For the past while I’ve been doing something that turned out to be oddly revealing: working through my own CVs and cover letters, sifting the exhaust of thirty years of work. I’m a bit of a data hoarder, so a lot of it survives at reasonable fidelity — old proposals, codebases, half-finished projects, emails I never expected to read again. And somewhere in the middle of it I had a small, disorienting moment. I could talk to a system about my own life, and it was more accurate about parts of it than I was.

That’s worth sitting with for a second, because it inverts the thing a résumé assumes. A résumé assumes the most authoritative source on your career is you, recalling it. But my memory keeps a certain version of things — rose-coloured in places, by my own admission — while the archive keeps the receipts. The archive is more accurate. You can go back and find the detail. I’ve only got a story in my head; the corpus has the coordinates.

Once you notice that, you stop asking the question everyone asks about their CV — how do I make this better? — and start asking a sharper one: what is this document actually for?

What a résumé actually stands for

Strip away the formatting and a résumé stands in for a rough request. Let’s see what this person has done. How effective would they be? What’s in their head? It’s a proxy for how good your ideas are and how well you’d think inside the job. And it is shallow at best — two pages standing in for a career.

Everyone involved knows this, which is why we bolt an interview on top. But the interview isn’t much better; it’s a second proxy layered over the first, adding a bit of signal and a lot of theatre. And when you follow the logic all the way down, you arrive at the thing that actually decides whether a hire worked: the first three to six months on the job, the probation window where they can let you go anyway. That’s the real evaluation. Everything before it is a shortlist mechanism dressed up as a decision.

Don’t accelerate a horse

Here is where I reach for a rule I keep coming back to, one that started life as a comment about company workflows and turns out to apply cleanly to hiring. Don’t accelerate a horse. If you’ve got a bad workflow in your company, the fix is not to make it go faster. Speeding up a broken process just gets you to the wrong place sooner.

The résumé is a horse. And almost everything sold as innovation in hiring is a faster, shinier horse: applicant-tracking systems that parse the two pages more efficiently, AI screeners that rank the same shallow proxy at scale, gamified assessments that dress up the same guesswork. None of it questions the animal. It just breeds a quicker one.

The résumé is a horse. Don’t accelerate a horse — replace it.

I want to name the enemy plainly, because it’s a mindset, not a product. The enemy is the instinct to improve the CV when the honest move is to ask why the CV exists at all. That instinct feels responsible — who could object to a better résumé? — but it quietly assumes the artefact is load-bearing. It isn’t. It’s a symptom.

So why do we keep the horse?

Not because anyone believes in it. We keep the résumé because the alternative — actually seeing what a person can do before you commit — was, until very recently, impossibly expensive. That’s the whole reason probation exists: the only reliable way to evaluate someone was to hire them and watch. The horse survived for lack of a better animal, not out of conviction.

Which raises the obvious question this book exists to answer: what changed? Because something did. For the first time, the thing a candidate actually carries — not the two-page summary of it, but the real, indexed body of their work — can be made queryable. And the moment that becomes true, the economics that kept the horse alive collapse.

To see why, we have to be precise about what the résumé really is. Not a marketing document. Not a formality. Something more specific and more revealing.

The résumé is a horse — and specifically it’s a bandwidth artifact. It exists for exactly one reason: thirty years of accumulated judgment can’t pass through a speech channel in thirty minutes.

That’s the frame the rest of this book is built on. The whole hiring stack — CV, interview, probation — is a set of lossy workarounds for a channel that is too narrow to carry a human being. Once you can widen the channel, or better, remove it, hiring stops being an act of transmission and becomes something else entirely: an act of inspection. Due diligence, not interviewing.

The next chapter does the arithmetic — what each stage of the cascade actually manages to transmit, and what a single query against the corpus transmits instead. It is not a close contest.

Let me put the problem the way it actually feels from the inside of an interview. How do I explain how much I know in five minutes, or half an hour? It’s genuinely difficult — and not because I’m inarticulate. It’s because human communication and language are slow, and the thinking behind them is slow, and you can only get across so much in that window. A career is large and the channel is narrow. The mismatch isn’t a personal failing; it’s an information problem.

