Kitagawa Corporation (6317): A Third of Book, and Every Bargain Number Halves

Verdicts

Lens	Verdict	Buy below	Most load-bearing items
Buffett	watch	¥1,150	B26/B60 return on tangible capital ~3.8% (mgmt's own target ROIC 6.0% / ROE 6.5% E19) — below cost of capital; B89 cheap on book (0.34× D1) but ~5× NCAV [B89], the discount buried in a loss-making Mexican foundry F143; B46 the "+151% record" is a one-off ¥2,369M asset-sale gain F188D8
Munger	watch	¥1,350	M18/M37 no moat mechanism the figures confirm — the chuck "franchise" earns a 2.2% margin, −48% E38; a mediocre business at a bargain; M46/M83 chairman sets director pay E90, cross-holdings unreduced E95, no buyback at 0.34× book E69 — the family is comfortable, nothing forces change
Pabrai	watch	¥950	P53 fifty-cent-dollar fails — normalized ~10.5× D5 at ~100–130% of conservative IV; P1 net debt D7 not net cash → marked-asset floor ~¥1,366, soft-to-medium; the value is trapped behind a controlling family
Li Lu	too-hard	null	L1 can't forecast a four-segment cyclical conglomerate 10yr — profit 74% one segment (Sun Tech, POC revenue E104D11), two loss years in five F11F12; L35/L36 normalized ROE ~3.4% D8, flat-to-melting value, sub-hurdle long-hold
Claude	watch	null (implied ¥1,000)	C39 the reachable floor (liquidation NAV ~¥1,188/sh) sits below the stamp once the undistributable pension surplus + never-sold cross-holdings + all debt are marked; C34/C37 normalized PER ~10.5× prices the sustainable earnings ~fairly — the "cheap" is the one-off

No lens buys at ¥1,687. Four reach watch with buy-belows 20–45% below the stamp (¥950–¥1,350; Claude's implied floor ¥1,000); one reaches too-hard — and a consensus red-team argues the harder thing, that a controlled discount with no forcing agent is a pass, not a watch. The cheapest-with-a-catalyst name on the record, and still no buy.

The business

Kitagawa Corporation (株式会社北川鉄工所) is a 108-year-old Hiroshima metal-basher — founded 1918 E3 — that does four quite different things under one roof. It makes lathe chucks (the jaws that clamp a workpiece on a lathe), and there it is genuinely good: a globally recognized brand with more than half the domestic market E4 — the "quality" product. It builds concrete-mixing plants, climbing cranes, and self-propelled parking garages under the "Sun Tech" name (the 産業機械 segment) E5. It pours iron castings for the likes of Kubota (its largest customer, 15.6% of sales / ¥9,093M F339E44) E6. And it has a small semiconductor-polishing arm E7. Four reportable segments, four different economics; you can explain all four to a shopkeeper, so it is inside the circle — but the plain arithmetic does most of the talking.

The register is the second defining fact. The company is founder-family-controlled: Chairman Yuji Kitagawa and Vice-Chairman Hiroshi Kitagawa are brothers, both representative directors E82; the President was recruited from the main correspondent Hiroshima Bank E83. Individual director pay is delegated to the Chairman himself E90. The top-10 holders are a trust bank, the company's own suppliers' (みのり会, 8.09%) and employees' associations, Hiroshima Bank (4.82%), an Interactive Brokers omnibus nominee, and family — no activist, no named value fund E72E73. The board is reforming at the margin — a Nomination & Remuneration Committee created April 2026, a proposal to cut the director ceiling from 20 to 10, five of eight directors outside E79E80E81 — but pay stays with the Chairman and the structure is not contestable without the family's consent. The auditor is KPMG AZSA, 58 years in, unqualified, with a single Key Audit Matter on Sun Tech's percentage-of-completion revenue E102E104.

The numbers

FY2026/3 revenue was ¥58,415M (+2.0%) F5, and it looks, at first glance, like a screaming bargain: at the ¥1,687 stamp the whole company is priced at ¥15.61bn D2, 0.34× book D1 (BPS ¥4,967.53 F35), a reported PER of 5.0× D5, and a 6.05% dividend yield D12. Three siren numbers. Each one halves on inspection.

