The Quantity Problem — HyperC Engineering Note

Quantity is a required input

P34 currently requires an explicit concept of volume, or quantity, per deal. A menu line is not just this opportunity — it is this opportunity, at this size, and the model's answer is a size as much as it is a yes or a no.

That fits a great many markets. A wholesale case pack has a unit count. A truckload has weight and lane capacity. A GPU block has hours. A loan has a principal. In all of those, "how much" is a real dimension of the decision, and the model can regularize along it.

On a market where every deal is one indivisible thing, that dimension collapses — and a load-bearing part of the model collapses with it.

A single used machine. One domain name. One storage unit at auction. One collectible. You either take it or you do not; there is no such thing as 40% of it. On these markets P34 still works, but it works with one hand tied.

Sizing falls back to portfolio-level risk

With no quantity to size, deal sizing no longer travels the regularized path. Selection falls back purely to portfolio-selection risk-adjusted return calibration — the portfolio is still assembled and still risk-calibrated, but the per-deal sizing signal that would normally shape it is absent.

The second consequence is the one worth stating plainly, because it is a capability that disappears rather than a number that gets worse:

All adversarial market action on the individual deal is lost.

Sizing is how the model acts against an individual counterparty or listing — how it presses, holds back or refuses at the level of the single deal. Remove the size dimension and that behaviour has nowhere to live. What remains is a well-calibrated portfolio built from all-or-nothing picks.

So the honest summary is: degraded, not broken. Members operating quantity-free markets should expect the model to select competently and to size not at all.

Three approaches, none of them free

We are working on solving this. Three approaches are currently on the table, and each one buys something by giving something up.

Option 1

Leave as-is

Set qty=1 and rely on portfolio-level risk. The model keeps its calibration and its predicted portfolio economics stay honest; it simply does not size, and per-deal adversarial action stays off the table.

Option 2

Artificial partial fills

Allow impossible partial quantity fills and use the predicted sizing as an additional greedy sort rule for final selection. Sizing information comes back — as a ranking signal, not as a quantity to execute.

Option 3

Pre-clustered identity lines

Cluster the dataset so that deals group into virtual continuous identity lines, then send those lines — with sizings — as P34 menu line items. The size dimension is reconstructed at the level of the group rather than the item.

Where each one hurts

Option 2 loses the portfolio risk calibration, and it makes the model under-predict portfolio profit. There is a second, less obvious cost: inventing a partial fill for a unit-sized item means modelling how part of that item would be sold, and that model is itself risky. A half of a used excavator is not half the price and half the buyer pool; it is a fiction the rest of the pipeline would then treat as data.

Option 3 moves the whole difficulty into the clustering objective. The target function has to properly describe how an item behaves across all significant business dimensions — not just the obvious ones. Get that wrong and the virtual line is a group of things that look alike and behave differently, which is worse than no grouping at all, because the sizing it produces will be confidently incorrect.

Approach	What it recovers	What it costs
1 · Leave as-is	Nothing — the current behaviour. Calibration and predicted economics stay intact.	No sizing; no adversarial action on the individual deal.
2 · Artificial partial fills	Sizing information, used as a greedy sort rule over the final selection.	Portfolio risk calibration; under-predicted portfolio profit; a risky model of how part of an indivisible item would sell.
3 · Pre-clustered identity lines	A genuine continuous size dimension, at the level of the cluster.	Correctness now depends on the clustering target function describing item behaviour across every significant business dimension.

If your market has no quantity

You can still point P34 at it, and the catalogue lists plenty of markets in this shape. Expect competent selection and no sizing, and read the predicted portfolio economics as the model's own — not as a claim about a per-deal position it never took.

If your market does carry a natural quantity — units, hours, weight, principal, capacity — supply it. It is not an optional field that improves precision at the margin; it is the input the sizing path runs on.

We will publish where this lands. If you are operating a quantity-free market and have a view on which of the three is right, tell us — that is exactly the kind of thing the membership community is for.

P34 needs a quantity. Some markets do not have one.