Grafton.
The Verifiable Frontier

Two fundamentally different AI business models

The same frontier AI sits underneath both — but the question the market keeps confusing is simple: who owns the risk of the AI being wrong?

Frontier labs
The software layerWho builds on top of the AI
Platform riskDoes the platform carry the AI risk?
Customer riskDoes the customer carry the AI risk?
Frontier AI
The raw capability. Sold to everyone; takes no responsibility for how it's used.
OpenAI
Anthropic
Google DeepMind
Meta
ToolsMode 1
Sell capability. A workflow that accelerates a professional who validates the output as part of their job.
CursorHarveyGitHub Copilot PerplexityMidjourneyElevenLabs
Takes no AI risk
The wrong-ness is passed straight through to the user. Capability in, responsibility out.
Customer owns the outcome
The buyer creates — and is liable for — the result. They pay for the tool whether or not it's right.
Service providersMode 2
Sell an outcome. The platform takes responsibility for the result — viable only where success is verifiable.
Sierra
Rare: essentially the only clean example at scale today.
Takes the AI risk
The platform stands behind the result and absorbs the cost of the AI being wrong.
Customer buys an outcome
No risk to carry — and pays only if the AI actually delivers the result.
The point
A tool sells capability — the customer carries the risk of the AI being wrong. A service sells an outcome — the platform carries it, and is paid only if it delivers. Every AI initiative is one or the other. Be clear which you're building.
Source: "AI Strategy — The Verifiable Frontier" · Grafton Labs · LegalHelp wiki
Grafton.
The Verifiable Frontier · Why it matters

Why verifiability matters

It's how we iterate and improve the output — a checkable result is the signal you can loop on. Two loops run on the same principle: one improves the system, the other improves each output before it ships.

1
Auto-research loop
Iterates the prompt & system to their optimum — offline, via synthetic-data runs
Synthetic dataset100s–1000s
Master prompt / system
Generate outputsat scale
Auto-research loop
Evaluate outputs
Identify weaknesses
Refine prompt / system
Iterate dataset & repeat
Repeat until performance converges
Improved system / prompt
iterate the prompt
2
Real-time quality loop
Improves each output before it's published — inline, every single time
User input / context
Generate output
Quality loop
Accuracy check
Consistency check
Completeness check
Issues found → revise output
Repeat until quality threshold is met
Human review / approval
Published output
improve the output
The point
Two loops, one principle. The auto-research loop iterates the prompt & system to their optimum — offline, once. The quality loop improves each output before it ships — inline, every time. Both run on a checkable signal to loop against: that's why verifiability matters.
Source: Content Library #113 "The Verifiable Frontier" · Grafton Labs
Grafton.
The Verifiable Frontier · The Mode 2 test

When can Mode 2 work?

An outcome business needs all three at once. Miss any one — even with the other two nailed — and the economics don't close.

Candidate outcome
1Stakes are lowIs the cost of any single error bearable?
2Accuracy is enoughDoes the model clear the bar for this task?
3Outcome is verifiableIs success binary & machine-checkable — no human judgement?
Verdict
Customer-service tickets
Sierra
Refund, not a lawsuit
Errors cheap to fix
Resolved vs escalated
Mode 2 worksResolved → they get paid. Escalated or unhappy → they don't.
Probate & LPA forms
LegalHelp
Bounded, correctable
Clears the format bar
Accepted by the registry
Mode 2 worksA freshly-crossed pocket — success is a binary receipt from the authority.
Will drafting
Consumer legal
Mega-high, irreversible
Frontier models clear it
Only on review or death
FailsAccuracy is there — but not provable until a human checks it, or the testator dies.
Clinical diagnosis
Hippocratic AI
A life per error
Below the bar
Needs clinical judgement
FailsFails all three. The $2B+ valuation prices capability hype, not deployment economics.
Autonomous coding
Devin
Bugs ship downstream
Strong
“Is this code right?” needs judgement
FailsSame domain as Cursor (a tool) — but tries to own the wrong-ness. One ✗ is enough.
Source: "AI Strategy — The Verifiable Frontier" · Grafton Labs · LegalHelp wiki
Grafton.
The Verifiable Frontier · Deep dive — time to truth

Verifiability is really a question of time

It was never a yes/no. The hidden variable is τ — the time to ground truth — and it runs from 0 to ∞. Where a task sits decides everything.