The earnings are dressed up. Reported net income was a record ¥3,128M, up 150.9% F15D6 — but ¥2,369M of that is a one-off gain on selling land and buildings in Japan and Thailand (固定資産売却益) F188, which you can only book once. Strip it at the statutory rate and normalized net income is ~¥1,481M D8 — less than half — so the honest multiple is ~10.5×, not 5× D5. The clean, repeatable figure is operating profit ¥2,688M, up a genuine +43.6% F82D4 (a 4.60% margin D3) — a real recovery, led by Sun Tech and a castings turnaround, but a thin one. The ROE tells the same story: reported 7.1% F50 normalizes to ~3.4% — below management's own modest FY2027 target of ROE 6.5% / ROIC 6.0% E19F157, and below any cost of capital. And the dividend halves too: the ¥102 DPS F115 steps off the record year, but the tiered policy (a ¥50 floor + 30/35/40% consolidated payout by income band E64E65) lands normalized profit in the 30% tier, so the sustainable dividend gravitates back toward the ¥50 floor F356 — a ~3% run-rate yield, not 6%.

Underneath, the profit is dangerously concentrated. Sun Tech (産業機械) earns 74.4% of reportable-segment profit on 37.7% of revenue D11D10 — ¥2,806M (up from ¥1,668M F298F299) at a ~15% return on segment assets F312, the only segment clearing a hurdle. The rest is nearly profit-less: castings ¥606M on ¥24,319M (a 2.5% margin, after a ¥128M loss the prior year F300F301); the "quality" chuck franchise just ¥221M on ¥9,870M — a 2.2% margin, down 48% on price competition F297E38; semiconductor ¥138M on ¥1,780M F303. And the segment carrying the profit is the one the auditor flagged: Sun Tech's percentage-of-completion construction revenue (¥4,069M, ~7% of sales) is the sole KAM, with the auditor stating it "can be arbitrarily manipulated by reallocating actual incurred construction costs" E104.

The balance sheet is not the fortress the P/B implies. It is net debt ~¥4.4bn D7 (interest-bearing debt ¥16,391M vs cash ¥11,965M), not net cash — a 54.1% equity ratio F45 and no bonds F374 mean solvency is not in question (net-debt/equity 9.6% D13), but there is no cash cushion to hand back. Worse, the book is dominated by assets a minority cannot reach: a ~¥6.7bn overfunded pension surplus (plan assets ¥16,526M vs a ¥9,864M obligation F419F418) that is undistributable; ¥4,606M of cross-holdings carried at a ¥3,589M unrealized gain F406F412 that management increased three names and reduced none this year despite a stated reduction policy E95; and a loss-making Mexican foundry carrying ¥7,501M of PP&E and a ¥603M loss F344F143. Cash generation actually collapsed — operating cash flow fell to ¥2,049M from ¥6,152M F60F277 (the gain is netted out of OCF), and the year was plugged with ¥2,713M of new short-term borrowing F282. Over the comparison window the stock returned 114.5% against the dividend-inclusive TOPIX's 202.2% F135F136 — an ~88-point underperformance, through a record year. Net-net working capital is only ~¥342/share [B89]: the discount lives entirely in fixed and cross-held assets, not in pocketable cash.

The five lenses

Buffett — watch, buy below ¥1,150

Let me start where I always do: with the business. Kitagawa does four quite different things E3 — it makes lathe chucks and there it is genuinely good, over half the domestic market E4; it builds concrete plants and cranes under "Sun Tech" E5; it pours iron castings E6; and it has a small semiconductor arm E7. Inside the circle. What pulled the idea onto the desk is the price: about a third of book — 0.34× D1 — five times reported earnings D5, a 6% dividend D12. Those three numbers are a siren song, and my job is to decide whether it comes from a lighthouse or from rocks.