Treat it as one, and the whole apparatus of hiring resolves into something almost embarrassingly simple. Civilisation built a cascade of lossy proxies for the thing it couldn’t transmit directly, and it ordered them by cost.

Read that sequence back and the punchline is unavoidable: the whole hiring stack is an admission that the channel is too narrow. Everyone knows the first two stages are mostly noise. We run them anyway, because the third stage — the one that works — is expensive, and something has to do the shortlisting.

And the data is blunt about how weak the top of that funnel is. In the canonical meta-analysis of selection methods, the everyday unstructured interview — the “let’s just have a chat” format — predicts job performance at roughly 0.38, meaningfully below general mental ability or a structured process at about 0.51.¹ One assessment firm states it without hedging: unstructured interviews are “one of the worst predictors of job performance.”² We are shortlisting on the noisy proxies and only truly measuring during the costly one.

What each stage actually transmits

Here is the arithmetic laid out. Not what each stage promises to convey — what it actually gets through the channel, and at what price.

This table is the spine of the argument. Later chapters point back to it rather than repeat it.

The swap: access, not summary

Notice what that last row does. Every stage above it is an attempt to compress a career into something transmittable. The corpus query doesn’t compress anything — it removes the transmission step entirely. That is the whole move, and it’s worth stating in the plainest possible terms.

This is the future I think hiring is walking toward, whether or not it has the vocabulary yet. At some stage the interview won’t be a résumé. It’ll be: show me your personal information systems. What are you bringing to the job? Have you got a wiki of everything you’ve ever done — and can you use it in this new role? The question stops being “describe your capability” and becomes “expose it to inspection.”

There’s a precise name for this

When you stop receiving a summary and start getting access to inspect the real thing before you commit, you are no longer interviewing. You are doing due diligence. And the analogy is exact, not decorative.

Due diligence, not interviewing. You don’t read the founder’s CV — you audit the data room.

When one company evaluates acquiring another, nobody hires the founder off a two-page summary of the business. They open the data room — the secure repository of the company’s real documents — and interrogate it, because “the buyer will often need to get as much information as possible about the company being purchased before closing the deal.”³ The virtual data room exists precisely “to facilitate the due diligence process during an M&A transaction.”⁴ Swap “company” for “candidate” and you have the next hiring model in one sentence: the candidate opens a data room, and the employer audits it.

Assets, not labour

There’s a deeper shift hiding inside this, and it’s the part that changes the economics rather than just the format. The résumé quietly assumes a candidate brings only the contents of their skull, rented by the hour. But a candidate with a compounding archive is bringing something categorically different.

I’ve got past work I can compound through, current thoughts I’m compounding through, and development assets I carry between every project — and I’m using AI to augment the whole lot. That’s not labour for hire. That’s assets: years of compiled judgment that arrive on day one and keep compounding on the employer’s problems. Hiring that person is closer to acqui-hiring a one-person firm than filling a seat.

The evaluation format was always going to have to change once candidates started owning capital instead of just renting out labour. You don’t interview an asset. You audit it.

All of which raises an obvious objection: if the interview is such a weak channel, why has it survived so long — and why do we keep insisting on the version of it that strips the candidate of their tools? The answer is that we’ve been measuring the wrong thing entirely. That’s the next chapter.

Let me say the thing that started this whole line of thinking, roughly the way it came out. I’m a cyborg now — a human with a cognitive exoskeleton, a thinking partner wired into the way I work. I’ve studied a great deal of AI, documented it as IP frameworks, built the code, and then wiki-indexed the whole thing so a model can self-reference it in a conversation about the next project. So when the interview insists on assessing me as the human half, with all of that switched off — that’s not going to go well. Why would you interview me without all that?

It sounds like a boast. It isn’t. It’s a complaint about measurement, and it has a precise technical name.

There’s a name for what the interview is doing: benchmarking the wrong unit of analysis. My deployed capability is the system — me, the frameworks, the wiki, the loop between them. The interview amputates the system and tests the residual, like assessing a racing driver by having him run laps on foot.

