The 28 items in a fund's first request

A private credit fund asks a non-bank lender for the loan level file first, then for eight blocks of documents that let it rebuild every number the lender has claimed. The data request we send runs to 28 items, and 8 of them ask the lender only to reconcile its own documents against each other.

The question has three links, and the two answers you find online belong to the links either side

One published checklist answers what a limited partner asks a private credit fund, in six blocks running from fund structure to operational due diligence. Another answers what a lender asks a single borrower. The link in between, what a fund asks the non-bank lender whose loans it is about to fund, is the one nobody has written down as a list, and it is structurally different from both of its neighbours.

The difference is not tone. A fund underwriting a lender is not underwriting a company and not underwriting a borrower. It is underwriting a machine that produces borrowers, and it has to keep valuing that machine after the money has left, which is why the request is loan level rather than entity level, why it repeats on a cadence rather than once, and why so much of it is not new information at all.

The list below is the data request we send, generalised away from the lender it was written for. In July 2026 we read a screening package from a non-bank lender with a book of about MXN 124 million and sent back 28 items in eight lettered blocks before the file could go to a debt fund at all.

Eight of the twenty eight ask the lender to reconcile its own documents against each other, and not one of the twenty eight asks for a forecast. A first institutional request is an audit of internal consistency, and consistency is the one property a lender can fix completely before anybody asks for it.

Block A asks you to reconcile your balance sheet with itself before anybody looks at credit

Netting is the first thing an analyst finds and the fastest. In the package we read, loans were netted against an identical collateral line down to zero, which moved 65 per cent of the company's debt off the face of a balance sheet signed under oath, and the same arithmetic implies either true leverage of 7.59 times or a portfolio overstated by 135 per cent, per our own published credit memo. There is no third reading, which is why the request opens with a gross restatement rather than with a question about credit.

Block B is the loan tape and the cohort curves, and nothing in the file substitutes for either

A loan tape is the loan level register: one row per loan at a stated cut-off date, carrying identity, terms, borrower, collateral, balances and payment behaviour. The fund also asks for cohort loss curves by origination year, that is cumulative charge-offs as a share of principal originated measured at each month on book, because an annual average hides a book whose behaviour changes cohort by cohort.

Then it asks you to reconcile your own weighted averages back to that file. In the package we read they did not reconcile: the weighted average yield was a hardcoded formula sitting on a base that did not equal the portfolio, and a claimed default rate of 0.02 per cent sat beside 1.41 and 1.34 per cent in the same workbook, a discrepancy of about seventy times recorded in our own published credit memo.

Block C asks what you reserve, what you restructure and what you actually recover

Provisioning is where an unaudited profit becomes a modelled one. On that anonymised book, a reserve of 1.5 per cent eats 42 per cent of annual profit, 3 per cent eats 84 per cent and 5 per cent turns the year into a loss, per our own published credit memo. A fund models the normalised figure, so the absence of a provisioning policy does not read as conservatism. It reads as missing risk accounting.

Roll rates, restructuring counts and recovery history complete the block, and each answers something the tape alone cannot. Roll rates show migration between arrears buckets over the last year, restructuring counts separate a cured loan from a renewed one, and a recovery series that reads zero in every single year is a statement about the collections function rather than about the borrowers.

Block D reads your funding stack as a queue, and finds out who is genuinely senior

In the package we read, 42 facilities were all marked senior secured while 37 of them recorded collateral as not applicable, and about 30 individual lenders held 64.2 per cent of the debt with no stated maturity at all, per our own published credit memo. A queue nobody has ordered is not a capital structure, and a fund that cannot find a free first lien will structure around your existing book rather than lend against it.

Block E establishes that the entity may lawfully lend, and a data feed closes none of it

Our own published checklist of 37 diligence items records that a live data connection closes 18 of them outright and 4 in part, and that the 15 it leaves are register printouts, notarised instruments and countersigned agreements. Every item in this block sits in that 15, which is good news about the calendar: procurement can run in parallel with everything else, and it does not wait on engineering.

Two of the items are specific to Mexico and both are cheap to answer early. Lending as a habitual, professional business is a vulnerable activity for anti money laundering purposes under article 17 of the governing statute regardless of licence category, and a SOFOM, the Mexican non-bank finance company category, is constituted under the Ley General de Organizaciones y Actividades Auxiliares del Credito rather than licensed by a banking supervisor, so there is no licence to show and the fund knows it.

Block F asks how repayment is enforced, and who is carrying the currency

Enforcement is a legal question with an arithmetic consequence. Where a consumer loan is repaid out of a state pension that cannot be attached, the security that legally works is a perfected lien over the borrower's withdrawable retirement balance, and the fund will ask what share of the pledged pool actually carries one. Currency is the second half: a facility drawn in a currency your book does not earn converts an exchange rate move into counterparty risk, so the fund asks for your own solvency at three named rates rather than for a hedging line item.

Block G reconciles the size of the ask against the capacity you already have

In the package we read, the lender asked for US$1.5M while about MXN 37.81m of committed capacity sat undrawn in its own funding file, at a target borrowing cost roughly equal to its existing weighted average cost of funds of 24.2 per cent, per our own published credit memo. Neither fact is disqualifying on its own. Both unexplained together are, because the fund cannot tell an origination constraint from a liquidity problem.

Block H reconciles the deck with the accounts, because the deck is the control sample

A deck is not evidence, it is what the evidence gets checked against. In the package we read, five months of revenue already converted to about US$1.95M against a claimed cumulative US$8.5M across four years, and the incorporation record disagreed with the stated operating history by about eight years, per our own published credit memo. A deck that cannot be tied to the accounts turns every other number into a claim.

Every answered item becomes a term, and every unanswered one becomes a punitive default

The list reads as an obstacle course and behaves as a pricing form. A filed auto securitisation states that «The Delinquency Trigger was calculated as a multiple of 4 times the previous historical peak delinquency percentage observed during the period», and the SEC filed language is explicit that the level sits above the historical peak so the trigger is not tripped by ordinary fluctuation.

A lender that cannot produce a clean delinquency series by cohort gives the fund nothing to multiply, and an absolute number is imposed instead. Your data is the covenant, and that is not a metaphor: the covenant is arithmetic performed on the history you can evidence.

The borrowing base works the same way, that is the running arithmetic saying how much you are allowed to have drawn against your collateral today. Eligibility decides what enters it: «a typical borrowing base eligibility criterion for a warehouse facility is a requirement for the underlying loan's documentation to permit the sale or transfer of the loan asset to the SPV borrower, the pledge thereof to the facility's administrative agent, and a foreclosure on the loan asset by the administrative agent». A form contract that forbids assignment does not fail credit. It funds nothing while remaining entirely yours.

The dictionary is a named deliverable in the same list, not a courtesy: one fund checklist asks the lender to «Include a data dictionary explaining all codes, column headings, and non-obvious terms», and a second names «Automated reporting capabilities» among the things to have working before the raise rather than after it.

The calendar is the last price. Twelve weeks is the published best case for a well organised issuer, a16z puts the realistic range at three to six months and warns that drafting the credit agreement alone «will take 2-3 months and can take longer depending on counsel availability», and the one measured Mexican datapoint ran 139 days from board approval on 1 May 2020 to signing on 17 September 2020.

Effect is not a bank, a lender, a broker-dealer, an investment adviser or a credit rating agency. Nothing here is investment, legal or tax advice, and no outcome is promised. The twenty eight items are our own data request generalised away from the lender it was written for, and the anonymised figures are our own reading of the package that lender supplied, at a 31 May 2026 cut-off.