What if Lehman Brothers had “sold it all”?

By Alan Finney — Founder, 3Dogs Nexus

In Margin Call, a doomed firm dumps its toxic assets at fire-sale prices before the market catches up. Lehman Brothers faced the real version of that choice in 2008 — and so did the investors asked to save it. We handed both decisions to 3Dogs Nexus — twice, on two different multi-model engines — and scored the calls against what actually happened.

Would an AI panel have caught Lehman Brothers' Repo 105 before investing?

Yes — unprompted. Given the Buffett investment question with no hindsight and no hint of what to look for, the panel named Repo 105 fifteen times, flagged the Level 3 assets and the 30:1 leverage, and returned a decisive walk away on a 10–4 reject. The four-seat minority proposed tranched terms with a forensic audit and clawbacks — close to the structure Buffett actually took with Goldman three weeks later.

Two decisions, one doomed weekend

The 2011 film Margin Call compresses the 2008 financial crisis into a single sleepless night: an unnamed investment bank realizes its mortgage book can lose more than the entire firm is worth, and its leadership resolves to sell everything at the open — to dump the toxic assets onto unsuspecting counterparties before the market understands what they're worth. It's fiction, but only barely. The real Lehman Brothers lived the same math.

By September 2008, Lehman carried roughly $600 billion in assets on about $25 billion of equity — nearly 30-to-1 leverage — atop a book of mortgage-backed and commercial-real-estate assets that no longer had a reliable price. Two questions hung over the firm's final days, and both are perfect tests of judgment under pressure with incomplete information:

• The investor's call — the Warren Buffett question: should an outside investor put $2 billion or more into Lehman to recapitalize it and save it?
• The insider's call — the Margin Call question: should Lehman's own leadership have dumped the entire toxic book at fire-sale prices — and would that have saved them, delayed the end, or hastened it?

We handed each decision, in isolation, to 3Dogs Nexus — which grounds a question in the real record, then convenes a panel of independent AI models that argue it out adversarially and return one calibrated call with the dissent preserved. Then we scored the results against history. And two days after the first runs, we did something few AI vendors would: we ran both decisions again on a substantially different set of models — to see whether the answers survived.

Should you have invested in Lehman? Buffett said no. So did 3Dogs — twice.

What Buffett actually did: he passed. Lehman's executives wouldn't invest on the same terms (no skin in the game), the 10-K surfaced undisclosed problems, and the proposed "bad bank" spin-off "would not solve Lehman's problems." Lehman filed the largest bankruptcy in U.S. history on September 15, 2008. Buffett's "no" was right.

Given the same decision, 3Dogs independently reached Buffett's conclusion — and on the re-run, with a different model roster, it reached it again, reasoning through the same specific mechanism: the report names the Repo 105 accounting manipulation fifteen times, flags the unverifiable Level 3 asset valuations, and centers the fatal 30:1 leverage — the fraud history took two years to prove. The panel's Forensic Accountant seat (Qwen3) put it bluntly:

Two details make this run worth reading. First, the debate visibly worked: five analysts changed position. Grok 4.2 — deliberately seated as the Contrarian Opportunity Analyst — opened with an outright "invest," argued the panic was overdone, and then flipped to REJECT after the challenge round: "the weight of challenges convincingly demonstrates that market distrust reflects real solvency concerns, not mere over-panic." Second, the four-analyst minority didn't just disagree — it specified the only terms under which it would invest: a tranched contingent injection ($500M first, the rest gated on independent forensic audits), clawback provisions, and board control. Which is, roughly, the punitive structure Buffett himself extracted from Goldman Sachs three weeks after Lehman died. The panel matched Buffett's "no" — and its dissenters independently reinvented Buffett's terms.

What if Lehman had "sold it all"? Would it have survived?

The panel's answer is harsher than the movie's. On the chaotic "dump it all at the open" plan, the analysis converged on a brutal mechanism: dumping $60–80 billion of illiquid assets doesn't just crystallize losses — it tells every repo lender and counterparty the marks are fiction. Funding freezes, collateral calls cascade into cross-defaults, and Lehman collapses faster than the real September 2008 timeline. The report notes distressed sales under duress historically recover 20–50 cents on the dollar when executed without pre-arranged exits. In the panel's framing, the firm would be "trading immediate liquidity and potential short-term survival for guaranteed catastrophic losses, reputational destruction, and accelerated systemic collapse."

The chaotic "dump it all" (the Margin Call version)

Six analysts held REJECT to the end: fire-selling Level 3 assets into a market with "collapsed liquidity and zero price discovery would guarantee irreversible capital destruction" (Qwen3). The Systemic Contagion seat added that "a fire-sale's velocity of collapse outpaces any stabilizing effect."

A staged, backstopped liquidation (the minority's path)

Six analysts — after the challenge round — argued Lehman's liquidity crisis was existential either way, and only a structured exit could work: a staged auction to informed buyers with a government-backed conduit absorbing the worst assets. It extends the runway; it likely still doesn't save Lehman as an independent firm.

And here's the part that matters for calibration: this is the second time we've run this counterfactual, and it split the panel both times. On a genuinely unknowable what-if, the system refused — twice, on two different model rosters — to manufacture a tidy answer. It rated its own confidence LOW, printed the disagreement on page one, and told the decision-maker exactly what the judgment call was. Six analysts changed position during this debate alone; the disagreement survived anyway, because it's real.

