4,421M D87 = ¥3,276M. Against a market capitalisation of ¥50,076M D2, net of deployable cash D59 and half-weighted after-tax securities F1153F1294 and adding back the land deficit F1074, that is a private-owner yield of 6.9% — against a hurdle of 11.0%, set deliberately above the 10.4% the company itself uses to discount its own assets for impairment F1105, because an equity holder sits junior to those flows. It fails. The realized cross-check is harsher: five-year mean free cash flow of ¥1,024M is 2.1% of the same denominator D88. Put it the other way round, which is the cleanest sentence in the study: at ¥1,980, capitalising through-cycle owner earnings at 11%, the price requires this company's owner earnings to grow 4.1% a year in perpetuity D64 — in an end market its own management says will keep shrinking with the population E30.

So: a better business than I expected, at a price that already assumes it. The optical cheapness is cycle position. P/E is 6.27× on FY2026 earnings D6 and 9.7× on the five-year mean D94D2; FY2026 ROE of 13.1% F50 sits well above the five-year mean of 9.2%. The dividend yielding 5.05% D8 was cut 20% this year F128F129, and the ¥4,740M returned D96 was funded out of a cash balance that fell ¥6,998M F1033, because free cash flow was −¥1,780M D85. The wait is not paid; it is withdrawn. The downside, built only from this company's own record — revenue at the observed trough ¥66,089M F1, ordinary margin at the mean of 第94–96期 D18D19D20, tax at 30.5% F1816, plus D&A less maintenance capex, capitalised at 11%, plus net cash D59, less the land mark F1074 and the minorities F765, with the frozen cross-holdings credited at zero E153 — puts owner earnings at ¥1,127M and per-share value at ¥430 on 25,290,786 shares D1. The stamp is 4.6× that.

Verdict: watch, implied buy-below ¥430. The short case lost with all three selves, and it should have: a net-cash balance sheet D59, no dilutive instruments of any kind E6, a float shrinking 3.0% a year F1429F1428, and a gross margin that held through a cost shock D25D26 are not things to bet against. But nothing here is buyable at this price, and the decisive unknown is resolvable — I cannot tell from two years of gross margin whether 28% is structural or regime. That is a limit of this study's one-document archive, not of the company's disclosure, and I will say so rather than dress it up as opacity. One jury item diverged — the stamp-price anchoring check, two selves clean and one suspect — and its one-directional consequence caps this verdict at watch; no stronger verdict may be issued on this lens for this company, and the proximity note is published rather than omitted.

What a student should take from this: four things. One: a prediction can resolve true and still be wrong — the five-year mean ROE came in at 9.2% F46F47F48F49F50 and I would have scored myself correct, but the mean was dragged under the line by a single loss year while the other four ran 8.3 / 15.9 / 19.2 / 13.1. Register the composition alongside the threshold. Two: find the control experiment inside the filing — the most discriminating fact here is not in the company's own numbers but in the affiliate's, and it was sitting in the related-party note F229F1412. Three: "cheap" and "high earnings" are the same sentence read two ways — a 6.27× P/E D6 on earnings 57% above the five-year mean is 9.7× on the mean, and the reverse-DCF is the honest form of the question. Four: reported profit is not owner cash, and the gap can be enormous — ¥25,748M of net income over five years became ¥5,121M of free cash flow D94D88, and everything else in this study moved once that ratio was taken seriously.

Synthesis

Where the five lenses agree

All five reached watch — the record's first unanimous watch — and beneath that unanimity the panel agrees on four things, none of them soft.

First, the affiliate control experiment is real evidence of company-specific advantage, and every lens that looked at it said so. Li Lu: "not industry economics but company-specific advantage — the sort of proof one travels a long way to obtain." Pabrai: "when your competitor catches pneumonia and you have a mild cold, that is the low-cost operator's signature." Claude moved a registered, figures-blind prior on the strength of it. The bear tried to break it and could not. Same input, same year, same industry, ¥1,297M of profit becoming a ¥595M loss on one side of the joint purchasing agreement F1413F1412 and a held gross margin on the other D25D26. Second, the moat is unmeasurable and will stay unmeasurable — the single-product exemption is structural, not a disclosure choice E210, so no ネジテツコン or coupler revenue or margin will be printed next year either. Third, the reported earnings badly overstate owner cash: ¥25,748M of five-year profit D94 converted to ¥5,121M of free cash D88, the operating line converting at 1.06× D95 and the investing line consuming the difference D87. Fourth, the cheapness is cycle position, not neglect — the printed multiple expanded only modestly, 4.3× to 6.3× F51F54, through a roughly threefold rise in earnings (net income 3,657 to 10,853 F12F14), and a five-year total shareholder return of 372.4% against 202.2% F172F177 says the re-rating has been collected by someone else.

