V₁
Full
Checkable at decision time, before the output is used. Compiler, calculator, type checker, formal verifier, deterministic transform. Verification is O(1) or O(input); the cost of checking is bounded.
The Deciding Axis
The line between demo-able and deployable.
Every AI system has a verifiability value. V is the axis nobody markets and the one that decides which markets you can enter. Most current AI is V₄ or V₅ being sold into markets that demand V₁ or V₂. The mismatch shows up as litigation, regulatory rejection, and quiet refusal to adopt.
Part 01 · The five values of V
Five ways an answer can be checkable.
V is the sixth of the seven axes. It takes five values, ordered from strongest to weakest. Most of the framework's commercial consequences fall out of which value a system can honestly claim.
V₂
Citation
Checkable against a retrieved authority the output points to. Legal RAG with the statute attached. Medical assistance with the guideline citation. The buyer can re-check by opening the source.
V₃
Delayed
Checkable after consequences play out. Medical diagnosis (was the patient OK?). Autonomous vehicle (did anyone get hurt?). Trading agent (did the position make money?). The system is wrong when something bad happens.
V₄
Statistical
Checkable only in aggregate, across a population. Recommendation systems. Credit scoring. Ad targeting. Individual outputs can't be checked — only the distribution can.
V₅
None
No check exists. Open-ended generation. Creative writing. Aesthetic output. The output is judged on taste. There is no ground truth.
Part 02 · Which V each market demands
Markets have V requirements the way they have power requirements.
A market's V floor isn't a preference. It's a structural property — encoded in regulations, in case law, in insurance contracts, in audit standards. A system below the floor cannot be sold into the market regardless of how compelling its demos are.
Requires V₁ · full, decision-time
Aerospace, semiconductors, payroll, tax computation, blockchain settlement, formal contracts.
"Probably correct" is not deployable. A bridge calculation, a chip mask, a tax return, an on-chain transaction — each must be checkable before the output is used. The substrate is deterministic computation, formal verification, or runtime assurance with a hard envelope.
Requires V₂ · citation
Legal practice, regulated medicine, insurance underwriting, compliance, audit, regulatory filings.
The answer must carry a pointer to its source: the statute, the guideline, the policy clause, the precedent. The cost function picks retrieval-anchored mechanisms; the receipt carries the citation. An LLM that asserts without citing is not deployable here, even if its answer is right.
Acceptable at V₃ · delayed
Medical assistance with human-in-the-loop, autonomous vehicles, robotic control with runtime assurance.
The system can fail safely. Consequences are bounded by an envelope (the clinician's judgment, the safety controller, the supervisor). The system is wrong when outcomes degrade; the cost function biases toward conservatism and the receipt carries the safety-envelope check.
Acceptable at V₄ · statistical
Search, recommendation, content moderation, ad targeting, A/B experimentation.
Individual decisions don't need to be checkable; the distribution does. The system is wrong when CTR drops, when the moderation FPR creeps up, when conversion stalls. Most consumer AI lives here — and most of the field's marketing tries to sell V₄ systems into V₁/V₂ markets.
Acceptable at V₅ · none
Creative writing, brainstorming, conversation, image generation.
Output is judged on taste. No verifier exists because the ground truth doesn't. The frontier model is exactly the right mechanism here — novel synthesis under soft constraints is what it's for. V₅ is not a defect; it's a category. The defect is selling V₅ systems where V₁ or V₂ is required.
Part 03 · The mismatch
Most current AI is selling V₄ into V₁ markets.
A statistical-verifiability system sold as a deterministic one creates a predictable failure pattern. The system passes demos. It loses in court. Regulators reject it. Procurement quietly stops adopting it. Insurance refuses to cover it.
None of this is a moral failure. It is a coordinate mismatch. The system cannot meet the V floor of the market because its mechanism is V₄ — statistical inference over learned weights, without a checkable substrate. Vendors compete on capability claims and hope the buyer doesn't ask for verification. Buyers, increasingly, ask.
The fix is structural. Match the mechanism to the V floor. Where V₁ is required, route to a deterministic substrate or a formal verifier. Where V₂ is required, route to retrieval with citation. Where V₃ is acceptable, use runtime assurance. The cost function takes Vrequired as a constraint; the router excludes candidates that can't clear it. The system that ships into a regulated market is the one that knows its V and routes accordingly.
Part 04 · The mechanisms that clear each V
Each V is a coordinate move, not a vague engineering goal.
Reaching V₁
Deterministic substrate · formal verification · runtime assurance.
The output is checkable in O(1). Compiler. Theorem prover. Lean. Simplex runtime envelope. A constraint-bound mechanism (search with a verifier) gets here when the verifier is formal.
→ Zone Z₁, encoding facts or grammar
Reaching V₂
Retrieval against a citable source · source-linked output.
The output carries a pointer to its authority. Legal RAG. Regulatory lookup. The mechanism is small-model adaptation grounded in retrieval. The receipt carries the citation.
→ Zone Z₂, encoding navigation
Reaching V₃
Runtime assurance · safety envelope · human-in-the-loop.
The system can fail safely. Simplex-style assurance, Lyapunov controllers, supervisor patterns. ISO 10218 / IEC 61508 territory. Failure is bounded by the envelope, not the model.
→ Active substance, signal/body interface
Part 05 · The unlock
The V-aware system enters markets the V-blind one cannot.
Auditable autonomy in regulated markets. Physical AI with V₁ runtime envelopes. Legal AI with V₂ citations. Medical AI with V₃ outcomes you can defend in court. Compliance AI with audit trails that hold. Each one is an empty cell on the Stack — catalogued on the frontier. Each one is buildable now.
These cells are empty because the field has been competing on capability rather than verifiability. Every empty V₁-or-V₂ cell that gets filled becomes an uncontested market — no incumbent can enter from a V₄ baseline without rebuilding the mechanism, and rebuilding the mechanism means abandoning the architecture they're optimizing.
Capability is one number. Verifiability is the gate. Pass the gate and the market opens. Fail it and no amount of capability gets you in.
The convergence in the field — energy-based models, joint-embedding architectures, formal verifiers — is a convergence on V. Each tradition is, in its own way, building toward auditable output. The cost function names the routing problem; verifiability names the gate; the receipt is what the buyer holds at the end.
Continue
The axis that opens markets.
V is one of seven axes. It is also the one that decides where the system can ship.