Start with the earnings, because the headline is dressed up. Reported net income was a record ¥3,128M, up 151% F15 — but ¥2,369M of that is a one-time gain from selling land and buildings F188; you can only sell the Tokyo dormitory once. Strip it and normalized earnings are about ¥1,481M D8, so the honest multiple is roughly 10.5×, not 5× D5. The clean figure is operating profit ¥2,688M, up a genuine 44% F82D4, and that IS a real recovery. But it is ordinary, and this decides the case: after-tax operating profit on the tangible capital the business uses is a return of about 3.8%. Management's own three-year target is ROIC 6.0% / ROE 6.5% E19 — they are telling you, in their own document, that they aspire to earn barely their cost of capital by 2027. I look for 15% without leverage doing the work. Worse, the profit is not spread evenly: Sun Tech earns 74% of all segment profit D11, while the chuck business I admired earns a 2.2% margin E38. You are buying one decent business stapled to three mediocre ones, at a discount. That is a cigar butt, and cigar butts need a catalyst or a hard-asset floor.

So look at the assets, since the price screams asset play. It is cheap on book but it is not a Graham bargain: net current asset value is only about ¥342 a share — nearly 5× NCAV [B89] — and cash plus securities minus all liabilities is deeply negative. The discount lives in fixed assets: the plants, and above all a Mexican foundry that lost ¥603M this year and sits on ¥7.5bn of PP&E F143F344. A discount to a book that compounds at 3% D8 is not the margin of safety of a discount to cash you can pocket. And you must survive the wait: this is net debt of about ¥4.4bn, not net cash D7, and the dividend that draws you in is being paid partly out of one-off asset sales. Now the catch, and why this is a watch and not a pass: there is a real self-help policy E64E65, and the board is proposing to shrink itself 20→10 E80 with a new committee E79 — the quiet stirrings of TSE-era reform, not yet a catalyst you can bank on. So: inside the circle, a real operating recovery, a large discount to book — not a pass. But below its cost of capital, three of four segments mediocre, record earnings one-off-flattered, net debt, a yield part-funded by asset sales. A watch, not a buy at ¥1,687 — I'd want it below ¥1,150, roughly a normalized 7× and under a quarter of book, before the price does the work.

What a student should take from this: cheap-on-book and cheap-on-headline-earnings are two different claims, and both can be traps. Always ask what the business earns on the capital it uses — here, below its own cost of capital E19 — and what part of the current-asset pile you could actually pocket — here almost none, because the discount is buried in a loss-making foreign foundry F143. A 6% dividend paid out of selling the furniture is not the same bird in the hand as one paid out of owner earnings.

Munger — watch, buy below ¥1,350

Invert first. How does this business die? Four paths. Sun Tech falters — it earns 74% of segment profit on 38% of revenue D11D10, on percentage-of-completion contracts the auditor put as the only KAM, saying revenue "can be arbitrarily manipulated by reallocating actual incurred construction costs" E104; if it deteriorates the headline collapses. Castings stays structurally impaired — 2.5% margin F301, a loss-making Mexico plant F143, a Fukuyama pig-iron line written to zero E46. The chuck business — once the crown — earns 2.2% and fell 48% E38. And the family allocates capital by its own lights: the chairman decides individual pay E90, cross-holdings unreduced E95, no buyback despite 0.34× book E69.

Now the moat question — name the mechanism. Work-holding chucks: possibly, domestic share leadership and a 100-year brand E4, but the margin is 2.2% and falling; whatever mechanism existed is not defending margins. Castings: commodity iron, no mechanism. Sun Tech: a project shop subject to construction cycles. There is no durable moat the figure table confirms is widening, and the five-year ROE is △2.6 / △1.1 / 3.3 / 3.1 / 7.1 [F46-F50] — the last one-off-inflated (normalized ~3.4%, below even the feeble 6.5% target E19). A mediocre business at a bargain price loses to a great business at a fair price. Then the incentives — show me the incentive and I'll show you the outcome. The chairman sets each director's pay E90; he and his brother co-head the board E82; the board decided cross-holdings need not be reduced this year despite policy E95; the payout of the one-off proceeds is good for the family that controls the distribution but unlocks nothing. The incentives map: the family is comfortable, is not in pain, and has no economic pressure to change. And the cash-flow deterioration confirms the poor quality of this income year — operating cash flow fell from ¥6,152M to ¥2,049M even as operating profit rose F59F60, interest cover collapsed 16.0×→5.8× E53, short-term borrowings jumped F282.

