You've watched a junior repeat a mistake you flagged in a meeting eighteen months ago. The fix lived in your head, then died in a Slack thread. Now it costs the firm $40K to relearn it.
That's the price of knowledge debt — the gap between what you know and what your organization can retrieve. For a senior expert billing $400/hr, every hour spent re-explaining judgment you've already given is a direct revenue leak. Multiply across a team and you're looking at six figures of recurring loss before anyone calls it a problem.
What would it be worth if your hardest-won judgment — the patterns you can't quite articulate in a meeting — became a structured asset you could query, license, or hand to an agent? Not a doc. An asset.
I'm Matt Cretzman, and that's the problem I built Skill Refinery to solve. Which is why the new wave of "GitHub for knowledge work" tools — Mesa being the loudest right now — is getting the metaphor exactly wrong.
Version Control Assumes the Artifact Already Exists
GitHub works because code is already structured. It's text. It compiles. It has a defined shape before anyone commits it. Version control sits on top of an artifact that exists.
Knowledge work isn't like that. The artifact doesn't exist yet. What exists is judgment — pattern recognition built over twenty years of doing the work, most of it never written down, much of it never even consciously articulated.
When you pitch "version control for knowledge work," you're solving a downstream problem. You're assuming the expert has already converted their judgment into a committable artifact. They haven't. That's the actual bottleneck.
Ask any senior consultant, surgeon, litigator, or operator to write down how they make a hard call. They'll give you a sanitized version that misses 70% of what they actually do. Not because they're hiding it. Because the real reasoning is compressed, contextual, and lives in the gap between explicit rules.
Versioning that sanitized version is worse than useless. It locks in the wrong thing.
The Three Extractions, Restated
In the book I'm writing, I argue every expert is already in one of three economies, whether they chose it or not.
One: LLMs ingested expert knowledge without consent. The UK copyright saga last year — 88% of surveyed creators wanted protection, the government punted — was the polite version of a wholesale extraction that already happened. Their terms.
Two: Platforms like Mercor pay experts hourly to dump IP into training sets. You sign up, you get $80/hr, you hand over the patterns that took two decades to build. Jacob selling his birthright for a bowl of stew. Their terms.
Three: Expert-owned extraction. You own the cards. You own the revenue. You own the distribution. Your terms.
A "version control for knowledge work" tool is a platform play. It sits between you and your IP. The question nobody asks at the demo is: who owns the commits? Who trains on them? What happens when the platform pivots?
If you're versioning your judgment inside someone else's pipeline, you're in economy two with a nicer UI.
What a Refinery Actually Does
A refinery doesn't store crude. It transforms it. Crude oil is worth $70 a barrel. Refined into jet fuel, plastics, and pharmaceuticals, the same barrel is worth several hundred. The value is in the transformation.
The Knowledge Delivery System — KDS — is the refinery I built for expert judgment. Here's the actual stack, not the marketing version.
Capture layer. Voice-first, low-friction, contextual. The expert speaks the way they think on a Tuesday afternoon between calls — not the way they'd write a white paper. Recording the actual reasoning, including the messy parts, the asides, the "well it depends on whether the client has..." The friction of typing kills 90% of capture. Voice removes the friction.
Refinement layer. This is the part nobody talks about and where the real work happens. Raw transcripts are crude. They need structuring — extracting the decision rules, the conditional logic, the pattern triggers, the exceptions. This is where AI earns its keep, not by generating content but by interrogating yours. Asking the second and third questions a junior would never know to ask.
Structuring layer. Refined judgment becomes queryable. Tagged by domain, by decision type, by client context, by confidence level. Not a document. A retrievable asset with metadata that makes it useful to humans, agents, and downstream systems.
Ownership layer. The refined IP sits in infrastructure you control. You decide what gets licensed, what gets fed to your own agents, what gets published, what stays private. The platform doesn't get a copy. The model vendor doesn't get a copy. You own the cards.
Version control happens at the end of that pipeline, not the beginning. By the time something is committable, the hard work is already done.
The Numbers That Matter
For a solo expert billing $300-500/hr, the math runs roughly like this. Twenty hours a week answering the same five categories of questions from clients and team. Ten of those hours are recoverable if the judgment is captured and queryable. That's $3,000-5,000 a week in time you could redirect to higher-leverage work or new revenue lines.
For a firm of fifteen experts, knowledge debt typically runs $2-4M annually in re-explanation, onboarding drag, and judgment that walks out the door when senior people leave. I've watched a thirty-person firm lose what they estimated at $1.2M in recoverable IP when one partner retired and took twenty-two years of pattern recognition with him. Nothing was written down. Nothing was structured. Nothing was owned by the firm.
A refinery prevents that. A repo doesn't, because there was never anything in the repo.
Why the Metaphor Matters
Metaphors decide architecture. If you believe "GitHub for knowledge work" is the right frame, you'll build features around branches, merges, pull requests, and history. You'll optimize for collaboration on existing artifacts.
If you believe "refinery for knowledge work" is the right frame, you'll build features around capture friction, judgment extraction, structuring, and ownership. You'll optimize for converting tacit knowledge into structured IP.
These are not the same product. They're not even adjacent products. One assumes the gold is already in the vault and you need a better filing system. The other assumes the gold is still in the ground and you need to mine and refine it before anyone files anything.
I think the expert who treats their lived judgment as a gift to be stewarded — not a commodity to be platform-versioned — ends up with something that compounds. The one who hands raw transcripts to whatever tool gets traction this quarter ends up funding someone else's training set. That's a stewardship choice, made early, with consequences that show up years later.
What I'm Asking You to Do
Before you pick a tool, pick a metaphor. Then pick the tool that matches.
If your judgment is already structured, queryable, and owned — congratulations, you're in the 2%. Version control might be your next move.
If your judgment lives in your head, your Slack history, and a few half-finished docs — version control is solving the wrong problem. You need a refinery first. The repo comes later, and it should sit on infrastructure you own.
This is the core argument of the book I'm writing — On Whose Terms: The New Expert Economy and the Fight for What You Know. If the thesis resonates, join the launch list and I'll send the first chapter when it's ready.
More on the full stack and what I'm building at mattcretzman.com.
Keep Building,
— Matt