Category · Field notes
What we found in 100 anonymized pilot files
Aggregate observations from the first hundred files Vyaso re-underwrote in pilot. Patterns, surprises, and what we now look for first.
Key takeaways
- Across the corpus, adjusted revenue diverged from gross deposits more than expected on the high end and less than expected on the low end.
- The most common flag was account kiting, not loan stacking. Multi-account merchants drive this.
- Loan stacking was concentrated in a small minority of files but tended to be severe when it appeared.
- Window dressing turned up at period boundaries in files where it was not the funding event the merchant described as such.
- Industry concentration of risk: post-COVID restaurant cohorts and 2024-cohort owner-operator trucking.
Context
These observations are aggregate. No specific merchant, funder, ISO, or pilot partner is identifiable in this post. No threshold values, formula details, or per-segment multipliers are disclosed. Numbers are rounded to ranges where stated.
The corpus is the first 100 anonymized files Vyaso re-underwrote during pilot engagements. The files were a mix of declined and funded, submitted by funder partners and ISO partners, drawn from across nine industries. The pilot covered late 2025 through mid-2026.
Observation 1: the adjusted-revenue gap was bimodal
We expected the adjusted-revenue gap (the percentage reduction from gross deposits to adjusted revenue) to follow a roughly normal distribution centered in the low double digits. It did not.
The corpus was bimodal. A large cluster of files showed a small gap, in line with the expected center. A second cluster showed a much larger gap, in the high double digits. The two clusters were separated by a thin middle band.
This matches what an experienced underwriter expects qualitatively: most files are clean or close to it, and a smaller subset are materially inflated. Vyaso's value lands hardest on the second cluster, where the gap is large enough to change the funding decision.
Observation 2: account kiting was the most common flag
We expected loan stacking to be the most common flag fired by the engine. It was not. Account kiting fired more often.
The reason is structural. Multi-account merchants routinely move money between their own accounts: from operating to payroll, from sales to capex, from owner draw to operating. Many of these movements are legitimate. Some inflate apparent revenue by counting the same money twice across account statements that are submitted together.
The kiting layer surfaces both. Underwriters then triage the legitimate from the inflating. The aggregate signal: any multi-account submission deserves a kiting check before the deposit total is taken at face value.
Observation 3: loan stacking was severe when present
Loan stacking appeared in a smaller subset of the corpus. Where it appeared, the depth was significant. Files with any stacking flag tended to have multiple stacking flags. The pattern is: stacking is uncommon, but a stacker is rarely a one-off stacker.
The actionable read: when a stacking flag fires, the underwriter should expect to find more stacking signals in the same file. The first lender deposit identified is often not the only one.
Observation 4: window dressing concentrated at quarter ends
Window dressing (large deposits at the closing days of a statement period that withdraw shortly after the period starts) showed a calendar pattern in the corpus. It concentrated at quarter ends, not at month ends.
This is consistent with funder underwriting cycles that pull statements at quarter boundaries. The detection signal is sharper there because the manipulation incentive is sharper there.
Observation 5: where industry concentration of risk lay
Two industry pockets carried the heaviest concentration of flagged files in the corpus.
The first is the post-COVID expansion cohort of restaurants and food service. Operators that opened or expanded in 2021 and 2022 carry layered debt now. Stacking flags concentrated in this cohort.
The second is owner-operator trucking from 2024. The combination of fuel price volatility and tight freight rates created a pocket of operators carrying multiple facilities, often disguised as factor settlements.
Other industries showed flagged files but not in concentration. Healthcare, professional services, retail, and SaaS files in the corpus were broadly cleaner.
What we changed about how we look at files
The corpus changed our priors. Three priorities moved up the underwriter's first-pass checklist:
- Run the kiting check before reading the deposit total. Multi-account submissions distort the headline number in ways the eye does not catch.
- When a stacking flag fires, expect more. Read the file for additional disguised lender deposits before drafting the credit memo.
- At quarter end, weight window-dressing signals heavier than at month end.
These are not rules embedded in the engine. They are underwriter habits the corpus reinforced.
What is in the next 100 files
We are running the same retrospective analysis on the next 100 pilot files. The first read of the next batch suggests the kiting prevalence holds, the stacking pocket has moved somewhat, and a new pattern around AI-generated forged statements is starting to register on the PDF integrity layer. We will write that up when the sample is large enough to be honest about.
Stacking is uncommon, but a stacker is rarely a one-off stacker. When a stacking flag fires, the underwriter should expect to find more stacking signals in the same file.