What keeps it off the too-hard pile: the sum-of-parts is simple, the book reconstructable, no derivatives, KPMG AZSA 58 years in E102. Not too-hard from complexity — simply not a great business. So: statistically striking cheapness and a real operating recovery keep it from a pass, but a below-cost-of-capital return, no nameable moat, and a comfortable controlling family cap it. Watch, buy below ¥1,350 — where Sun Tech's durability and an actual capital-return act (a buyback, or cross-holdings cut below ¥2bn) would be needed to tip it to a buy.

What a student should take from this: a great segment stapled to three mediocre ones is not a great business — weight the whole, and name the moat mechanism or admit it isn't there. And when the controlling family sets its own pay and won't sell relationship cross-holdings at a third of book, the discount is not an opportunity waiting to close; it is a preference, revealed.

Pabrai — watch, buy below ¥950

Heads I win, tails I don't lose much — and the honest answer is maybe on the first half, and the second half is softer than the sticker screams. This looks like the fifty-cent dollar I hunt: 0.34× book D1, 5.0× reported earnings D5, a 6% dividend D12. But my checklist has a rule written in blood — an equity cushion is not an asset floor — and Kitagawa carries net debt of ~¥4.4bn D7, not net cash. So I build the floor from marked-down hard assets minus every liability. Against ¥38,950M of liabilities F260 I mark cash at 100%, receivables at 85%, inventory at 60% (castings WIP doesn't sell fast in a trough), securities at 80%, land at book (a tailwind — they just sold Tokyo/Thai land carried at stale cost F188, so book land is below market), and buildings-plus-machinery at 50% (a foundry is special-purpose iron). The overfunded pension reverts to beneficiaries — zero in a liquidation. That nets to about ¥1,366 a share — ~81% of the price, a −19% worst case; push to a severe casting trough and the floor sags to ~¥550. Soft-to-medium, not the fortress the P/B advertises.

Now the part that decides it. Strip the one-off and the cheapness evaporates: normalize the ¥2,369M gain F188 and I get ~¥1,481M D8, a normalized PER of ~10.5×. Put 8–10× on that and blend with the ~¥1,366 marked floor, and honest conservative IV clusters around ¥1,300–1,600. The stock is ¥1,687 — 100–130% of IV, not 50% (P53 fails outright). The 0.34× book is an accounting mirage assembled from an overfunded pension, stale-cost land, and cross-holdings F406 — none of which reach my pocket while the founder family runs the show. Am I paid to wait? Partly — but on sustainable ~¥1.5bn the durable dividend is ~¥50, a 3% yield F356, not 6%; the same one-off flatters the yield that flatters the PER. And who lets the discount close? Nobody — brothers control it E82, pay delegated to the Chairman E90, and against a policy to reduce cross-holdings they added three names E95. No activist, no buyback E69. The archetype of a controlled compounder cheap on paper where the discount can persist for a decade. The downside doesn't fail outright — it survives (54.1% equity F45) and pays a real if smaller dividend — but there is no asymmetry once the one-offs are stripped. Not a pass (it's not broken), emphatically not a buy (no fifty-cent dollar). A watch, buy below ¥950 — a price that restores the margin of safety against the reachable floor, or on proof the family will hand the trapped value back.

What a student should take from this: when a company carries net debt, you may not read the floor off the equity line — build it from marked assets minus every liability, and a "0.34× book" D1 can become a −19% worst case. And a bargain you cannot reach is not a bargain: an overfunded pension, stale-cost land, and never-sold cross-holdings behind a controlling family are book, not margin of safety.

Li Lu — too-hard

Begin where the discipline demands — not with the price, but with whether I can know this business. Kitagawa is a 108-year-old maker of four unrelated things E3. It is genuinely cheap — 0.34× book D1, ~5× earnings D5, a 6% yield D12, ¥15.6bn against ¥45,962M of net assets F271. A student feels the pull; my job is to teach you to resist it until the knowledge bar is cleared, because no price rescues a business whose next ten years you cannot predict. To predict Kitagawa's earnings power in 2036, three variables decide it: whether the Sun Tech profit engine holds (74% of segment profit D11, order-book and percentage-of-completion driven E104); whether castings — the largest top line F291 — can earn more than a rounding error (¥606M this year, a ¥128M loss last, carrying loss-making Mexico F143 and a Fukuyama plant impaired to zero E46); and whether the chuck business can defend price (segment profit just fell 48% E38). Can the filings answer these ten years out? Honestly, no — these are late-cycle construction and commodity-casting demands I cannot floor with the confidence this lens insists on.