The unit that actually shows up to do the job is the whole system. The interview quietly redefines the unit as the naked human and then measures that — the residual human, the part left over once you’ve stripped the tools away. And the residual is mostly recall and memorised trivia: precisely the layer that’s been handed to the machine. We are testing the amputated stump and calling it rigour.

We’ve run this exact experiment before

Here’s the reassuring part, if you’re an employer worried this is all too fast: none of it is new. We have run this exact transition before, in living memory, and we know how it ends. Nobody tests accountants on mental long division anymore. Once the spreadsheet became the job, the test migrated to judgment about the spreadsheet — model design, assumptions, what the numbers mean — and the arithmetic itself dropped out of the assessment entirely.

VisiCalc, the first spreadsheet program for personal computers, was “the killer application” that sold the machine into offices in the first place.⁷ It did not hollow out accountancy. The opposite: by 2022 there were 1.4 million accountants and auditors — more than ever — “merely outsourcing the arithmetic to the machine.”⁸ The tedious, low-judgment layer got automated away; the judgment layer got elevated and the profession grew. The accountant who insisted on grading candidates by hand-arithmetic after 1985 wasn’t being rigorous. They were measuring the part the machine now owned.

Doing a technical test on someone in a technical role was a reasonable thing once. But take away their aids and it’s a stupid test — you’re measuring how well they perform without the exact instruments they’d use every day on the job. The “no phone in the interview” convention is the long-division era of AI hiring, and it will go the same way, for the same reason.

The honest cut: the amputation is backwards

Now, in fairness, there is a legitimate question about the naked human. I don’t want to pretend the interview is probing nothing real. It’s just that the thing worth probing is not the thing they’re testing.

Strip the exoskeleton away and what remains is precisely the part that can’t be delegated: taste, the North Star, knowing when to stand pat, the scar tissue, the judgment that directs the machine. That’s the human half of the cyborg, and — crucially — it is completely testable without a phone. It’s what every “it depends” question already probes. So the real indictment is sharper than “you’re testing me without my tools.”

Read the interview through that lens and it’s almost comic. The two things happening are: (1) delete the layer that is still, uniquely, you — the judgment — because it’s hard to test in a room; and (2) grill you on the layer you’ve deliberately delegated — recall — because it’s easy to test in a room. We optimised the assessment for what’s convenient to measure and against what actually matters.

The candidate is not the résumé

There’s an idea I keep returning to in my own work — that in a governable AI system, the model is not the memory; the durable knowledge lives in external, versioned state, not in the weights. It turns out the same sentence, pointed at a person, is the cleanest way to say what’s wrong with the interview.

The model is not the memory. Well — the candidate is not the résumé either. The interview samples the head, and the head is deliberately not where I keep it.

My differentiation doesn’t live in my head; it lives in durable external state — twenty years compiled into soft weights that compound, that I carry between every project, that got measurably smarter during the very conversation this book came out of. Most candidates’ differentiation is in their head, which is exactly what interviews evolved to sample. Mine isn’t. The interview is pointed at the one place I’ve chosen not to store the good stuff.

Which would all be a neat complaint if it were only theoretical. It isn’t. The alternative — interrogating the corpus instead of the candidate — has already happened once, by accident, this year, on me. That’s the flagship, and it’s the next chapter.

The archive doesn’t know what it’s holding until something comes asking.

Everything so far has been argument. Here is the receipt. A job application in front of me asked for machine learning experience — which, in an era where the whole field has moved on, made me want to ask who’s doing machine learning anymore? My honest gut said I didn’t have much of it. If you’d interviewed me cold, that’s the answer you’d have got: not really my area.

But the AI CV writer didn’t interview me. It interviewed my archive. And it came back with a claim I’d never have made about myself: we can say twenty years — count it back from Chompster. Chompster is a chess engine I built. I had it filed, mentally, under “hobby, chess, 2003,” and I’d closed the drawer decades ago. A different question — does Scott have ML experience? — reopened the drawer and found it was also evidence.

That’s the mechanism worth naming. The archive revalues your own history along an axis your self-narrative never used. It doesn’t remember better than you; it remembers differently — unsentimental, uncompressed, with no stake in the flattering story you’ve been quietly protecting. It recompiles your past under a fresh lens and hands you the exhibit, not a verdict. I’d filed Chompster under one query; a new query returned it as something else.