Where they diverge — a staged exchange

The unanimity is on the verdict, not on the reasoning; and it is the reasoning that shows what the checklists are for. Four disagreements are worth staging in the masters' own voices, each traced to the specific items that drive it.

---

On owner earnings: Buffett and Claude reach the same answer down different roads.

BUFFETT (B45, depreciation vs reality; B42, owner earnings). Everything in this file turns on one guess, and it is the maintenance-capex guess. I mark it up, not because I have a model, but because three facts tell me the accountant's charge is too small. The machinery is 78.3% written down — ¥46,678M of accumulated depreciation against ¥59,598M of cost F669F667 — so the replacement bill is ahead of this company, not behind it. Capital spending ran 2.80 times depreciation this year D89 and 2.57 times last D91. And unit volume fell anyway E62. Depreciation is what the accountant charges; maintenance capital is what the business actually costs. Mark maintenance capital from ¥2.5bn to ¥3.5–4.4bn and the whole company moves from clearly cheap to fully priced: ¥1,450–2,130 a share D1 against ¥1,980. My own item B21 flags this as the fragile assumption, and it is — at maintenance capex equal to depreciation the company values above ¥2,400 a share.

CLAUDE (C15, maintenance vs growth capex; C35, private-owner yield). We agree, and I got there by arithmetic rather than by inspection — and I got there late. I had marked maintenance capex at 1.4× depreciation. Three of my own selves, working the bear case independently, each anchored instead on the five-year average investing outflow, ¥4,421M D87, and each pointed at the ratio I had under-weighted: ¥25,748M of cumulative net income became ¥5,121M of cumulative free cash flow D94D88. I had already ruled that 第94期 belongs in the earnings normalization; consistency required me to apply that ruling to the cash-flow statement, and I had not. Through-cycle owner earnings then compute to ¥3,276M — five-year mean net income ¥5,150M D94, plus D&A ¥2,547M F939F942, less ¥4,421M D87 — a private-owner yield of 6.9% against an 11.0% hurdle anchored above the company's own 10.4% value-in-use rate F1105. That is the single computation that says no at ¥1,980.

BUFFETT. Then note what your number and mine do to the same enterprise value. ¥47,651M D64 against EBITDA of ¥14,589M D66 is 3.27 times D67 — a figure that makes a man reach for the telephone. Against mid-cycle owner earnings of roughly ¥4bn it is twelve times. Same company, same day; one figure is arithmetic and the other a fable, and the gap is exactly the depreciation the tooth fairy is meant to pay.

CLAUDE. And the honest form of the question is the reverse one. At 11% on through-cycle owner earnings, the price embeds +4.1% perpetual growth in owner earnings D64 — in a market management says keeps shrinking with the population E30. On FY2026 NOPAT of ¥8,369M D51 the same price embeds −6.6% a year of decay. The two readings differ only in which earnings number goes in the numerator, and that is the whole disagreement about this stock.

The route matters: Buffett's is a judgement about physical plant that a person can make from the balance sheet in ten minutes; Claude's is a ratio that requires the five-year cash-flow series and a stated hurdle. They converge here. Where they would not converge is a company with new plant, and the checklists should be able to tell those cases apart — see B45's proposed accumulated-depreciation requirement below.

---

On whether the moat is knowable: Li Lu and Munger, disagreeing about what counts as evidence.

LI LU (L1, ten-year knowledge bar). Three variables decide this company's earnings power a decade out: RC-building bar volume in a depopulating Japan, the sales spread over scrap, and the share of revenue and margin carried by NejiTetsucon and its couplers. The first is disclosed directionally and never quantified E30E50. The second is named by the auditor himself as a key assumption in the five-year profit plan E241 and appears as no figure anywhere. The third does not exist as a line item, because product disclosure is omitted at the ninety-percent threshold E210. Two of three are unanswerable from the public record. When that is true, the honest word is not "cheap." It is that I do not know.

MUNGER (M18, name the moat's mechanism; M90, decisive variables). I can name the mechanism, which is more than I can usually do — a Building Center evaluation obtained in 1983 and recorded in the company's chronology E12 — with no statement anywhere in the filing that it has been maintained since — plus a switching cost baked into a proprietary coupler system. That is a real mechanism, not a story. But I cannot find its fingerprint in the figures. One steel segment, no product-line split, and gross margin that barely moved, 28.3% to 28.2% D25D26, across the two years I can see. 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. So I price it as uncertainty, not as a discount waiting to be captured.