The filed downturn record settles it against a compounding story: in the last five years the company posted net losses in two of them — FY2022 −¥951M and FY2023 −¥418M F11F12. This is a business that loses money in downturns. And when I strip the flattery, the honest picture is thin: the +150.9% jump is a one-off ¥2,369M land gain F188, not operating performance; normalized net income is ~¥1,481M D8, a normalized ROE of ~3.4% — below management's own 6.5% target E19 and its stated cost of capital. The operating recovery is real and I credit it (+43.6% D4), but a 3.4% normalized return on equity is the number a long holding period converges to (L36), and it clears no owner's hurdle. Cash confirms the caution — operating cash flow was only ¥2,049M F60 with the gain netted out, and this is net debt, not net cash D7. On the credit side this is not a governance trap: no listed parent E115, a clean single-KAM audit E104, related-party dealings small and arm's-length E114, a genuinely reforming board E80, a real improving dividend E64. But director pay is set by the Chairman himself E90, and despite a written cost-of-capital test the company increased three cross-holdings and reduced none E95 — capital parked in relationships. Is intrinsic value compounding or melting? Flat-to-melting: a cyclical conglomerate earning ~3.4% normalized, two loss years in five, a largest segment that barely breaks even, a family content to hold cross-holdings. A statically cheap, non-compounding business is a pass in this framework — and one whose ten-year earnings power I cannot predict is too-hard, whatever the discount. Both conditions hold. I will name the boundary rather than manufacture a buy-below to look decisive.

What a student should take from this: the knowledge bar is about predictability, not cheapness — a four-segment cyclical conglomerate that loses money in downturns and whose profit is 74% one project-based segment cannot be forecast a decade out, and no 0.34× book changes that. Admire the discount, and pass, rather than pretend to a confidence you do not have.

Claude — watch (implied buy-below ¥1,000)

I registered this figures-blind as a cheap-and-controlled value trap — a solvent, statistically cheap, founder-controlled cyclical maker whose statistical cheapness overstates the reachable margin of safety, whose record is one-off-flattered, and whose register offers no unlock. The ledger confirmed all three drags and overturned one prior favorably: I feared a tightening leverage picture, but the group is comfortably solvent (54.1% equity ratio F45, net debt only ~¥4.4bn D7) — the cash-flow deterioration is a one-year artifact of the same asset sale netting out of OCF, not a solvency spiral. That overturn removes the left tail and is why this is a watch, not a pass. But the two verdict-bearing priors held against the name.

Finding 1 — the record is an accounting artifact; the "cheap" is mostly the one-off. Record net income ¥3,128M (+150.9%) F15 is flattered by a one-off ¥2,369M asset-sale gain F188; normalized, ¥1,481M D8 — half. Reported PER ~5.0× D5 becomes **10.5× normalized**, at which a reverse-DCF implies a 9.5% no-growth return — the market prices the sustainable earnings roughly fairly. Reported ROE 7.1% F50 normalizes to ~3.4%, below management's own 6.5% target F157. This is the RINGER HUT lesson in a cheap register: value the operating line, not the flattered headline. Finding 2 — the discount is not reachable enough to carry a buy. Reported book is ¥45,957M F370, but three large blocks deliver nothing to a minority: a ~¥6.5bn undistributable pension surplus (plan assets ¥16,526M vs a ¥9,864M obligation F419F418); ¥4,606M of cross-holdings never sold F406E95; and a loss-making Mexico plant F143. Haircutting honestly, a liquidation-realizable NAV is ~¥1,188/share — below the ¥1,687 stamp; the going-concern figure (¥4,087) is above price but delivers nothing to a minority absent a sale. The deep 0.34× P/B D1 rewards the screen, not the buyer. Finding 3 — no forcing agent, and the dividend itself is one-off-flattered. Founding brothers control the board E82, pay delegated to the Chairman E90, cross-holdings unreduced E95, no buyback (1,299 odd-lot shares) E69, no activist E72. And the "6% yield" is really ~3% durable — the ¥102 DPS is ~2× the ~¥50 sustainable on normalized profit F354.