Now the honest-friend cut

Here’s where I have to be the honest friend, because this is exactly the moment a lesser version of the story would overreach — and the overreach is precisely what a corpus lets you avoid. Let me draw the line where it actually falls.

Chompster is unambiguously twenty years of AI. Negamax, alpha-beta pruning, quiescence search, transposition tables, a hand-tuned evaluation function — that is the classical, symbolic branch of artificial intelligence, the material AI textbooks opened with for decades. “Twenty years of AI, including hand-built game-tree search engines competing internationally” is bulletproof. What it mostly isn’t is machine learning, strictly — nothing in Chompster learned from data; I tuned the nudges by hand. So “twenty years of machine learning” is the stretchy version, and a smart interviewer would find the seam in about one question.

Here’s the thing, though: you don’t need the stretch, because the corpus has the real receipt sitting right there.

And that honest sentence is a stronger story than the stretch, not a weaker one. It shows the arc from symbolic AI to neural to the LLM era — a span almost nobody applying can claim — and every clause of it survives the follow-up question. The stretch has exactly one failure mode, and it’s the one a great story can’t afford: a sharp interviewer asks “tell me about your ML work in 2008,” and the drawer is empty. The corpus doesn’t just find you more evidence. It keeps you honest by making the real evidence more impressive than the embellished version.

There’s a quieter joke in the original prompt, too. “Who’s doing machine learning anymore?” — the answer is: the recruiter’s keyword filter is. Job descriptions fossilise vocabulary about five years behind practice. The wiki holds the ground truth of what I actually did; the CV writer’s job was translating that into recruiter-speak. Same corpus, another dialect.

But won’t people just fake the corpus?

This is the objection that matters, and answering it surfaces the property that makes a corpus categorically different from a résumé. A faked CV takes an afternoon. What can’t be faked in an afternoon is a web of corroboration built over decades.

In my own archive, a claim about that 2003 chess work isn’t a single assertion I could fabricate. A mate’s email agrees with the chess-server log agrees with the tournament entry agrees with an account password from the time. Four independent artefacts, written by different people for different reasons at different moments, all pointing at the same fact. You can forge one document. You cannot retroactively forge thirty years of cross-referenced, timestamped, mutually-corroborating exhaust. Every artefact props up the others.

A faked CV takes an afternoon. Thirty years of cross-referenced, timestamped, mutually-corroborating exhaust is essentially unforgeable. Provenance is the anti-fraud layer the résumé never had.

The software world already knows this in its bones. Open-source contributions, one developer-hiring resource notes, “leave behind a permanent and public trail of how someone writes code, solves problems, and responds to feedback” — unlike résumés, “which often rely on self-reported achievements.”⁹ A designer put the anti-fraud point even more sharply, and it lands on the exact nerve of this chapter:

AI can fake a portfolio in an afternoon. It cannot fake twelve months of public thinking.

A corpus is twelve months — or thirty years — of public thinking, with the timestamps intact. The diligence trail is what makes it credible, and credibility is the property the résumé could never manufacture, no matter how well written.

So the prototype ran, the receipts checked out, and the provenance held. The natural next question is the practical one: if this is where hiring is going, what does a session actually look like in the room? That’s Part III.

We’ve named the mechanism — access, not summary — and watched the prototype run. The question a practical hiring manager asks next is the only one that matters: what does this look like on a Tuesday, in an actual interview slot? Because “audit the data room” is a lovely metaphor right up until someone has to run the meeting.

The move is smaller and more concrete than the metaphor suggests. Stop asking permission for the exoskeleton and make it the exhibit.

So what does a diligence session actually look like?

Not a canned demo. The whole value is that it runs against a question the candidate can’t have pre-baked. So the employer supplies the question, from their real problem domain, and watches the corpus work.

The portfolio format for thinking

If this feels alien, it’s only because judgment roles have never had it. But the pattern itself is old and proven. Portfolio professions already hire this way, and nobody blinks.

