Proof of work in hiring: what it proves, and what it doesn't

Hiring is shifting from what people claim to what people have made, and that shift is correct. A repository, a deployed product, a campaign that actually ran: these are harder to invent than a line on a resume, and any process that weights them above keywords will produce a better shortlist than one that doesn't. This is not an argument against proof of work. It is an argument about where proof of work stops.

What a portfolio link actually establishes

A link establishes that a thing exists, and that someone put it in front of you. That is genuinely more than a resume does. But read the claim carefully, because it is narrower than it appears: the artifact is the evidence, and the artifact says nothing about who made it. The connection between the person and the work, the only part a hiring decision actually rests on, is asserted by the candidate and corroborated by nobody.

Call it the attribution gap. It is not a rare edge case or a bad-actor problem. It is the default condition of every portfolio on the internet, including the honest ones.

Four ways proof of work gets gamed, none of them exotic

• A forked repository. The commit history is real, the code is real, and none of it is theirs. A fork renders as a link like any other.
• Team work presented as solo work. A product genuinely shipped by six people, described in the first person singular. Nothing in the artifact contradicts it.
• Commits authored by someone else. The git author field is a self-reported string; setting it to your own name is one command, and no reviewer checks.
• A real contribution with overstated scope. The candidate did touch the project. They did not architect it. Looking at the deployed thing cannot tell you which.

Only the first of those requires any intent. The last two happen constantly by accident, through ordinary self-flattering summary, which is the point. All four survive a portfolio review intact, because a portfolio review examines the work, and the work was never the thing in question.

Manual review checks something other than what you think

The usual answer to this is a human step: someone reviews every profile before it reaches employers. That is worth having, and it is not verification. Look at what such a review can actually determine from a profile: is it complete, is it coherent, is it plausible, does it match the roles on offer. Every one of those is a question about the profile. None of them is a question about whether this person did this work, because that question cannot be answered by looking harder at the work.

A careful human reviewing a link is a careful human looking at the same link you would have. The gap does not close because someone senior looked through it.

A link is an assertion. A check is a result.

That is the whole distinction, and it is worth stating flatly: proof of work is a link; verified evidence is a check. A link points at something. A check is something that happened: a test that could have failed and didn't, a named person who put their reputation behind one specific statement, a record confirmed against a source with no stake in the answer. The difference is not one of degree. It is a difference of category, and no amount of curation converts one into the other.

What checking actually requires

Once you accept that artifacts alone cannot close the attribution gap, the design follows. Evidence has to be scored by how it was obtained rather than by how impressive it looks, and no single route can be allowed to reach full confidence on its own.

• Score by modality, by how a skill came to be known. Demonstrated live under fresh questions is not the same species of evidence as a self-reported outcome, and any model that averages them has thrown away the distinction it most needed to keep.
• Cap each route independently. On insiderOne a produced artifact tops out at 70 of 100 confidence by itself, precisely because of the attribution gap this post is about. Attestation from other people caps at 60: ten vouches from ten friends structurally cannot reach expert, because piling up more evidence of one kind never repairs the weakness of that kind. A self-reported outcome caps at 45.
• Discount repetition from the same source. Three signals from one manager are one signal wearing three hats, and a model that counts them as three is a model that can be farmed by anyone who notices.
• Reward corroboration across routes. A skill that appears as a live demonstration and a real artifact and a named attestation is credible in a way none of them is alone, and the number should move because of the agreement, not because of the volume.
• Recompute it. A standing earned two years ago and never revisited has quietly turned back into a claim.

Only live demonstration escapes the ceiling, and only because passing it is expensive in the one currency that cannot be borrowed: you have to actually be able to do the thing, in real time, on questions nobody has seen before.

If you're hiring

Keep asking for proof of work. It is the right filter and it will improve your shortlist this week. Then ask the second question the portfolio cannot answer: what did this person specifically do, and who other than them will say so by name. If no step in your process could return a different answer than the candidate gave you, then you have not verified anything, you have collected claims in a better format.

If you're job hunting

Build the portfolio. It remains the highest-return thing you can do, and everything above is a reason to make yours checkable rather than a reason to skip it. Then close the gap yourself, because most people won't: state your exact contribution instead of letting the reader assume the flattering version, get one person to attest to one specific skill by name, and verify the single skill your target role screens hardest for. You are competing against a stack of links. Being the one that can be checked is a bigger advantage than being the one with the most projects.

The industry is right that resumes stopped working. It is early in noticing that portfolios inherit most of the same problem in better-looking packaging. Proof of work is where evaluation should start. It is not where trust can end.