Core Primitive
Every AI output you accept without validation is verification debt: trust extended without collateral. Like financial debt it is sometimes the right trade — speed now, checking later — but it compounds quietly, concentrates in the claims you reuse most, and eventually some decision defaults on it. The discipline is not zero debt; it is visible debt: know what is unverified, cap it, and pay it down on the claims your decisions actually load-bear.
Naming the loan
Software engineering gained enormous leverage from one metaphor: technical debt. Ward Cunningham's framing did not tell engineers to stop taking shortcuts — it told them shortcuts were loans, with interest, requiring a ledger. The name converted a vague guilt into a manageable quantity.
Knowledge work in the AI era needs the sibling term. Every time you accept a machine-generated claim without checking it, you are borrowing: you receive the speed of not verifying now, against the obligation of it being true later. Call it verification debt. The claim sits in your notes, your deck, your decision — functional, load-bearing, and uncollateralized. Nothing about it looks different from verified knowledge. That invisibility is the whole problem this lesson exists to fix.
Why this debt compounds
Financial debt compounds by arithmetic; verification debt compounds by reuse. An unverified claim that sits untouched in a note is a small, stable liability. The dangerous ones are the useful ones — precisely because you found a claim compelling, you cite it, build on it, repeat it to colleagues, let a second artifact assume it. Each reuse multiplies the eventual correction cost: not one fix but a propagation-tree of fixes, plus the credibility tax wherever others received it from you.
And AI shifts the whole curve. The borrowing rate used to be limited by how fast humans could produce claims at you; now it is limited by nothing. Fluency — the historical proxy for "someone competent checked this" — is exactly what the machine manufactures at zero cost. Without explicit tracking, a modern knowledge base does not drift into debt; it is born leveraged.
Debt is a tool; hidden debt is a trap
The point is not abstinence. Verifying everything at intake is the debt-phobic bankruptcy: it re-caps your throughput at your reading speed, which the variety law already proved cannot hold. Borrowing is how you move fast — accept the claim provisionally, keep working, verify when the stakes arrive.
The point is the ledger. Three properties make debt manageable in finance, and the same three transfer: it must be visible (the unverified are marked, not camouflaged among the verified), capped (there is a level of outstanding unverified claim you refuse to exceed on decision-critical topics), and scheduled (payment is a cadence, not a mood). The artifact format from Phase 2 provides the instrument — provenance says who you borrowed from, confidence says whether you have repaid, and expiry keeps even verified claims from pretending to be permanent.
Where the interest is highest
A ranked payoff strategy beats a virtuous one. Three factors set a claim's interest rate.
Decision load. A claim a live decision rests on charges more than one in cold storage. Pay these first, always — the census exercise exists to find them.
Category risk. Numbers, legal and medical assertions, quotes, and anything with a named source are high-rate categories: precise, checkable, and expensive when wrong. Vibes-level claims ("users prefer simpler onboarding") carry lower, slower interest.
Propagation. The moment a claim is about to leave your system — into a deliverable, a decision, another person — its rate spikes to infinity. Hence the one covenant worth making absolute: nothing unverified propagates. Verify at the exit gate, or attach the caveat visibly. This single rule catches the majority of real-world defaults, because defaults almost always happen in public.
The gate that collects
Connect this to the machinery you built in the last two lessons. Verification debt is what the jidoka gates measure: an output that passes format and consistency checks but carries "provenance: AI, unverified" is precisely a recorded loan. The exception queue is your collections department; the alarm-audited review cadence is the payment schedule; the collision ledger records the defaults you caught early.
Seen this way, the whole gate architecture has a financial summary: it converts an invisible, unbounded liability into a visible, priced, scheduled one. That is the entire difference between people whose AI leverage compounds and people whose AI leverage one day detonates — not how much they borrowed, but whether anyone was keeping the books.
Sources
- Cunningham, W. (1992). "The WyCash Portfolio Management System" — the original debt metaphor for deferred work.
- Reason, J. (1990). Human Error — latent conditions accumulating toward active failure.
- Ohno, T. (1988). Toyota Production System — defects caught at source cost least.
Put it into practice
Exercise
Take a census of your outstanding debt. Search your notes and recent documents for claims that entered via AI in the last month. For each, mark V (verified — you checked it against a primary source or test), U (unverified — accepted on fluency), or P (propagated — unverified AND reused somewhere downstream). Count the P's especially; they are debt earning interest. Then rank the U's and P's by blast radius: which ones does a real decision currently rest on? The top three are your payoff queue for this week. Most people find the census takes an hour and changes how their stomach feels about their own notes.
Watch for the failure mode
The two bankruptcies. Reckless borrowing: accepting everything fluent, verifying nothing, until a decision defaults publicly and takes your confidence in the whole system with it. And debt-phobia: insisting on verifying every claim at acceptance, which caps your intake at your reading speed and quietly returns you to the variety mismatch the cascade exists to solve. Both come from refusing the actual discipline — treating verification as a per-claim moral question instead of a portfolio to be managed with limits, priorities, and scheduled payments.
Make it stick
Make the debt visible at the point of borrowing. From today, the confidence field of any AI-sourced artifact gets one of two values at save time: verified, or unverified-with-a-date. Your review cadence (weekly is enough) queries the unverified list and pays down by priority — decision-load-bearing first, oldest second. Set one hard covenant: nothing unverified propagates. The moment a claim is about to be reused in a deliverable, a decision, or someone else's hands, its debt comes due in full — verify or visibly caveat. Reuse without verification is how small loans become systemic risk.