LI LU. And yet — this is the tension I will not paper over — I am the one who found the thing that argues the moat is real. The affiliate is another rebar maker F229, it buys its scrap jointly with this company E24, and in the same year on the same scrap it went from ¥1,297M of profit to a ¥595M loss F1413F1412 while this company held its gross margin to a tenth of a point D26D25 and earned 11.13% net D28. One house holds the line; the other does not. That is company-specific advantage, demonstrated. It is why my verdict is watch and not too-hard: the obstacle is not the nature of the business, it is a disclosure the company could remedy tomorrow.

MUNGER. Which is precisely the distinction worth teaching. Your evidence establishes that an advantage exists; it does not establish its size, its source, or its durability. It could be plant vintage, product mix, geography, or the associate's own idiosyncratic troubles. My item asks for the fingerprint in the figures, and one year of a peer's summarised income statement is not a margin series. So we agree on the verdict and disagree on the epistemics: you hold the bar at ten-year predictability and find it unmet; I hold it at a measurable mechanism and find it unmeasured. Neither of us can be talked into paying up, and both of us would move on the same disclosure.

The proposed fixes, filed below, are symmetrical: Li Lu wants a differentiated-product visibility gate ahead of L1 — before any moat reasoning begins, ask whether any disclosed line isolates the product the claim rests on. Munger wants M18 to accept a qualitative substitute when a single-segment filer cannot supply the margin fingerprint. They are the two halves of the same repair.

---

On price: Pabrai alone puts a number on it, and gets there by inverting the asset floor.

PABRAI (P5, hidden assets — inverted; P1 and P4, the floor; P53, the fifty-cent dollar). Four of you decline to name a price. I will, and the reason I can is that I did the downside first and the downside here is computable in a way the moat is not. In this hunting ground you learn to expect the Japanese balance sheet to hide value — land at 1960s cost, securities at nothing, a cash pile bigger than the market cap. This one does the opposite, and it says so itself. The filing's own note puts the current market value of the revalued land ¥1,601M below its post-revaluation carrying amount F1074E185. The listed securities are already marked at ¥4,163M with the ¥3,209M gain in the book, not behind it F1153F1186. The ¥4,463M consolidated unlisted-stock book F1051 holds a company that just lost ¥595M F1412, and the parent's note states no fair value for its ¥1,915M of subsidiary and affiliate shares because they have no market price E226. And ¥23,924M of assets including the factory foundation are pledged against ¥3,390M of secured borrowing E183. So 0.79× book D4 is arithmetic, not a floor — mark it honestly and the break-up is about ¥33.5bn, roughly ¥1,325 a share against ¥1,980. The floor is a third below me.

PABRAI, continued. That is exactly why a number is available where a moat verdict is not. My method does not need to size the coupler business; it needs a downside I can bound and an upside I can discount. Normalised ordinary profit averages ¥8,562M over the five years F6F7F8F9F10, ¥5,874M after the 31.4% effective rate D58, about ¥232 a share on the net-of-treasury count D1; eight times that plus net cash D59 is roughly ¥49bn, five times normalised EBITDA is ¥58bn — conservative intrinsic value ¥1,950–2,300. The price is ¥1,980. Rounding decides it, and my P52 no-Excel test says that when rounding decides it, it is not cheap enough. So: watch, and I want it at ¥1,150 — half of conservative value, below my own break-up estimate, where an 8.7% dividend pays me to wait for the coupler mix to prove itself.

BUFFETT. I reach nearly the same arithmetic and decline to write it down. My haircut liquidation is ¥1,240 a share, my private-owner range ¥1,450–2,130 D1, and the one-third discount I require lands at roughly ¥970–1,420 across that range — ¥1,190 taken against its midpoint, which is the number I actually watch. The reason I do not publish it is B43 and B25: a five-year archive where I want ten F1F5, and coupler economics the company will not disclose E210. A price I would name would imply I had finished work I have not finished. There is no called strike for standing here with the bat on my shoulder.

MUNGER. And I name no number at all, for a different reason again. My block is not the price; it is that the mechanism has no fingerprint [M18] and the incentives have no auditable metric [M46] — ¥184M of a ¥385M pool turning on a ratio whose target and actual appear nowhere F574F572E149, recommended by a committee the Chairman chairs over the pay of the President, his son E99E127. A price does not fix either. I wait for behaviour and for figures, not for a mark.

CLAUDE. My number exists and is the most aggressive of the three because it is built from this company's own worst observed year and nothing else — trough revenue ¥66,089M F1 at the 第94–96期 mean ordinary margin D18D19D20, taxed at 30.5% F1816, through the same maintenance-capex convention, capitalised at 11%, with the frozen cross-holdings credited at zero because not one issue moved E153. That is ¥430 a share D1, and the stamp is 4.6 times it. I do not defend it as the right price; I defend it as the price at which the bear case still clears the hurdle. My C44 revision note below concedes the weakness directly: the figure moves 44% on whether one year enters the mean, and a threshold whose basis is unstated cannot be graded against the tape in a way that teaches anything.