τ = 0
τ = ∞
truth exists at decision time
no fact of the matter — ever
τ = 0
Knowable now
A rule, an authoritative gate, or a self-evidencing output. Machine-checkable at decision time.
payment clearedaccepted by registryreconciles to zerotests pass
Sierra & Mode 2 Test 3 = the τ = 0 slice
0 < τ < ∞
Knowable eventually
A fact exists, but only the future reveals it. Verifiable-eventually — the slow loop.
void length — weekstenant default — the tenancylitigation — yearsdid it pay?
τ = ∞
Never knowable
No fact of the matter; correctness is contested opinion. The hard floor — can't be engineered away.
"was the advice good?""the right strategy?""is it maintainable?"
Where the residual human lives — accountability, not verification
What sets τ?
The oracle you can reach
1A rule — a deterministic check
2An authoritative gate — registry / acceptance (Land Registry, HMRC, court)
3A future world-outcome — did it pay, did it win, did it not flood
The point
Stop asking "can we verify it?" Ask "how long until we know?" The shorter τ, the faster you learn — and the tighter the loop you can build. That, not raw capability, is the binding constraint.
Source: Content Library #113 "The Verifiable Frontier" · Grafton Labs
Grafton.
The Verifiable Frontier · Deep dive — time × risk

Verifiability is one dial — risk is the other

How fast you can know (τ) is the first axis. What it costs to be wrong is the second. Cross them, and you get the real map of what to automate.

Stakes — cost × irreversibility ▲
Verifiable · but irreversible
Keep a human on it
You can learn — but one error is unrecoverable, and verification doesn't undo the harm.
Insurance — available, but priced for a real, unrecoverable loss.
gas-safety sign-offirreversible filing
Unverifiable · high stakes
The residual human
No oracle, ever — and you can't afford to be wrong. The job is accountability, not verification.
Insurance — barely possible: no oracle exists to price the risk.
"was the advice good?"clinical judgementstrategy
Verifiable · recoverable
Full automation — the Mode 2 heartland
Learn instantly, errors are cheap. "Pay only if it works" is possible here.
Sierrapaymentsbookingregistry filings
Slow-verify · recoverable
Learn slowly, survive
Truth lags but errors are bearable — run a proxy + calibration loop. This is where private data compounds.
default ratestribunal win-rates
0 · nowτ — time to ground truth ▶∞ · never
The point
τ says whether you can learn; stakes say whether you can afford to be wrong while learning. Automate the bottom-left. The top-right keeps a human — and resists insurance.
Source: Content Library #113 "The Verifiable Frontier" · Grafton Labs
Grafton.
The Verifiable Frontier · Deep dive — the edge

You can move the dial

τ isn't fixed by nature — it's partly constructed. First disaggregate the process; then engineer each part down toward τ ≈ 0. That's where the alpha is.

Disaggregate first
One process is many τ. Split it, and treat each part on its own terms.
e.g. a Section 8 possession claim
Draft the notice
grounds & dates · τ ≈ 0
✓ automate & self-verify
Serve & file
court acceptance gate · τ ≈ 0
✓ automate
Win at the hearing
world-outcome · τ = months
◑ human + proprietary data
Then lower τ on the verifiable parts — three ways:
1
Regulation → a procedural oracle
AML asks "did you run the check?", not "is the person clean?" — turning an unverifiable outcome into a verifiable procedure that also bounds liability.
Procedural gate = a gift. Standard-based gate = a trap.
2
Design → a self-checking structure
Force structured outputs, add an acceptance gate, or build a reconcile-to-zero identity so the output proves itself.
Closed output beats open — the strongest single predictor of low τ.
3
A private proxy model
Train on your own slow outcomes to turn a τ = 2-years truth into a τ ≈ 0 calibrated estimate.
This is just underwriting — a moat, not a feature.
The trade
Don't buy what's already τ ≈ 0 — it's priced in. The edge is buying work the market prices τ-high / ∞ that you can cheaply re-engineer to τ ≈ 0.
The moat
AI-on-AI verification only harvests an oracle the world already supplies. So public τ ≈ 0 work commoditises; private slow-verify data compounds — a fresh entrant can't time-travel to acquire it.
Source: Content Library #113 "The Verifiable Frontier" · Grafton Labs