The answer we are most proud of is “no”

Give a valuation system a receivable with six comparable settlements and it will produce a price. It will look exactly like the price it produces from two hundred comparables — same format, same confidence, same number of decimal places. Nothing about the output tells you that one of those two is an artefact of which six matters happened to close.

The median of six is not a property of the population. It is a property of the sample, and with a sample that small the two are unrelated in a way that no amount of presentation can repair. A system that prices it anyway has not made an estimate; it has made something up in a format that resembles an estimate.

So EliAI refuses. Where the retrieved cohort is too thin, the mark is recorded as indicative, carries a range and no midpoint, recommends no advance, and is excluded from book value rather than assumed to zero — because assuming zero is also a made-up number, just a conservative-sounding one. The book reports the exclusion beside the total instead of quietly shortening it.

The important part is that this is not a policy anyone remembers to follow. The database refuses to store a priced mark without a band. The refusal is in the schema, which means it survives a rushed quarter, a new hire, and a product manager who wants the dashboard to look complete. A control that depends on discipline is not a control.

There is a version of this argument that sounds like humility and is really marketing. The test of whether a vendor means it is whether the refusal costs them anything. This one does: it means a demo where a position comes back unpriced, and a book where some rows are blank. We show that on the worked valuation deliberately, next to the priced case.