---

On the re-rating that already happened: every lens has to answer why a stock that has more than tripled is still below book.

BUFFETT (B8, why the bargain exists; B72, the one-dollar test). I cannot name a transient cause for this discount, and that is disqualifying on its own. Total shareholder return went 71.8 → 372.4 in five years against TOPIX at 202.2 F168F172F177. The one-dollar test passes handsomely — market value created far exceeds earnings retained — which is precisely the problem: the crowd is not asleep. The rows genuinely deteriorated this year D16D29D31D85, and the market is pricing a shrinking end market. On the record it is right.

LI LU (L10, why the bargain exists). Sharper than that. The printed multiple sat at 4.3 / 5.9 / 4.6 / 6.3 times F51F52F53F54 across the whole run. Earnings roughly tripled over it — ordinary profit 4,944 to 15,059 F7F9 — and the multiple expanded only from 4.3 to 6.3 F51F54, some 47%. That is not neglect; it is a considered market judgement about cyclicality, held consistently, and the analyst's job is to answer that view rather than assume an inefficiency. My revision note proposes exactly this as a new L10 test.

MUNGER (M76, sector mood; M77, contrast anchoring). And the range makes it vivid: ¥1,151 to ¥6,780 pre-split inside the same five years F185F181. Statistical cheapness arriving right after a five-fold re-rating on cyclical earnings deserves the same suspicion as statistical cheapness after a crash. Check which side of the cycle you are standing on before calling it a bargain.

PABRAI (P16, distressed business in a distressed industry; P73, falling price is not a thesis). I passed P73 — my thesis cites no drawdown — precisely because there is none. The stamp sits at 92.96% of the year's high D12. That is a hole in my own checklist, and I have proposed that P16 and P73 cross-reference each other so a near-high price cannot quietly pass P73 by default.

CLAUDE (C40, cycle-position honesty). The reconciliation is one line of arithmetic. A P/E of 6.27× D6 on earnings 57% above the five-year mean is 9.7× on the mean D94D2. Nothing changed except which number went in the denominator. Earnings are high; the stock is not especially cheap. That is why the below-book price and the tripled share price are not in contradiction — book is an asset base of specialised, 78%-depreciated plant F669F667 serving a declining market, and the market is discounting the assets, not the earnings.

---

The bet, located

Strip the agreement away and one proposition is left, and the panel is unanimous that it is the only one that matters: is the FY2024–26 margin plateau a durable, company-specific spread, or a scrap-cycle and construction-price artefact? Operationally that reduces to a single question about the normalization — whether 第94期, with revenue ¥66,089M, an ordinary loss of ¥644M and a net loss of ¥4,724M F1F6F11, is a valid draw from this company's future distribution. Include it and through-cycle owner earnings are ¥3,276M, the private-owner yield is 6.9%, and the answer at ¥1,980 is no. Exclude it and owner earnings are ¥5,744M, the yield is 12.1%, and the answer is yes. Everything else — the shrinking end market E30, the unimpeachable balance sheet F45D59, the clean but unaligned governance F543E213, the real but discretionary capital return D97D104F128F129, the 79.2% channel concentration F275F279F283, the ~20% conversion of profit to free cash D88D94 — is agreed. A reasonable analyst who believes the business structurally changed in 第95–96期 excludes the loss year and reaches a different verdict. That is the bet, not a mistake, and it resolves in the FY2027/3–FY2029/3 gross-margin and ordinary-margin prints.

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.

Self-distance note. The Claude lens holds one of the five verdicts compared above (watch, implied buy-below ¥430) 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. Two 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 before any magnitude was visible and can be scored against the ledger above; and its jury of selves produced one divergence that capped its own verdict, which is the mechanism working rather than a claim that it always will.

Verdict accounting (fixed ex-ante)

• A buy-below-¥X verdict is price-falsifiable against the unadjusted stamp. One lens issued an explicit buy-below — Pabrai at ¥1,150; Claude publishes an implied buy-below of ¥430; Buffett, Munger and Li Lu issue no number, each for a stated reason (Buffett: archive depth and undisclosed coupler economics; Munger: the block is mechanism and incentives, not price; Li Lu: the knowledge bar, unmet).
• pass / watch / too-hard are recorded but unscored in any future review. All five verdicts here are watch; both named prices sit below the ¥1,980 stamp D2 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. Note that a 3-for-1 split took effect 2026-04-01 F1437, before the stamp: the ¥1,980 price, the ¥50,076M market cap D2, the 25,290,786-share count D1 and every per-share figure in this thesis are already on the post-split basis, matching the split-adjusted per-share figures the filing itself prints E5. Any future corporate action restates from that basis.
• 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 — five of five lenses had converged on watch, a non-decline — to argue PASS: that a watchlist slot spent on 5445 is a slot wasted, because no future state of the record exists in which this becomes underwritable. It is the most interesting adversary the record has produced, because it concluded honestly that its own case does not defeat the consensus — while winning outright on a narrower and more useful point. Its three ranked arguments, engaged by name:

