A $449M ambiguity in Carvana's new ABS-EE filings

Two tapes, one trust

On August 11 at 22:02 UTC, Carvana's shelf filed a complete ABS-EE loan tape for Carvana Auto Receivables Trust 2026-P3: 52,339 loans, $1,550,309,023.48 of initial pool balance. Six minutes later, at 22:08, the same shelf filed a second complete tape for the same trust: 37,104 loans, $1,101,231,512.91. Both parse cleanly. Both tie to the preliminary prospectus to the cent. Only one of them is the deal, and nothing public says which.

This is not a filing error. The Rule 424(h) preliminary prospectus filed the next morning identifies both schedules explicitly: the smaller tape is the pool "if the Base Transaction is issued" and the larger is the pool "if the Upsize Transaction is Issued." The offered notes come to $1,046,167,000 in the base case or $1,472,790,000 in the upsize case, and the document states plainly that "the Sponsor will make the determination regarding the aggregate initial principal amount of the Notes based on, among other considerations, market conditions at the time of pricing."

I compared the two tapes loan by loan. The base pool nests strictly inside the upsize pool: all 37,104 base loans appear in the upsize tape, 15,235 loans appear only in the upsize tape, and zero loans appear only in the base. The difference between the two candidate trusts is 15,235 loans and $449 million of principal.

As of August 13, EDGAR shows no 424B5 and no 8-K for this trust. The election has been made in a pricing call somewhere, but it is not yet public information. Anyone loading both ABS-EE filings today faces the same question we do: which 2026-P3 is real?

What our pipeline does with a question like that

It refuses to answer.

Every new trust that appears in our ingestion goes through a quarantine lane before any of its loans reach the published data. A first filing carries no continuity evidence: there is no prior month to reconcile against and no subsequent month to confirm it. Our automated pool resolver admits a lone new-trust candidate only in narrow, provable cases, and a two-candidate pair like this one has a specific rule: a later filing that contains the earlier one is treated as a superseding refiling. Here the superset arrived first, at 22:02, and the subset second. Acceptance order contradicted the refiling reading, so the rule never fired and both tapes fell through to quarantine for review.

The review agreed with it. Both schedules are legitimate; the ambiguity is real and external. So both accessions sit in a pending state, visible in our internal monitoring with their age in days, and neither contributes a single row to the published dataset.

If we picked one tape now, we would have even odds of publishing a trust that is wrong by 15,235 loans and $449 million. Every statistic for that trust would carry the error: pool balance, delinquency denominators, weighted-average coupon. Included loans also acquire stable identities and begin accruing monthly history, so unwinding a wrong guess a month later means surgically removing a phantom population from every table it touched. We have done that kind of cleanup before, on a lease trust filing that reached the data through an older, more permissive version of this pipeline. That is the most expensive kind of mistake a loan-level dataset can make, and catching it before ingestion costs nothing at all.

Multi-pool filing sets are routine for new issues, which is why the lane exists at all. Honda's most recent trust filed three complete tapes in one day, a statistical pool of 65,408 loans, a pricing pool of 76,270, and a final pool of 87,213. Only the final pool belongs in the data, and only the final pool is in ours.

Same day, opposite outcome

A filing is held until the evidence decides it, and sometimes the evidence is already there. World Omni Auto Receivables Trust 2026-C arrived the day after Carvana's pair and cleared review the same day it landed.

World Omni habitually files base and upsize tapes in pairs, minutes apart; 2026-B, 2025-C, 2025-B, 2024-A and 2023-D all did. So when 2026-C appeared as a single tape with base in the filename, the natural reading was a missing sibling. The evidence said otherwise. EDGAR lists exactly one ABS-EE filing for the trust, with no second tape in the following days. The preliminary prospectus contains zero occurrences of "upsize" in 682,345 characters, where prior World Omni deals disclose both alternatives. And it discloses exactly one pool: 34,343 receivables, $1,084,440,413.63 as of a July 9 cutoff. That count matches the filed tape exactly, and the disclosed balance sits between the tape's July 1 beginning balance ($1,092.4M) and July 31 ending balance ($1,059.5M), consistent with a mid-month cutoff.

One disclosed pool, one filed schedule, counts equal to the unit. That is sufficient evidence, and WOART 2026-C was admitted on the strength of it: 34,343 loans now live in World Omni's July coverage. Two new trusts landed a day apart, one went in and one is still waiting, and the only thing separating them is what their documents could prove.

How the Carvana question answers itself

Trusts file monthly servicing tapes as ABS-EE exhibits alongside their Form 10-D distribution reports. Carvana files on roughly a 12th-to-14th cadence, so 2026-P3's first servicing tape, covering August, is expected in mid-September. That tape will report the pool the trust actually owns. Its population will match exactly one of the two candidates sitting in our pending lane, and our resolver will process the three filings together: one candidate matches the successor population and is admitted, the other does not and is excluded. Nobody has to guess, and nobody has to sign anything off.

Two things could resolve it earlier. If Carvana files the final prospectus or an 8-K naming the issued transaction, the review can be closed on that document immediately. And if the sponsor upsizes or restructures before the expected August 26 closing, the filings themselves will say so. Until one of those happens, 2026-P3 is absent from our dashboards, and that absence is the honest reading of a deal whose composition is not public yet.

For readers tracking the new-issue calendar, there is one number to watch in September. If the first servicing tape carries roughly 52,000 loans, the upsize was issued; roughly 37,000 and it was the base. Either way, the trust enters our data with its identity proven on its first day.

Data note

Filing facts come from EDGAR: acceptance timestamps, accession numbers 0001770373-26-000036 and 0001770373-26-000038, and the Rule 424(h) preliminary prospectus for the trust, all pulled August 12 and 13. Loan counts and balances come from parsing both EX-102 tapes in full; the nesting analysis compares distinct loan identifiers across the two schedules. The World Omni comparison uses the same sources for its trust. Prospectus quotes are verbatim. The one fact this post cannot state is the one that does not exist publicly yet: which transaction Carvana priced. The loan-level data behind our published trusts, each of which passed this same admission process, is available on every plan.