Loans that come back: 33 held ABS-EE trust-months

33 trust-months, no bad filings

The July 2026 monthly ABS-EE wave arrived on August 17 and 18. Of the 102 July-period filings, 69 passed the continuity gate and were applied. The other 33, one per trust, were held: six CarMax, one Drive, 17 Ford, nine Santander, every one with a report date of July 31, 2026. The nightly review email began counting them against a five-day page threshold, and Santander's August shelf data was held out of the published history behind them.

I went through all 33. None was a wrong filing. Every loan the gate flagged as new was original collateral, present on the trust's first tape. What the gate had caught was loans coming back: loans that paid off, charged off or terminated in an earlier month and reappeared on the July tape in a shape the rule did not yet recognise. This post covers those shapes, what each one means economically, and how a fail-closed gate was widened to admit them without admitting anything else.

A tape is not a ledger

A monthly servicing tape, filed as an ABS-EE exhibit alongside the trust's Form 10-D, is a snapshot of the pool the servicer reported that month. It is not a running ledger in which every loan gets one row per month until a terminal event closes it out. Loans drop off when they pay to zero. They come back when something happens to them afterwards: a borrower reinstates, repossession proceeds arrive, a leased vehicle is sold at disposition, or a residual of a few dollars is collected.

Our admission rule for a monthly filing compares the population on the new tape against the trust's existing history. The version in production for the July wave was the third, and the first that was fully fail-closed: an exact-population match with narrow exceptions, and quarantine of the whole trust-month for anything else. That is deliberate. I wrote last month about why a new trust with two candidate pools is held in quarantine rather than resolved by a guess. The same rule that refuses to guess on a new trust also refuses to guess when a known trust's tape contains a loan it last saw at zero balance.

The difficulty is that "gone" is not a state a loan tape enforces.

The four shapes

Reinstated loans

The Santander Drive 2024-4 loan was last seen at $0 on July 31, 2025, and came back on the July 2026 tape with $6,256.34. The v3 exception for a returning loan required the returning row itself to be all-zero. That works for an inert re-listing and fails for a reinstatement, which by definition carries a balance. Of the four shapes this one has the largest effect on published metrics: a loan that was out of the pool for a year is back in it and will be in every delinquency denominator from here on.

NULL-amount placeholders

Ford Credit Auto Owner Trust filings re-listed between one and 50 previously dropped loans per trust with every amount field NULL. Not zero, absent. The v3 rule required a not-NULL zero balance. A blank re-listing carries no economics; the servicer is keeping the identifier on the tape.

Recovery-proceeds returns

CarMax and Drive re-report a zero-balance loan in the month repossession proceeds arrive: $1,445 in one case, $3,770 in another, both above the $1 threshold the rule used to decide whether a returning row was inert. Ford lease trusts show the same shape in larger numbers. Leases terminate with the contract balance still populated, $999 to $97,562 across the wave, and a termination date set, then reappear months later carrying the disposition proceeds: between 24 and 151 such returns per Ford lease trust this wave. These rows are where the recovery side of a cumulative net loss figure comes from, so a gate that excludes them excludes the recoveries.

Residual-balance payoffs

Santander Drive 2022-5 and 2022-7 each carried a loan amortized to $3.94 and $1.88 respectively as of June 30, 2026. Both paid off and dropped from the July tape without a terminal marker. The active-loan test treated anything above $1.00 as active, so each was an active disappearance, and each quarantined its entire trust-month. One $1.88 loan held one trust-month.

How the rule got here

Each version of the rule was widened by a specific case, not in anticipation of one.

Rule v1 compared the new tape only with the immediately prior month. Fifth Third Auto Trust 2023-1's June 2026 filing exposed the gap: 37,084 loans, 37,073 present in May, 11 not. All 11 had been in the trust's initial July 2023 population; ten were last reported at zero balance in March or April 2026 and one in May 2025. Every one re-entered at zero beginning balance, zero ending balance and zero activity of any kind. None matched any of the 1,242 legitimate May exits under an exact identifier, a normalized identifier or an economic fingerprint. May's ending balance and June's beginning balance were both $378,324,430.05. Rule v2 admitted exactly that shape: a same-identifier, economically inert re-entry from the initial population whose last observation ended at zero. Across the four pending filings at the time, that one decision changed and no other did.

Rule v3 turned that into exact-population continuity with the v2 carve-out attached, and that is the version that was in production for the July wave. Its carve-out tested the returning row: is it all-zero, is it not-NULL, is there under $1 of activity. Three of the four shapes above fail that test for reasons that have nothing to do with whether the loan belongs to the trust.

