Does this company have proprietary technology, or a proprietary evaluation?
The cheapest thing to improve is not the system but the measurement. A metric a company controls can move without the product moving — and the same fact makes a good private evaluation one of the more defensible things a team can own.
A company reports a strong number on a benchmark you have not heard of. Before anything else, it is worth being clear about what you have just been told: that a measurement they control returned a good value for a system they built.
That is not an accusation. Private evaluations exist for good reasons — public benchmarks are often a poor match for a real product, and a company that built its own is frequently the one taking evaluation most seriously. But the arrangement moves the burden. With a standard benchmark, the check is external and you can compare. With a private one, the check is internal, and what you are really assessing is the team’s judgement about how to measure their own work.
So the question is not whether the evaluation is proprietary. It is whether the team can tell you how their own measurement could mislead them.
Metrics move for reasons that have nothing to do with the product
The field my group works in had a version of this problem baked into its standard metric.
Explanations of what a unit inside a network responds to are scored by how well the explanation’s coverage overlaps with the unit’s behaviour. A reasonable measure. But the explanations are logical formulas, and the score can be inflated by lengthening the formula: adding terms broadens what the explanation covers and the number rises, without the explanation becoming a proportionally better account of anything. A metric everyone in the area was optimising rewarded something adjacent to what it was meant to capture.
Our response was not to propose a better single metric, because we did not believe one existed. We proposed a set of criteria to be read together, on the argument that each of them has a blind spot in isolation and only the combination gives a usable picture. That is a less satisfying answer than a single headline number, and it is the honest one.
I bring this up because that failure was in a public metric, scrutinised by a research community, with everyone’s incentives roughly aligned. A metric that is private, defined by the party being measured, and used in fundraising material is subject to the same failure with less external correction available by default.
The three cases, and what each one is worth
A standard benchmark, well chosen. Comparable and externally checkable. The risk here is different: the result may be real and irrelevant, because the benchmark measures something adjacent to what customers pay for. Comparability is not the same as significance.
A private evaluation. Potentially far more relevant, since it can be built around the actual task, and unverifiable from outside. Its worth depends almost entirely on how it was constructed and how honestly it is reported.
Neither. More common than you would expect, dressed as a demo. That case is covered by the space around the demo point.
What makes a private evaluation credible
The question I use is simple: how could this evaluation make you look good without your product being better? A team that has thought about it can usually name a specific weakness they have decided to live with. A team that has not treats the metric as a fact about the world rather than a construction they made.
Beyond that, a few things separate an evaluation that is doing work from one that is producing numbers.
Was it fixed before the results came in? An evaluation defined after the system’s behaviour is known has already been shaped by it, whether or not anyone intended that.
Does it ever disagree with the public alternative? If it never meaningfully diverges from the public one, I want to know what additional information it is supposed to add; if it always diverges, the divergence is the thing to explain. Either way I want to see both numbers, not a choice between them.
Is it one number or several? A single figure is easy to move. Several figures with genuinely different blind spots are harder to game simultaneously — though a suite whose measures are correlated, or designed together without that in mind, buys less than it appears to.
Who can run it? An evaluation only the team can execute is not an independent check. If a customer, an auditor, or a diligence process can run it and get the same answer, it becomes something closer to evidence.
The part that inverts the argument
Everything above reads as caution, so it is worth stating the other half plainly: a well-built private evaluation can be one of the more defensible assets a company owns.
Evaluation infrastructure is expensive, unglamorous, and often harder to reproduce than the model itself. A competitor can read your paper and reproduce your method — but if your advantage lies in knowing, faster and more reliably than anyone else, whether a change made the product better, that compounds. It is also often less portable than the model itself, because it lives in accumulated test cases, infrastructure, failure taxonomies and operational know-how rather than in a file. This is the same shape as the moat in the benchmark note: the defensible part was not necessarily the model.
Which gives the cleanest version of the question. If a company’s evaluation is proprietary, that can be a strength or a liability, and the difference is observable: is it something they can show you, hand you, and tell you the weaknesses of — or is it something you are asked to accept?
The first is infrastructure. The second is a number.
This is the kind of question I work through in technical due diligence — what the measurement establishes, what it can’t, and what would have to be true for it to be evidence.
Evidence
- NeurIPS 2023 senior author