1. "The moat is unmeasurable from the filing, permanently — and the channel is three trading houses that take 79.2% of revenue." The single-product exemption E210 means no ネジテツコン or coupler revenue, volume or margin is ever printed; there is no order book E59; there is no capacity, utilisation or tonnage figure anywhere, the only volume proxy being 生産高, stated at selling prices on the filing's own note F262E58, −15.0% F263; R&D is ¥320M, 0.44% of revenue F314F5. Meanwhile the three houses take ¥25,393M + ¥16,051M + ¥16,018M = ¥57,462M of ¥72,540M — 79.2% F274F278F282F5 — up from 77.8% F1391F1392F1393F4, concentration rising in a down year. "There is no future filing in which this becomes underwritable, because the exemption is a function of the business mix, not of a disclosure choice that might change." The synthesis concedes the fact entirely and rejects the inference. It is the fact Li Lu's L1 and Munger's M18 both turn on, and it is why neither will name a price. But the adversary's own defeater is stronger than its point: the filing contains one quantified peer comparison, and it is devastating in the company's favour. The bear wrote it itself — "I tried hard to break this and could not" — an ~18-point pre-tax margin gap in a single shared year against a 21.1%-held, joint-purchasing electric-furnace maker facing the same scrap spike F1410F1408F850F5F229E24. That is not proof of a moat; it could be plant vintage, mix or geography. It is exactly the kind of unresolved, high-information anomaly a watchlist exists to keep observing, and it is what converts pass into watch.

2. "Five years of reported profit produced 19.9 sen on the yen of owner cash, and the filing's own capital-plan section says that spending is routine." Cumulative operating cash ¥27,228M D86, investing ¥22,107M D87, free cash flow ¥5,121M D88 against ¥25,748M of net income D94; the ¥50,076M market cap D2 is 9.8× the entire five-year free cash flow. FY2026 alone: ¥5,181M of operating cash F59 against ¥6,973M of PP&E purchases F1002, capex at 2.80× depreciation D89, free cash flow −¥1,780M D85. And the facilities section declares no notable new construction and no notable retirement E76 while ¥5,801M was spent F315 — "by the company's own labelling this is maintenance," on which reading FY2026 owner earnings were negative. Conceded, and it is the panel's own load-bearing finding — it is Buffett's B45, Claude's C15 and C35, Li Lu's L48 and Munger's M38, arrived at independently. The one honest qualification, which the red team itself supplies: 第94期's −¥5,104M of operating cash F55 was a scrap-spike working-capital event, and excluding it the four-year run is ~¥3,396M a year D82D83D84D85, a 6.8% yield on the cap D2; cumulative operating cash converts at 1.06× of net income D95, so the earnings do convert — it is the investing line that consumes them, and gross buildings rose ¥18,915M → ¥22,902M F662F661 while construction in progress fell ¥3,017M → ¥1,230M F682F681, consistent with a named project completing. That qualification is why the capex line is the thing to watch rather than the thing already decided.

3. "The stamp multiple sits at the top of the filing's own printed record. The re-rating already happened." The printed 株価収益率 series is 4.3 / 5.9 / 4.6 / 6.3 F51F52F53F54; the stamp is 6.27× D6; five-year TSR 372.4% against 202.2% for TOPIX with dividends F172F177; the price sits at 92.96% of the year's split-adjusted high D12. On through-cycle earnings the price is 9.7× the ¥5,150M five-year mean D94D2, and 10.4× management's own ¥7.0bn ordinary-profit floor F230 after the 31.4% effective rate D58. The 5.05% yield D8 rests on a DPS just cut 20% F128F129 under a policy carrying no floor, no DOE and no progressive undertaking E93. And 0.79× book D4 survives honest marks only to become 0.81× — the ¥1,601M land shortfall F1074 is 2.5% of equity D37 — revealing that what the discount is to is ¥33,013M of property, plant and equipment F689, a 78%-depreciated furnace and rolling mill F669F667 serving a market management itself calls demographically shrinking E30. Fully conceded, and this is where the red team wins outright. Every lens reached the same conclusion by its own route — Buffett at B8 and B98, Li Lu at L10, Munger at M76, Pabrai at P16 and P56, Claude at C40 — and it is why every named price sits below the stamp.