Rule v4 tests the loan's history instead. The carve-out keys on the loan's last observed state, never on the shape of the returning row. Two conditions: the loan was original collateral, present since the trust's first period, and it was not materially active when last seen. A reinstatement with $6,256.34 passes because the loan was at $0 in July 2025. A NULL placeholder passes for the same reason. A Ford lease return passes because the termination date was set when the lease was last seen. And material activity now shares one $10 de-minimis floor, raised from $1, on both disappearance tests, so the $3.94 and $1.88 payoffs are no longer active disappearances.

What v4 still refuses: loans never observed in the trust, collateral added after the first period, loans that were active at last sight, and materially active loans that disappear. Fifty-five execution tests cover the four shapes and those refusals.

Re-run once against the same settled data, v4 flipped 32 of the 33 quarantines to include: 31 as a prior-period match with returning original loans, and Santander Drive 2022-7 as a plain prior-period match once the $1.88 loan stopped counting. CarMax six of six, Drive one of one, Ford 16 of 17, all nine Santander.

Rule v5 came two days later, from a case v4 was built to refuse. GM Financial Consumer Automobile Receivables Trust 2026-1's July tape has 29,653 loans: 29,648 overlap June, there are no active disappearances, and five rows are returns. Four were in the trust's initial period. The fifth first appeared in January, last appeared in April at zero balance with a terminal marker, and returns in July at zero balance with recovery activity only. Nissan Auto Receivables 2025-B Owner Trust had the same pattern with 13 loans out of 46,239, all first seen in a reviewed November launch-normalization filing, each observed for five to seven months and last seen at zero between March and May.

V4's requirement that a tolerated return be original collateral is what held both. V5 adds a second, narrower class: a later-first-seen loan is admitted only if it was observed in at least two earlier periods, was settled at last sight, is de-minimis at both beginning and end of the returning month, and carries recovery activity only. The refusal list does not change.

What stayed held, and the cost in each direction

By August 21, 25 of the 30 filings on the pending list had applied under v4. Five remained. GM and Nissan Receivables 2025-B are the v5 cases above. The other three are not rule problems, and they show what each kind of error costs.

Ford Credit Auto Lease Trust 2026-B is a correct quarantine. Our June population for the trust is the union of two non-overlapping launch schedules, 43,991 and 35,185 leases, both ingested before the gate existed. The July tape continues the 43,991 exactly, and the other 35,185 read as active disappearances. The defect is in our June, not in Ford's July. The remedy is the one used for Carvana 2026-P3 and BMW 2026-A: include the monthly, record the losing schedule as excluded, and remove its 35,185 rows from the history, where they distort every denominator if left.

Nissan Auto Lease Trust 2026-B is the same defect with the cost already visible. June in our history held both the statistical pool (35,407 leases, $1.448 billion) and the final pool (47,185 leases, $1.931 billion), with zero overlap between them. June coverage for the trust reported 82,592 active assets and $3,379,369,089.10, the exact sum of both schedules. The July filing carries 47,185 leases and $1.898 billion, an exact continuation of the final schedule. The fix is a scoped deletion of the 35,407 statistical-pool rows (the precondition check found zero references to those identifiers from any other month) and a rebuild of June. That is what publishing a population that does not exist costs: a $1.45 billion overstatement of one trust's June coverage, corrected by deleting rows from the last complete month and re-running it.

The cost in the other direction is the one the wave started with: 33 trust-months held from the wave until August 19, a shelf's August data behind them, over returns worth $1.88 at the low end. The gate has to tell a $1.88 residual payoff from a $103 million pool mismatch, and it has to do that from the tape alone.

Exeter Automobile Receivables Trust 2026-4 is that mismatch. The XML carries 42,065 assets and $1.005 billion. The same-day Rule 424(h) prospectus discloses one cutoff-date pool of 45,996 contracts and $1,108,004,343.39. No upsize language, 3,931 loans and roughly $103 million apart. It stays held until a corrected tape or an exact successor resolves the population.

Once GM and the two Nissan filings apply, 28 of the 30 are in the July data, and Ford Lease 2026-B and Exeter 2026-4 are the two that remain.

Two as-of dates

The July wave is also why the dashboard now carries two separate as-of dates. Analytical charts stay pinned to the latest month in which every trust has applied, June 2026. Coverage advances trust by trust as July reports apply. That keeps a delinquency rate from being computed over a denominator that is 69 trusts of July and 33 of June, while still showing which July reports are in.

Data note

Every figure here comes from three internal evidence records: the Fifth Third zero-balance re-entry follow-up of July 30, the v4 re-resolution record of August 19, and the v5 and Nissan split-pool record of August 21, 2026. The Exeter comparison uses the trust's Rule 424(h) prospectus as filed. Rule v4 was executed once, manually, against the July wave after a preview showed exactly 32 flips. The next monthly wave arrives around September 15, and it is the first to run through these rules in production. The loan-level data behind every admitted trust-month is available on every plan.