The XO whitepaper · in twelve slides
Agents now do real work — but the token bill only measures what they burned, never the impact they generated. quirq counts that impact: every verified, owner-valued outcome minted as a unit you can price, compare, and compound.
Suraj Sharma · XO Labs · ▮
01 · The problem
The token bill tells one side of the story perfectly: what the machine burned. It says nothing about the other side — what the work was worth. Every AI program today runs on half a dashboard: a precise meter for spend, and no meter at all for the thing the spend was for.
We need a second meter — one that reads the right thing. Building it is this paper.
02 · The question
Every business is spending more on AI this quarter than last. The question that actually matters is harder: is my team getting more efficient over time using AI — or less? Today, nothing on the invoice can answer it.
The answer this paper proposes
quirq — a unit of work that measures human and AI combined productivity by outcomes, not tokens, messages, or prompts. Count the outcomes delivered and what they were worth, and the efficiency question finally has a number.
03 · Why tokens fail
i · Non-monotonicity
More tokens is not more work. An agent that fails and retries burns more tokens than one that nails it first try — so a bigger bill can mean worse work. Every team wants to justify token spend per outcome, but in the traditional sense that number is nearly impossible to calculate.
ii · Model-relativity
Tokens don't mean the same thing across models. The same job can cost 10× more tokens on one model than another — so you can't compare the cost of two agents by their bills. Swap the model and the ruler changes size.
iii · Value-blindness
Cost and volume aren't results. A business needs to measure what AI actually did for the business — not just what it cost or how much it produced. A $4 ticket and a $25k contract look identical in the token bill.
Tokens still meter the machine side completely: cost, energy, carbon. They just can't cross to value.
04 · The missing half
Nobody buys gasoline to own liters — they count the miles. Every form of energy we've ever metered comes as a pair: a unit for what goes in, and a unit for what comes out.
AI is the first industrial input we bill by the unit-in with no unit-out at all. The paper's answer to the question mark: the quirq — a unit of verified, owner-valued outcome.
05 · The vehicle
A budget needs a vehicle — something ownable, checkable, settleable. A prompt is none of these: stateless, unowned, uncheckable. The unit of work lives in an environment instead.
S₀ · the before-state
A snapshot of world state at creation — files, tickets, records. Whatever the workspace observes, captured by the environment, not reported by the agent.
G · definition of done
Weighted, decidable checks {(g₁,w₁)…(gₙ,wₙ)} over a state snapshot. Stated before execution.
B · the budget
What reaching the outcome is worth, priced once, by its owner. The unit's potential quirqs.
The environment hosts the runtime, memory, files, tools, budget, and record — and is therefore the effective specification, the scorekeeper, and the binding constraint.
06 · The lifecycle
01
Define
The outcome and its checks, before any execution.
02
Budget
What reaching it is worth — B, set by the owner.
03
Execute
Inside a workspace that meters every token, second, and call.
04
Verify
Compare captured before- and after-state. Never the agent's account.
05
Settle
The budget against metered cost. Quirqs mint; the ledger appends.
History is a hash chain — tamper-evident by recomputation.
07 · A relatable example
A business owner doesn't care about tokens, messages, or prompts. They care that the site they priced at $1,000 actually ships. So the deal is written as a unit of work: the budget up front, and a checklist that decides "done."
| Definition of done — set before any work starts | Weight | Delivered? |
|---|---|---|
| Site is live on the owner's domain | 0.5 | ✓ |
| All five pages built, looks right on a phone | 0.3 | ✓ |
| Contact form emails the owner | 0.2 | ✗ not yet |
What got delivered
V = 0.8 → $800
80% of the checklist passed, so $800 of the $1,000 counts as delivered — or nothing counts until the form works, if the owner said "all or nothing."
What it cost, all-in
$38
Every token, compute-second, and API call the agent used — metered automatically. Nobody counted prompts.
When the form is fixed
$962 margin
$1,000 delivered on $38 of cost — a 26× return, in dollars the owner already understands.
The token bill appears exactly once in this story — as a cost line. The unit the owner thinks in, budgets in, and settles in is the outcome.
08 · The mint
Upon verification at score V, the environment mints V·B quirqs of delivered work. Within a company, a quirq is one dollar of verified, owner-valued, delivered work — directly comparable to payroll and revenue.
atomic settlement (τ = 1): mints 0.00 until every check passes — the owner declared partial completion worthless
09 · The cost side
The environment meters all of them as a side effect of hosting the work — so the cost ledger is complete by construction.
The last term matters most and is most often omitted: outputs that each need twenty minutes of senior review pay their largest cost in a currency the token bill never sees. Under quirq accounting it can't hide — interventions are exactly the V < τ events the scoring rule already counts.
10 · The company ledger
QER · efficiency
Σ Q / Σ C
Is the output generated by AI getting better over time?
QV · velocity
Σ Q / |T|
The overall ship rate and velocity of the system.
IR · intervention rate
|{V < τ}| / |U|
How often human time was needed to guide the AI and intervene.
All dimensionless ratios — quirqs per dollar, per kWh — so the unit travels across companies and currencies.
11 · The thesis
As a sentence
Price the outcome up front, verify it by state, and count the verified dollars — not the tokens.
As an equation
Q = V · B
impact = was it delivered × what it was worth. Cost lives in a different equation entirely.
As the missing unit
Liters get miles, kilowatt-hours get lumens — tokens finally get their output unit.
As two meters
Tokens meter the machine: cost, energy, carbon. Quirqs meter the work. Run both, and the ledger is whole.
12 · Don't take the paper's word
Point the calculus at a real GitHub repository, set the budget, and let the arithmetic return the verdict. Every claim in the paper is tiered — sourced, derived, measured, open — with a public validation program.
quirq · XO Labs · suraj@xo.builders