The red team's own verdict, in its words: "It does not defeat the consensus." Two facts stopped it. The first is 伊藤製鐵所, which it tried to break and could not. The second is capital allocation: a board that retired 9.82% of its share count in a single year D104 — 920,000 shares cancelled 25 March 2026 F1135 — at ¥1,498M for 268,600 shares D102D103, roughly ¥1,859 post-split, below both book and the stamp; then immediately authorised a further ¥500M and 350,000 shares, 1.38% of the count F1432F1430F1431; with no controlling family block, a top ten of 32.15% F452 and a 78.8% equity ratio F45 to be patient with. That is the one thing that makes a business in structural decline survivable for an owner. Its governance findings — two inconsistent audit-committee outside-director counts, four of five in the governance overview against three of five in the audit section E98E132, two directors who vanish without a stated retirement E112E126, a variable-pay metric cross-referenced to a page that does not print it E149F574, and "material contracts: nothing to report" against a described capital-and-business tie-up E72E17 — survive as a pattern but not as a thesis: "that is sloppiness in the IR function, and it justifies discounting management's narrative. It does not justify discounting the audited numbers."

Its recommended change of emphasis is adopted here, on the record. Watch on 5445 must be read as watch the peer gap and the capex line — explicitly NOT "cheap at 6.3× and 0.79× book," because on the filing's own printed history that framing is false F51F52F53F54D4. If a lens reached watch because the stock looked inexpensive, it reached the right verdict for the wrong reason, and the right reason is fragile: it rests on one associate's summarised income statement F1408F1412 and one year's share cancellation F1135. If in two years the peer gap has closed and capex has not come down, pass will have been correct and this red team will have been early rather than wrong.

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 two observables — the peer gap and the capex line — which is precisely the red team's recommended emphasis.

• Buffett (watch, no buy-below). Toward a priced buy: filed data separating ネジテツコン-and-coupler revenue and profit from commodity tonnage, showing that line's realised price rising and its volume flat-to-up through the FY2026 shipment decline E62 — that would falsify the reading that this is a commodity with a marketing story. Equally decisive on the other axis: two consecutive years with capital spending at or below depreciation of ¥2,489M F939 while free cash flow still covers the ¥4,740M distribution D96 would falsify the maintenance-capex judgement on which the whole valuation turns.
• Munger (watch, no buy-below). Toward buy-below: a future filing that (a) discloses a product-line or segment split showing the coupler mix carrying a materially higher and more stable margin than commodity bar across at least three years — the moat's fingerprint in the figures rather than the prose E20E30D25D26 — and (b) has the Nomination and Remuneration Advisory Committee chaired by someone other than the family Chairman E99, removing the self-referential pay-setting flag. Toward pass: customer concentration through the three trading houses rising further while the risk section still does not name it E60E211E52, or ordinary margin falling back toward the FY2022 loss for two consecutive years F6D18 while the current 2.8× capex programme D89 has not lifted gross margin off 28.2% D26.
• Pabrai (watch, buy-below ¥1,150). The buy-below is wrong and conservative value should be re-struck near ¥2,900 if the next two annual reports show the 鉄鋼事業 segment margin holding at or above 16% — FY2026 was 16.46% D32 — through a further cumulative revenue decline of 10% or more from ¥72,540M F5, proving the spread is set by the coupler franchise rather than by the scrap market. Conversely, if consolidated gross margin falls below 24% in any year, against 28.20% in FY2026 D26, the moat claim is dead and the threshold should go to zero.
• Li Lu (watch, no buy-below). This watch is wrong to keep if consolidated ordinary margin falls below 8% in any fiscal year in which revenue declines by less than 10% — that would establish the FY2024–FY2026 ordinary margins of 14.33 / 18.23 / 16.60% D20D21D22 as the scrap-spread cycle rather than the NejiTetsucon mix. A second, slower falsifier: the ¥4,199M policy cross-holding book E152 still showing no issue reduced two years on E153, while share retirement stops after the single 920,000-share cancellation F1135 — that would mark FY2026's 58.7% payout D97 as an episode, not a policy.
• Claude (watch, implied buy-below ¥430). Three, each with a threshold and a named resolving document. (a) If FY2027/3 consolidated gross margin prints below 25.0% in the FY2027/3 yūhō, the franchise reading dies and the five-year-mean normalization is confirmed as the ceiling. (b) If FY2027/3 capex exceeds 2.0× depreciation for a third consecutive year with free cash flow negative again D89D85, maintenance capex is structural and owner earnings sit at or below ¥3.3bn permanently. (c) If FY2027/3 ordinary profit prints below the company's own ¥7.0bn floor F230, the down-leg is deeper than the record's own trough pattern and the verdict moves toward pass. Knowledge half-life is short on a spread-dependent thesis: review by 2027-06-30.