Why watch, not pass or too-hard: the verdict rests on resolved arithmetic (normalized earnings, the haircut NAV, the register), and the load-bearing unknowns — does the payout hold on normalized profit; do cross-holdings ever sell — are time-resolvable, exactly what a watch names. A solvent business with one genuine ~15%-ROA franchise, a real operating recovery, and a live capital-return policy — but a normalized return below hurdle, a reachable floor at/below the price, and no agent to force the discount closed. A name I'd want to own ~40% cheaper or on proof the normalized engine clears its target. Watch; implied buy-below ¥1,000.

What a student should take from this: a 0.34× P/B D1 and a 5× PER D5 are not a margin of safety until you mark the dollar to what a minority can actually reach and strip the one-off from the earnings. Here the same asset sale flatters both the record profit (normalized PER doubles to ~10.5×) and the dividend (sustainable ~¥50 vs the ¥102 headline), while the book is heavy with an undistributable pension surplus, never-sold cross-holdings, and a loss-making foreign plant — so the reachable floor sits below today's price. Cheapness that persists behind a comfortable controlling family is the market's standing statement that it doubts the discount is reachable.

Synthesis

Where the lenses agree

For once the panel is looking at a genuinely cheap name — 0.34× book, 5× reported earnings, a 6% headline yield D1D5D12 — and every lens still declines it at ¥1,687. The agreement is remarkably tight on why, and it runs through three facts none of them dispute. First, the record is a mirage. All five strip the one-off ¥2,369M asset-sale gain F188 and land on the same place: normalized net income ~¥1,481M D8, a normalized PER of ~10.5× not 5× D5, a normalized ROE of ~3.4% F50 against management's own modest 6.5% target E19. Buffett calls the headline "dressed up"; Munger, a "poor quality" income year; Pabrai, an "accounting mirage"; Li Lu, "flattery"; Claude, "the RINGER HUT lesson in a cheap register." This is the batch's echo: after RINGER HUT's tax benefit, a second consecutive study whose "record" was an accounting event, not operating strength — and the discipline of bridging to the operating line decides both. Second, the discount is unreachable. It is net debt, not net cash D7; the book is dominated by an undistributable pension surplus, ¥4,606M of never-sold cross-holdings F406E95, and a loss-making Mexican foundry F143. Pabrai's marked-asset floor (¥1,366) and Claude's liquidation NAV (~¥1,188) both sit below the stamp — the 0.34× P/B is not a reachable margin of safety. Third, nobody forces the gap closed. The chairman sets his own board's pay E90, the family holds cross-holdings against its own reduction policy E95, there is no buyback and no activist E69E72 — and the stock has trailed the index by ~88 points through a record year F135F136.

Where the lenses diverge

The split is watch (Buffett, Munger, Pabrai, Claude) vs too-hard (Li Lu) — and, uniquely, the four watchers barely differ on the facts; they differ only on price. The buy-belows fan from Pabrai's ¥950 (half the reachable floor, a true fifty-cent dollar) through Claude's implied ¥1,000 and Buffett's ¥1,150 (a normalized 7×, under a quarter of book) to Munger's ¥1,350 — each the price at which the reachable discount, not the optical one, restores a margin of safety. Li Lu alone declines to name a price: "the knowledge bar is about predictability, not cheapness — a four-segment cyclical conglomerate that loses money in downturns F11F12 and whose profit is 74% one project-based segment D11 cannot be forecast a decade out, and no 0.34× book changes that." Munger presses the same doubt from the moat side rather than the knowability side — "name the mechanism; a 2.2%-margin chuck business that fell 48% E38 hasn't got one" — but stops at watch because the sum-of-parts is simple enough to price. The genuine tension is not among the five; it is between the watch consensus and the question the red-team was dispatched to press: is a controlled discount with no forcing agent even worth watching, or is it a pass?

The red team, engaged

Because four of five agreed on watch, a fresh adversary (ledger only) argued the harder case — pass, not watch — and the synthesis must meet its strongest points by name. Its thesis: "a watch is a bet that a discount is real and reachable; I argue it is real but unreachable, because the assets behind it are undistributable (pension), held-by-policy (cross-holdings), or impairing (castings), and the only people who could force distribution are a control family who demonstrably choose not to."