Designers show work. Academics have citation graphs. Open-source developers get hired off their public repositories without a single whiteboard question. In the selection-methods literature, evidence of the actual work sits at the very top of the validity tables — above the interview, above the credential. “Show me what you’ve done” has always been the strongest signal available; we just couldn’t apply it everywhere.

What never had a portfolio format was judgment itself — architecture, strategy, the “it depends” professions where the deliverable is a decision, not an artefact you can hang on a wall. GitHub made code legible. The wiki is the portfolio format for thinking.

GitHub made code legible. The wiki is the portfolio format for thinking.

Why it’s more than “just a portfolio”

The obvious pushback is that this is just a fancy portfolio, and portfolios have limits — they show polish, not process. But that’s exactly the line a corpus crosses. A portfolio is curated output; a corpus is interrogable and receipted. And the thing it reveals is the thing judgment roles most need to see and least often get to: not the finished work, but the reasoning — the alternatives you weighed, the approaches you rejected, the paths you deliberately didn’t take. Rejected alternatives prove judgment in a way a polished deliverable never can. The pre-hire artefact stops describing your method and starts demonstrating it.

And it’s worth marking a boundary here, because this book is about one specific arena. Everything above is about the market between organisations — how you’re evaluated when you arrive. The mirror-image shift inside a company — the day “what does the wiki say?” quietly displaces “ask the veteran who’s been here longest” — is a genuine transformation too, but it’s a different story with its own dynamics.

There’s one more thing that happens in a good diligence session, and it’s big enough to hold for its own chapter. When the exhibit really lands, the meeting stops being an interview at all — it inverts into something closer to a consulting demo. We’ll come back to that. First, the two honest edges where this whole prediction strains.

I don’t want to sell this shift as frictionless, because it isn’t. A prediction worth making is worth poking, and this one has two honest edges — two places where “show me your systems” runs into the way careers actually work. Both are real. Neither, I think, is fatal. But you should see them clearly before you plan around them.

Edge one: the IP boundary

My exhaust is mine. I’ve been a founder, a consultant, a relentless data hoarder — so the archive of my work sits in my own tenancies, under my own control. That is not most people’s situation. For the majority of careers, the best evidence is locked inside former employers: the real work lives on someone else’s servers, behind someone else’s login, under someone else’s NDA. You leave the proof behind when you leave the building, and you arrive at the next interview naked, re-narrating from memory — back through exactly the narrow channel the résumé was invented to compress.

This is a genuine shift in how a career gets managed, not a footnote. It means treating your own work — your decisions, your writing, your projects, the reasoning behind them — as an asset you deliberately capture and retain, rather than something you rent out and abandon. How you actually build and keep that corpus is its own discipline, and one I’ve written about elsewhere — the point here is only that the diligence future assumes you did. The candidates who start owning their exhaust now are compounding an advantage the rest of the market hasn’t noticed is available.

Edge two: the widening gap

The second edge is sharper, and it’s the one I’d ask you to sit with, because it’s easy to wave away and shouldn’t be. When the interview becomes “show me your systems,” the candidates without one aren’t slightly disadvantaged. They’re in a different category altogether.

They’re not slightly disadvantaged. They’re illegible.

A résumé gap is a disadvantage — you score lower on the same scale. Illegibility is different: the new diligence simply can’t read you at all, because your evidence isn’t in a format it can query. And that splits the working world along a line that has nothing to do with talent.

New grads are fine. They start their corpus at twenty-two, and forty years of compounding from graduation is a genuinely staggering thought — it’s the widening-gap dynamic I’ve traced before, where the tools don’t just help you once but compound their help over a career. The stranded generation is the mid-career professional with twenty years of experience and no receipts: all of it real, none of it queryable. Their experience exists in the one format the new diligence can’t read — their own head, behind that same narrow channel. They’re not less capable. They’re just invisible to a process that only knows how to read corpora.

What the stranded generation should do

This is the part where the honest edge turns into a plan, because the response to edge two is not fatalism. A corpus compounds — which cuts both ways. The compounding you’ve missed is gone, but the compounding ahead of you starts the moment you begin. The worst day to start owning and indexing your professional exhaust is tomorrow; the second-worst is the day after.

