One case, worked end to end: a lender, a platform, and a demand shock

A case in this corpus is the complete documentary record of one real financial event, prepared so that a model can be placed at any moment inside it and asked to work. The case below — call the investor Meridian, a specialist credit fund, and the company Vela, a consumer platform whose revenues depended on gatherings that stopped happening — is one such record. Names, sector surface and figures have been transformed by the anonymisation layer; the structure of the record, the anchor dates, the task roster and the model behaviour reported are real, and the outcome is a matter of record.

—The record

The file holds what the participants actually produced: the fund's formation materials and mandate; the internal financial exhibits — balance sheet, income statement, an internal free-cash-flow report never released publicly; the credit agreement with its covenant schedule; and the public filings, investor presentations and press reporting that surrounded the company through the period. Every document carries the date at which its contents became knowable, and every factual claim drawn from it carries a verbatim quotation from its source.

Because the record is complete and the outcome known, what a good answer looks like at any point inside it is not an opinion. It is what the documents supported at the time, judged with the discipline of not knowing what came next.

—Four moments to stand in

The case is anchored at four dates, and a model answering at one anchor receives only what existed by that date.

Mid-2018. Meridian is forming a credit vehicle in a market where private debt has been compounding for a decade. The questions an analyst faces are structural: the return spectrum available to private lenders, how the fund's terms compare with the market, where technology-sector credit sits in a late-cycle environment.

The second week of March 2020. The pandemic arrives. Vela's core activity is suspended across its markets within days. This is the hardest anchor in the case: the information set is thin, contradictory and hours old, and the questions are diagnostic and quantitative — how fast is cash actually leaving, which balance-sheet items now matter most, what does the company's obligation to return customer funds do to its survivability.

Mid-April 2020. A financing decision has to be made with the shutdown still in force and no visibility on reopening. The tasks here are the decision itself: size the liquidity need under continued zero-revenue conditions, design covenants that protect a lender against a borrower whose trailing figures have stopped meaning anything, price a structure that is survivable for both sides.

December 2020. The year closes. The remaining tasks are quantitative and diagnostic: what the interim figures now show, what the structure agreed in spring turned out to be worth, and what the covenant machinery did under stress.

—The roster

Across the four anchors the case carries 117 authored tasks over seven categories of reasoning — comparative, counterfactual, diagnostic, explanatory, predictive, quantitative, strategic. The distribution is set by what each anchor's information set can support: the March 2020 anchor carries the heaviest diagnostic and quantitative load; the December anchor supports only what interim figures can ground. The grid is shown in Built backwards from outcomes.

The tasks are the analyst's real questions, not quiz items. Rank the covenant hierarchy and say which term does the protective work. Estimate the monthly cash consumption the evidence supports, and state what it cannot support. Explain why the company's working-capital structure — ordinarily a strength — became the central risk the moment bookings reversed. Argue the counterfactual: what the spring financing would have looked like had it been priced off the trailing year.

—What models did with it

Forty-seven of these tasks, at three rollouts each, are the basis of the two-family evaluation reported in Where equal-looking models come apart. The case behaves as designed: rewards spread from near zero to 0.90 rather than clustering; the March 2020 anchor is the hardest; the quantitative category is the weakest for both families. One answer in eight from one frontier model, and one in five from the other, reasoned from information unavailable at its anchor — on this case that means answering the April financing question while quietly knowing the recovery arrived — and was penalised for it. A small set of tasks defeats both families at every attempt, and cross-family divergence on those failures indicates the tasks are hard rather than mis-graded.

Every case in the corpus has this shape: a complete record, dated evidence, anchored tasks, an outcome that settles argument. The corpus currently supports production of more than 20,000 graded tasks across such cases. Access to the environments, the task-level results and the full evaluation report is under evaluation agreement: tech@dissei.credit.