Verification has a cost. So does uncertainty.

The first question every employer, investor and regulator asks about a worker record is whether it is verified. It is asked as a yes or no, in the same tone one might ask whether a document is signed.

There is no honest yes available, and the no is misleading.

A record of a housekeeper's capability might contain a practical assessment scored by an accredited centre, an attestation from a training partner, a work history she stated herself, and a language claim nobody has examined. Four claims, four completely different epistemic statuses. Answering "yes, it is verified" flattens them into one and asserts something false about at least two. Answering "no" discards the assessment that cost real money to produce.

The question feels well formed because of an analogy that does not hold. Identity documents genuinely are binary: the passport is authentic or forged. Capability is not like that, because there is no single issuing authority and no single fact of the matter. A worker can cook and not clean. Verified is not a property the record has. It is a property individual claims have, in degrees, at a cost.

Which turns the design question from how do we verify everything into something much more tractable.

The formal treatment of this is nearly fifty years old and sits in a literature most people building worker records have never read.

Townsend's 1979 model of costly state verification sets up exactly this problem. One party knows the true state. The other can learn it only by paying a fixed fee. Townsend's result is that the optimal contract does not verify always, and does not verify never. It has the informed party report, and triggers verification only in the states where the report would otherwise be unreliable.¹ Gale and Hellwig developed the argument further, showing the verification-cost-minimising contract has a standard debt form, and that investment stays strictly below the first best precisely because verification is costly.²

Williamson's transaction cost work reaches the same shape from a different direction: reducing the cost of safeguards raises exposure to opportunism, and raising safeguards costs money, so the question is always where to sit on that trade rather than how to eliminate it.³

So the sharper definition is this. Verification is not a state a record achieves. It is a budget allocated across claims, and the allocation is the design. Any system claiming a fully verified record has either spent more than the decision was worth or is describing something narrower than it sounds.

If verification should be selective, the practical question is where. There is a natural experiment that answers it with unusual clarity, and it comes from tax rather than hiring.

Kleven and colleagues ran a randomised audit study covering roughly 40,000 Danish tax filers. Income reported by third parties, employers and banks, showed an evasion rate of 0.3 per cent. Income that filers reported themselves showed 37 per cent.⁴

That is a ratio of about 123 to 1, and it is the single most useful number I have found for designing a verification budget. Misreporting is not spread evenly across a record. It concentrates, almost entirely, in the channel where the subject reports on themselves with nobody else attesting.

Hiring data points the same way with less precision. Across three samples, Henle, Dineen and Duffy found 72 per cent of respondents admitting to embellishing a CV, 61 per cent to omitting, and 31 per cent to fabricating at least to some extent.⁵ Those are anonymous self-reports against a low threshold, so the levels should be read cautiously. The structure is what matters: misrepresentation clusters where nobody else is attesting.

Translate that onto a worker record and the allocation writes itself. A practical assessment scored by an accredited centre is third-party reported. A partner attestation is third-party reported. Work history stated by the worker, with no contactable employer, is self-reported. That last category is where almost all of the risk sits and where almost none of the verification spend currently goes.

Concentration tells you where the risk is. It does not by itself tell you what to check, because a claim can be unreliable and irrelevant at the same time.

Decision theory supplies the missing half. Howard's formalisation of information value establishes that information is worth precisely what it changes about the decision taken.⁶ A perfectly rigorous check on a claim that cannot alter the outcome has a value of zero, and it costs whatever it costs.

So the allocation rule has two terms rather than one. Verify a claim when it is both load-bearing for the decision at hand and self-asserted rather than third-party attested. Claims that are load-bearing and already attested need no further spend. Claims that are self-asserted and decision-irrelevant should be labelled untested and left alone.

The load-bearing half is brief-specific, which is what makes a standing universal verification policy incoherent. For a household hiring for an elderly parent, the elderly care claim is load-bearing and the ironing claim is not. For a household with three children under six, it is the reverse. A record that verifies identically for both has spent the same money to answer two different questions, and answered neither of them well.

There is a further reason to resist the urge to verify everything, which is that nobody has demonstrated it works. A review of the background-check literature found the legal-liability guidance well developed and empirical evidence linking checks to reduced turnover, theft or violence essentially absent.⁷ That is not proof that checking is useless. It is a warning that the industry has been selling an outcome it has not measured.

The serious objection is that selective verification is an announcement of where not to look. If workers, agencies and training providers learn that self-asserted language claims are checked and self-asserted work history is not, the falsification simply migrates. On this account partial verification is worse than none, because it lends the whole record a credibility it has not earned.

It is a real risk and I have seen a weaker version of it in practice. But the Danish study speaks to it directly, and cuts the other way.

Alongside the audits, the researchers sent randomised letters announcing the probability that a filer would be audited. Compliance rose in proportion to the stated probability.⁴ Deterrence did not require certainty. A known, non-trivial chance of being checked moved behaviour, which is what an insurance model would predict and what the gaming objection implicitly denies.

That suggests the right design is not verify everything but verify a declared random sample in addition to the targeted claims, and publish the sampling rate. The randomness restores deterrence across the claims that targeting would otherwise leave alone, at a small fraction of exhaustive cost.

Two honest limits on that. Tax filers are a different population with different incentives from workers seeking a placement, and a deterrence result in one does not transfer automatically to the other. And the whole argument assumes the verification is competently done; a sampled check that is careless is worse than none, because it produces the credibility without the substance.

One thing stopped this paper from going further, and it is worth reporting as a finding rather than hiding as a limitation.

I could not locate credible independent cost data for verification in hiring. Not for reference checks, not for credential verification, not for practical skills assessment. What exists is vendor pricing pages and screening-industry marketing, which state prices rather than costs and have an obvious interest in the answer. Following the standard this series has tried to hold, none of it appears here.

That absence matters more than it might seem. The allocation rule in section four requires two inputs: the risk concentration, which the Danish study supplies, and the cost per verification type, which nobody publishes. Without the second, everyone building in this space is optimising against a number they have privately estimated and cannot check against anyone else's.

It is also a straightforward thing to fix. Any operator running assessments at volume knows their own cost per unit by method. Publishing those figures, even roughly, would let the field reason about allocation properly rather than each participant guessing in isolation. I intend to publish ours once the volumes are large enough for the numbers to mean anything, which is a commitment with a date I have not yet set and should be read accordingly.

Three things follow for anyone building or buying one of these records.

Stop answering whether the record is verified. Answer which claims are in which state, and show the date beside each. That is more information than a flag, it is honest, and it takes no more space on a screen.

Allocate the verification budget by the two-term rule: load-bearing for this brief, and self-asserted rather than attested. Then add a published random sample across the remainder, because the deterrence evidence says a known probability works and certainty is not required.

And treat any counterparty who answers the verification question with a yes as having told you something useful about their record, which is that they have not thought about it. A record with no untested claims on it is not a rigorous record. It is a record that has quietly relabelled its untested claims.

Uncertainty is not the absence of a cost. It is a cost that has been moved onto whoever makes the decision, usually the household, usually without being told. The work is not to eliminate it. The work is to price it, place it where it does least damage, and say out loud where it still sits.