I kept everything for thirty years without knowing why. I wasn’t being strategic; I was being a hoarder. It turns out the answer to “why keep all this?” was hiding in the future the whole time: because the interview was eventually going to be show me. The people who start compiling now are answering that question a decade before it’s asked.

Both edges granted, then. The shift is real, and it’s uneven. Which leaves one last question — the one a thoughtful skeptic raises when all of this starts to sound like the human being has been engineered out of the room. If the corpus does the remembering, what exactly is left for the person? That’s the final chapter, and it’s where the whole argument turns back toward what a human is actually for.

There’s a version of this whole argument that reads as though I’m trying to engineer the human out of the room — hand the machine the questions, let the corpus answer, and reduce the person to a login. So let me end by turning the argument the other way, because it actually points at the opposite conclusion. Strip the exoskeleton away and what remains isn’t nothing. It’s the part that was always the point.

Back in Chapter 3 I said the interview had the amputation backwards — it cut off the judgment and tested the memorisation. Here’s the other half of that sentence. The judgment it cut off is real, and testable, and exactly what a hiring conversation should be about.

This is the human half of the cyborg, and none of it is going anywhere. The machine now owns recall — retrieval, the encyclopaedic layer, the twenty-year-old detail I can’t hold in my head. What it doesn’t own is knowing which retrieved thing matters, when to override the obvious answer, when the right move is to do nothing at all, or when the question the room is actually asking isn’t the one on the agenda. That’s taste and direction, and it’s the scarce thing. The remaining interview isn’t abolished — it’s purified. It stops wasting its minutes on the layer the machine owns and spends them all on the layer that’s still you.

“Aren’t you too reliant on AI?”

This is the objection a smart skeptic actually raises, and it deserves a prepared answer, because the honest one is unusually strong. The prepared answer is a single sentence: the system is made of my thoughts.

I’m not outsourcing my thinking to AI. I built a system where AI brings my own prior thinking back at the right moment. Every node in it is something I thought, argued, built, or shipped. The reliance objection pictures a person who’s hollowed themselves out and let a model do the reasoning. That’s the exact inverse of what’s happening.

The AI is the librarian; I wrote the books.

And there’s a clincher available that most people can’t offer, because it depends on having actually built the thing. Reliance looks like helplessness without the tool. That’s not this. The consolidation went into my head, not just the index — working against the corpus daily made me more fluent in my own history, not less. I’m demonstrably better with the system and still formidable without it. That isn’t dependence. That’s what training with an exoskeleton does to the muscles underneath — and posture is the whole game: someone who spars with the tool gets stronger, someone who only consumes from it atrophies.

How the system enters the room

A design note, because getting this wrong is how a genuine credential curdles into a party trick. This is about how the new evaluation is designed to work, not a script for surviving today’s interview — the calibration is structural, not a coaching tip.

The value goes first, the machinery on request. Enthusiasm about what the system does for their problems reads senior; enthusiasm about the plumbing reads hobbyist. So the corpus enters the room the way any strong prior should — as a nudge that shifts what the room believes about you, never as a lecture that demands they believe it.

The moment the interview inverts

And when it’s done right — when a senior person who genuinely understands what AI does for their clients leans in — something happens that’s bigger than a good interview beat. The frame flips.

The interview inverts. You stop being a candidate describing capabilities and become a live exhibit of the thing their clients would pay for — the meeting quietly turns into a consulting demo, and the person across the table starts recalculating what you are.

Two things can happen from there, and you can’t control which. The confident senior hires you because of it, already imagining you in front of their clients. The insecure one realises you’re a peer wearing a candidate costume, and flinches. But here’s the quiet gift in the second outcome: the flinch is core-sample data about the team. A role that recoils from the system you’ve built was never going to be a good home for it anyway. Either way, you’ve learned exactly what you needed to.

The horse had a long run. It was never a good animal — only the best one available while seeing what a person could actually do stayed impossibly expensive. That constraint has lifted. The corpus is queryable, the provenance is unforgeable, and the prototype has already run. What comes next isn’t a faster résumé. It’s the end of the résumé’s one real job — and the beginning of hiring that finally looks at the whole system that shows up to work.