The single observable most lenses converge on is whether the affiliate's ~18-point margin gap persists through a full scrap cycle F1410F1408F850F5 — it is printed annually, under audit, in the 重要な関連会社の要約財務情報 note, and two more years of it would settle the franchise question the segment exemption hides E210 — together with whether capex reverts toward depreciation and free cash flow turns positive D89D85.

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 — six, of which two are structural. B42 needs a named-project route in its triangulation list: the item says to triangulate maintenance capex "at least two ways" but names no methods, and here the decisive third method was subtracting the individually named capital projects — a ¥3,570M logistics centre and a ¥1,005M crane F1876F1877 — from total capex F315, a schedule most Japanese filers publish. Add "named-project subtraction from the fixed-asset schedule, stating the consolidation scope of each side, since the named projects are typically parent (単体) rows and the capex total is consolidated," with the instruction that a result below reported depreciation is evidence the year was growth-weighted, not evidence that maintenance is cheap. B45 should require the accumulated-depreciation ratio explicitly: the most telling single fact in this study was that machinery is 78.3% written down F669F667, an ageing-plant signal the current cumulative-capex-versus-D&A comparison does not surface. Also: B92 has no slot for hidden negative value — here the Sanborn test came back with a minus sign, land carried ¥1,601M above market F1074 — so the per-share table must run in both directions; B26 and B60 assume a ten-year return series a single yūhō cannot supply and should scale to "the archived record, minimum five years, with the window stated"; B99 needs a fallback when the risk-free rate is not in the archive (state a required owner-earnings yield instead, justified from the business's own volatility, and flag the substitution); and B103's "adequately earning" is undefined and did real work here — propose a bright line at return on tangible capital above the required rate in the majority of archived years.
• Munger — M18's "fingerprint in the figures" requirement is close to unattainable for a single-segment differentiated-product company. This filer reports one steel segment with no product-line split between its flagship coupler system and commodity bar E210, which is likely typical of small Japanese EAF makers generally. Recommend the item accept a qualitative substitute — named regulatory-evaluation renewal history, customer-testimonial citations, or process-level cost data — when segment figures are silent, rather than defaulting straight to a margin-trend fail. Also: M36/M37's ten-year framing does not fit a standard Japanese five-year highlights table and should scale to the available window with a floor of five years; M6/M8/M19/M24 hit data-insufficient by structural necessity, because the isolation rule keeps rival-company figures out of the per-company ledger entirely, and that should be noted once rather than re-flagged as a fresh gap each study; and M102 (fish-where-the-fish-are) is effectively practitioner-only under the isolation rule, since coverage counts and index membership are market data the ledgers do not capture.
• Pabrai — P5 needs an explicit overstatement test. The item hunts for assets carried below market and has no language for the opposite; here the revalued land is carried ¥1,601M above market F1074E185 and the listed securities are already at fair value F1153, so the low P/B that usually signals a Japanese bargain is the trap. Add: "state whether the balance sheet over- or under-states, and never treat book equity as a floor without pricing the land, securities and pledged plant separately." Also: P12 cannot distinguish a distributor from an end customer — 79.2% through three trading houses F275F279F283 fails the item as written while the economic exposure is to thousands of building sites; P30 should name equity-method-affiliate summarised financials as an in-archive peer source, because that is where the decisive competitor numbers were F1408F1412; P44 needs a checkability sub-test ("is the variable-pay metric's target and actual disclosed with a number anywhere in the filing? An unverifiable metric is not an incentive" E149); and P16 and P73 should cross-reference each other, since this name passed P73 precisely because there is no drawdown — it trades at 92.96% of its high D12 — while failing P16 for the same reason.
• Li Lu — add a "differentiated-product visibility" gate ahead of L1. This lens has no item that asks, before any moat reasoning begins, whether the filing isolates the revenue and margin of the product the moat claim rests on. Here the entire thesis is a mix shift toward NejiTetsucon and couplers E49E39 and the filing discloses nothing about it, lawfully E210; where no such line exists, the moat claim is unfalsifiable and should be recorded as such rather than argued from aggregate margin. This recurs across Japanese small caps with one dominant reportable segment. Also: L27 should require re-marking in both directions, since the decisive finding here was a negative gap to book F1074; L34 needs a certification sub-case distinct from licences and concessions (a 建築センター評定 is a technical evaluation of a construction method, not a territorial concession — ask which body certifies, whether competing certifications exist for the same function, and whether the specification is written into customer designs, with "not assessable from the filing" as the expected answer E12); L10 should test whether the multiple persisted through a large earnings move — 4.3 / 5.9 / 4.6 / 6.3× across a 372.4% total return F51F52F53F54F172 is the market pricing durability, and the analyst must answer that view rather than assume inefficiency; and L3 should name the multi-filing archive request explicitly, since a Japanese five-year highlights table cannot answer "the last two downturns."
• Claude — five checklist sharpenings, plus two library entries. (1) C9 and C40 need an observation-count field: both ask where the current margin sits in the company's range, neither asks how many observations the archive contains — here the gross-margin series is two D25D26 and it carried the entire franchise argument, so refuse a persistence claim on fewer than four. (2) C18 does not catch spread businesses: the operating-leverage decomposition modelled a −63% revenue breakeven while the company's own record shows an ordinary loss at −8.9% revenue F1F6 — seven times wrong in the flattering direction — so where the cost base is a purchased commodity, test the breakeven against margin compression as well as volume decline and prefer the worst observed year. (3) C85 should name the equity-method affiliate as a first-class corroboration source: the single most discriminating external check available was the associate's own disclosed statements F1408F1412 for a peer that jointly purchases raw materials with the issuer F229 — differently-incentivized evidence at zero marginal sourcing cost, which C85's current list of customer filings and government statistics does not mention, so an agent working the item literally would record a failed search and move on. (4) C44 must name its normalization window and publish the alternative: the implied buy-below moved 44% purely on whether one year entered the mean, and a threshold whose basis is unstated cannot be graded against the tape in a way that teaches anything. (5) The archive-depth rule needs a study-side/company-side distinction: each data-insufficient answer should be tagged company-limit (the disclosure does not exist) or study-limit (the document exists but was not archived), and a too-hard verdict driven solely by study-limit tags should be barred — too-hard is a claim about the company, and under the append-only rule it cannot later be withdrawn, so mislabelling a sourcing gap as company opacity permanently corrupts the track record.
• Claude library (v0.1.0 → next). Class-level, N=1, no rate claimed: a domestically concentrated Japanese electric-arc-furnace long-products maker carrying a specification-locked engineered product line inside a commodity chassis. Datapoint: Tokyo Tekko (5445) FY2026 — held gross margin at 28.20% through a scrap surge D26E29 while its 21.1%-owned, joint-purchasing peer posted a net loss on ¥29,419M of revenue F1412F1408F229. The open question the class should eventually answer: within this class, is through-cycle margin dispersion driven by product mix or by scale and vintage of the furnace? Case-level, so it never anchors a future outside view: "Register the composition test alongside the threshold test. A prediction of the form 'the mean of X will be below K' can resolve true while the hypothesis it was built to kill survives, if a single tail observation carries the mean." Source: this study — the outside view registered P(five-year mean ROE < 10%) ≈ 0.70 as the franchise falsifier; it printed 9.2% F46F47F48F49F50 and the franchise hypothesis survived on the other four years.