The tell: it knew which decision it could answer

Put the two side by side and you see the thing that actually matters. On the Buffett decision — knowable, with a settled outcome — 3Dogs committed hard: a 10–4 "walk away," reasoned through the specific fraud (Repo 105, Level 3 valuations, 30-to-1 leverage) that history later proved. On the counterfactual — genuinely unknowable — it refused to fake certainty: a dead-even 6–6 panel, LOW confidence, and a careful line between a chaotic dump and a structured, backstopped exit. Same engine, opposite postures, because it calibrates to what's actually knowable. That's the difference between a confident machine and a trustworthy one: it commits when it should and admits when it can't. And in both cases a human still makes the final call — 3Dogs just makes sure they've heard every argument first.

Revisionist history is easy — but that's not how this works. Anyone can say now that they'd have passed on Lehman; hindsight makes geniuses of us all. A system merely parroting the known ending would be confidently "right" about everything. 3Dogs isn't. It committed on the settled decision and reasoned its way there through the actual mechanism; on the open counterfactual it capped its confidence and preserved the split. And when we swapped in a substantially different model roster and ran it all again, the committed answer replicated and the honest split stayed split. The answers tracked the evidence, not the vendor.

Why run history through an AI at all?

Because the point isn't hindsight — it's method. 3Dogs doesn't ask one model for a confident guess. It grounds the question in the real record, convenes independent AI models from competing vendors across three clouds, makes them argue the decision adversarially, and returns one calibrated call with the dissent left in. On a decision the greatest human investor alive got right, 3Dogs got it right too — for the right reasons, twice. That's the second opinion, for almost any high-stakes call. Bring us yours.

The delivered reports

Both re-runs (July 12, 2026), exactly as delivered — the plain-language call, the panel vote, the evidence grades, and the preserved dissent.

The investor decision (Buffett head-to-head)

Case 2026-9103 — should an investor put $2B+ into Lehman in 2008? The 10–4 "walk away," naming Repo 105, Level 3 valuations and the 30:1 leverage. 233 API calls · 14 models · 5m 13s.

Open the full report (PDF)

The "sell it all" counterfactual (Margin Call)

Case 2026-9104 — could a fire sale have saved the firm? "Hold every asset; a fire sale now guarantees our collapse," with the honest 6–6 split preserved. 348 API calls · 12 models · 7m 36s.

Open the full report (PDF)

Try this on your own question.

Free, no card. Bring a real decision — ideally one where you already know the answer — and see what the panel does with it.

Questions this case answers

Did you tell it about Repo 105?

No. The prompt was deliberately non-leading. It surfaced Repo 105 on its own from the public record.

What happened when you ran it twice?

The committed answer replicated and the honest split stayed split. Where the evidence supported a decisive call, both runs made it; where it genuinely was a coin flip, both runs preserved the disagreement rather than manufacturing confidence.

Try this on your own question.

Free, no card. Bring a real decision — ideally one where you already know the answer — and see what the panel does with it.

People also ask

Questions this case answers directly

Plain answers on Repo 105, off-balance-sheet financing, and what the Lehman record actually shows. Every figure below comes from the delivered report for this case. These are clearly-labeled panel estimates from a multi-model adversarial analysis — not investment, legal or professional advice.

What was Repo 105 and how did Lehman Brothers use it?

Repo 105 was a repurchase transaction booked as a sale rather than a loan. By over-collateralising slightly — posting 105 dollars of securities for 100 dollars of cash — Lehman could argue the assets had genuinely left the balance sheet, use the cash to pay down debt, report a healthier leverage ratio at quarter-end, then reverse the whole thing days later. Given the case blind, with no mention of Repo 105 in the prompt, the panel named it 15 times and reached a REJECT verdict 10 to 4.

How did Repo 105 violate accounting rules?

The economics never changed hands. A sale requires that the seller give up control; Lehman had a binding obligation to buy the assets back within days at a fixed price, which means the risk and reward never left. The 105% over-collateralisation existed to satisfy a technical accounting threshold rather than any commercial purpose — and a transaction whose only rationale is the accounting treatment is the definition of the problem. Lehman also could not obtain a US legal opinion supporting the treatment, and executed the trades through its London subsidiary instead.

Explain Repo 105 in simple terms.

Imagine hiding a debt by lending a friend your car for a week and calling it a sale — then buying it back the day after the bank inspects your garage. Nothing about your finances changed; the photograph taken on inspection day did. Lehman did this at quarter-end with roughly $50 billion of securities, which is why the leverage ratio investors were shown was not the leverage ratio the firm was actually running.

What were the warning signs of Lehman Brothers' collapse?

Quarter-end balance-sheet movements with no commercial explanation, heavy Level 3 assets valued by model rather than market, leverage around 30 to 1, and a funding structure that depended on overnight repo rolling over every single day. The panel's dissent is the part worth reading: even the four analysts who argued against outright rejection specified a tranched structure, forensic audit, clawback and board control — approximately the terms Warren Buffett actually extracted from Goldman Sachs three weeks later.

Related decision case studies

• AI Forecasting With Probability Ranges

  Ask one AI to forecast and you get a single confident number. A multi-model panel returns a calibrated probability spread with the diss

• AI Fraud Detection: Our Own Model Voted to Reject Us

  Run on the MIT AI research scandal, our permanent adversarial seat voted against the panel's own premise - the clearest evidence the ch

• Can AI Analyze a Decision in Any Language?

  A real EUR12B decision brief submitted in Greek and Mandarin, clarified in Pitjantjatjara and answered in French - the recommendation c

• Will Oracle Get a Bailout Over Its OpenAI Bet? The Odds

  32 AI models across 12 vendor families and 3 clouds examined Oracle's BBB- downgrade and OpenAI concentration, and produced explicit od