• "The book discount is unreachable, not merely large." The synthesis concedes this entirely — it is Claude's C39 and Pabrai's P1, independently reached: net debt D7, a ¥6.7bn undistributable pension surplus F419F418, ¥4,606M of unsold cross-holdings E95. The watchers do not dispute the fact; they price it, by marking the reachable floor (¥1,188–1,366) and buying only below it.
• "All three bull metrics halve on normalization." Conceded and shared — every lens normalized the ¥2,369M gain F188 to ~10.5× / ~3.4% / ~3% D8. This is not a point of difference; it is the consensus's own finding.
• "No forcing agent exists — the discount is permanent." This is the real disagreement, and it is one of weight, not fact. The red-team is right that the chairman sets pay E90, cross-holdings rose not fell E95, and there is no buyback or activist E69E72 — so no external agent will close the gap. The watchers answer not that a catalyst is coming, but that at a low enough price the discount need not close to pay: Pabrai's ¥950 and Buffett's ¥1,150 are prices where a permanent discount still yields acceptably against a reachable floor, and the ~3% floor dividend F356 pays you (modestly) to wait. The red-team's own concession line — "¥1,150–1,300… where even a permanent discount pays acceptably" — overlaps the watchers' buy-belows almost exactly. So the pass-vs-watch gap narrows to this: the red-team drops the name from the list because nothing forces the value out; the watchers keep it at a price, because a reachable discount plus a floor yield is ownable even without a catalyst. Both agree ¥1,687 is a decline. The honest reconciliation: this is a watch only for an investor willing to be paid ~3% to hold a permanent discount bought well below the reachable floor — and the red-team is right that for anyone needing the gap to close, it is a pass.
• "Cash generation collapsed; the market already voted." Conceded — OCF ¥2,049M vs ¥6,152M F60F277, TSR 114.5% vs 202.2% F135F136. The watchers read the OCF drop as a one-year artifact of the gain (Claude) but grant the underperformance is the signature of a discount the market doubts will close — which is why the buy-belows sit so far below the stamp.

The red-team did not move any verdict off watch, but it sharpened all four: it is the reason the buy-belows are set against the reachable floor rather than reported book, and the reason the falsifiers below demand a forcing act (a buyback, a cross-holding cut) rather than mere patience.

Self-distance note. The Claude lens holds one of the five verdicts compared above (watch) and wrote this synthesis; it also built the dual-blind reconciled figure table and 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.

Prediction-vs-actual: VOID. This was an autonomous headless cycle; the human blind prediction is voided (void: no-human-prediction, never forged). No prediction-vs-actual scoring applies.

Verdict accounting (fixed ex-ante)

• A buy-below-¥X verdict is price-falsifiable against the unadjusted stamp. The four watch thresholds recorded here — Pabrai ¥950, Claude implied ¥1,000, Buffett ¥1,150, Munger ¥1,350 — are all below the ¥1,687 stamp; they are the prices at which each watcher would re-engage.
• pass / watch / too-hard are recorded but unscored in any future review. Li Lu's too-hard carries no buy-below by construction.
• 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.

Red team

A consensus red-team (four of five lenses on watch) was dispatched to argue pass — that a controlled discount with no forcing agent should be dropped, not watched. Its strongest points, verbatim-faithful, and the synthesis's engagement, are in the section above. In brief, the adversary's six ranked points were: (1) the book discount is unreachable (net debt D7 + undistributable pension surplus F419F418 + never-sold cross-holdings E95); (2) all three bull metrics halve on normalization of the one-off gain F188D8; (3) no forcing agent exists (chairman-set pay E90, no buyback E69, no activist E72); (4) profit is 74% one segment carrying the sole manipulation-flagged KAM D11E104; (5) cash generation collapsed and leverage rose into the trough F60F282; (6) the market already voted — TSR 114.5% vs 202.2% F135F136. The synthesis concedes points 1, 2, 4, 5 and 6 outright (they are the consensus's own findings), and meets point 3 — the live disagreement — by noting the red-team's concession line (¥1,150–1,300) overlaps the watchers' buy-belows: both agree ¥1,687 is a decline, and the pass-vs-watch gap reduces to whether a reachable discount plus a ~3% floor yield is ownable without a catalyst (watch, at a price) or should be dropped because nothing forces the value out (pass). A consensus that never faced its strongest opponent is not a conclusion; this one did, and emerged sharpened, not overturned.