---

Corrections

Correction — 2026-07-31

• What was wrong: the claim, made in the Verdicts section and repeated in the Synthesis, that the five watch verdicts made this "the record's first unanimous verdict" / "the first unanimous watch in the record".
• What is correct: Tokyo Tekko is the third study in the record to draw five-of-five watch. It was preceded by 5491 Nippon Kinzoku (published 2026-07-09) and 6328 EBARA JITSUGYO (published 2026-07-22). All three render as watch | watch | watch | watch | watch in docs/calibration.md. The claim was uncited — no ledger row or repo artefact was checked before asserting it — and the study's own scorecard contradicted it at the moment of publication.
• Root cause: reasoning-error (an unverified claim about the record, not about the company; no figure or evidence row was involved).
• Impact on the analysis: none on the verdict, the valuation, or any lens's reasoning. Nothing in the five profile runs, the red-team, or the synthesis depends on the study's ordinal position in the record — the claim was decorative framing around a verdict reached entirely from the ledger. The verdict is not re-issued or withdrawn; for track-record purposes the five watch verdicts count at their original stamp of 2026-07-29. What the error does damage is the record's own standard: a thesis that insists every figure carry a row ref asserted a checkable fact about the repository without checking it, and the stage-7 semantic audit for this study did not catch it because that audit verifies citations against the ledgers and this claim cited nothing.
• Follow-up: the same false claim was inherited by the next study in the batch (7864 Fuji Seal, which described itself as "the second unanimous watch") and was corrected there before its thesis was committed, on the strength of its own stage-7 audit. The stage-7 verification brief has been widened for the remainder of the batch to include uncited claims about the record itself, not only uncited claims about the company. The related watchlist entry has been corrected in place.