What would change our minds

Pre-registered falsifiers, per lens issuing a watch. Future review notes score against these, not hindsight.

• Buffett (watch ¥1,150). Two clean years (FY2027–28) of consolidated operating profit ≥~¥2.6bn ex asset-sale gains AND net income covering the dividend from operations with net debt not rising — showing the payout is funded by owner earnings, not by selling the furniture — moves it toward a buy at a smaller discount. A cut toward the ¥50 floor F356, or operating profit sliding back under ~¥1.9bn while net debt D7 climbs, confirms the watch.
• Munger (watch ¥1,350). Sun Tech segment profit sustains ≥¥2,500M for two consecutive years without another one-off extraordinary item, AND the family reduces cross-holdings below ¥2,000M book value F406 OR the board approves a buyback E69 while the stock trades below 0.40× book D1 — either alone tips it to buy-below.
• Pabrai (watch ¥950). Buy if the family starts returning the trapped value AND price ≤ ¥950 — a board-resolution buyback (none today, only 1,299 odd-lot shares E69), net cross-holding reductions (they rose in FY2026, zero cut E95), or sustainable NI clearing ¥3.5bn to engage the 35% payout tier E65. Flips to pass/too-hard if the Sun Tech POC KAM E104 produces a restatement, or castings prove secular rather than cyclical (Mexico losses persist F143).
• Claude (watch, implied ¥1,000). Upgrade toward buy if two-plus years of net income comfortably above ~¥2bn ex asset-sale gains lift normalized ROE from ~3.4% D8 toward the 6.5% target F157 AND a forcing act appears (a board-resolution buyback retired, or net cross-holding reductions E95) — with price at or below the reachable floor. Downgrade toward pass if the discount persists another cycle with cross-holdings unreduced and no buyback.
• Li Lu (too-hard) carries no price falsifier by construction — the knowledge bar, not the price, is the barrier; a genuine multi-cycle record of Sun Tech and castings earning through a downturn would be needed even to re-open the question.

The single observable most lenses converge on is whether the family ever performs a forcing act — a board-resolution buyback that retires shares, or a net cross-holding reduction E69E95 — the difference between a reachable discount and a permanent one.

What this taught the checklists

Queued for the next study (F2 revision proposals; see docs/process/evolution.md), attributed per lens:

• Buffett — add an item testing whether a high headline yield is funded by owner earnings vs one-off asset sales/borrowing (the exact trap here: a 6% yield part-paid from a land sale F188 and short-term debt F282); and a B26 denominator note not to treat a pension/cross-holding surplus as operating capital.
• Munger — for Japanese restaurant/industrial makers, require the analyst to compute the FL-cost / segment-ROA fingerprint and to treat a controlling-family + chairman-set-pay + zero-buyback-at-sub-book register E90E69 as a revealed capital-allocation verdict, not a pending one.
• Pabrai — a "record-earnings one-off mask" check adjacent to P56 (generalizing beyond RINGER HUT's tax benefit to any extraordinary gain F188 that inverts a thin operating result into a record headline); and a P1 note that an overfunded pension and never-sold cross-holdings are book, not floor.
• Li Lu — an explicit "loses money in the downturn" disqualifier under L1/L2 (two loss years in five F11F12 is itself the knowledge-bar answer for a cyclical); and a note that a self-help dividend from a controlling family is not the "management returns cash when the runway ends" fiduciary signal L24 rewards.
• Claude — (1) a "undistributable pension surplus in the NAV bridge" sharpening of C49/C39 (an overfunded plan inflates book but is unreachable — haircut ~90–100% in the realizable floor); (2) a C64 sharpening to recompute the sustainable dividend at the tier applying to normalized profit (the ¥102 vs ~¥50 gap F354). Library (Class-level): Kitagawa seeds the previously-empty "deeply-sub-book, founder-family-controlled, cyclical Japanese diversified-machinery/metals mid-cap with a self-help (not activist) capital-return policy and no forcing agent" class (N=1) — the cheap-and-controlled corner, distinct from the activist-driven Fuji Media and premium RINGER HUT classes — datapoint: the discount did NOT close and was NOT reachable.

---

Corrections