XO Labs · Research · Draft v3, July 2026
A unit of work for intelligence.
The token meters what AI consumes. Nothing meters what it delivers. The quirq is the missing output meter: value budgeted by a human owner, minted by machine verification against captured world state, costed all-in from inference to intervention, and read over time as a trajectory.
A genuinely good unit, for exactly half the question.
The number a business has been missing.
Divide one by the other and every number an operator needs falls out: cost per quirq quirqs per dollar quirqs per kWh and, tracked over time, the AI effectiveness trajectory of a company.
A quirq is never claimed, it is minted. The environment that hosts the work snapshots the world before execution and after, scores completion V against a weighted definition of done, and mints V · B quirqs from what verifiably changed. The environment is the prism: nothing passes through unrefracted.
A machine-checkable definition of done: weighted predicates over world state, stated before execution.
The owner prices the outcome: the amount they would pay for it to exist. The task carries B potential quirqs.
The agent works inside an environment with runtime, memory, files, tools. Every token and API call is metered as a side effect.
Snapshot before, snapshot after. Completion V scored from the world itself, never the agent's account of what it did.
Verification mints V · B quirqs of delivered work into a tamper-evident ledger, alongside everything production consumed.
The whitepaper's worked example: a support ticket, budgeted by its owner at B = $4.00, with three weighted checks. Toggle the checks, move the budget, and watch what the environment would mint. Nothing here is self-reported: each check is a predicate the environment evaluates against the after-snapshot.
Incomplete work is automatically more expensive per quirq, which is the incentive pointing the right way. A unit that settles below threshold is an intervention event, counted for free.
Against minted quirqs the environment meters everything production consumed. The same worked ticket, at all-in cost C = $0.128:
An AI program whose outputs each require twenty minutes of senior review is paying its largest cost in a currency the token bill never sees. Under quirq accounting it cannot hide.
A worked quarter for a mid-size support-and-engineering operation, every number computed from the calculus. The token bill, read alone, says costs nearly doubled. The quirq ledger says something else entirely.
| Month | Units | Minted Q | All-in C | QER |
|---|---|---|---|---|
| April | 2,100 | 15,770 | $5,020 | 3.1× |
| May | 3,400 | 26,310 | $6,480 | 4.1× |
| June | 4,800 | 38,000 | $6,830 | 5.6× |
Illustrative arithmetic from the whitepaper, exhibiting the full calculation chain. The validation program's job is to replace it with production ledgers.
In tokens alone, spend rose +83%: a number indistinguishable from waste. In quirqs, verified value per all-in dollar rose +81% while human rescue fell a third: the difference between an expense line and an investment case.
Because the cost model retains token counts, physical accounting composes with value accounting. Tokens and quirqs are duals: coupled only through the world, which is exactly where a business wants them coupled.
A month at 2.1B tokens and 1.5 J/token is ≈875 kWh. If it minted 148,000 quirqs, that is the plant's efficiency frontier.
The sustainability number a board can act on: not "we used less AI" but "more verified value per unit of energy."
"Is AI working here" becomes a trend line instead of an anecdote: the AI effectiveness trajectory of a company.
Quirq accounting does not escape Goodhart's law; it is engineered so that gaming leaves marks where activity metrics fail silently. The attack surface, in order of severity:
AttackIf quirqs are the KPI, inflate B.
DefenseThe budget is set by the party paying it, so inflation is self-taxing wherever budgets clear against real money. Notional budgets are benchmarked against historical human cost, market rates, and sampled value audits. The ledger makes inflation visible as a drifting ratio of budget to audited value.
AttackManufacture units whose checks are trivially green.
DefenseGold-check sampling on a random audit set, plus the owner's acceptance authority at settlement. A unit whose checks pass but whose owner rejects is an audit event, recorded, and the check set is re-specified.
AttackEdit the test rather than fix the code.
DefenseFully mechanical, fully countered: the verification surface is itself under state comparison, byte-identical across the unit. Experiment E2 shows one such check converting 100% silent success into 100% detection.
AttackConvince the scorer to read the agent's summary.
DefenseRuled out by construction: the mint consumes only environment-captured state. Experiment E1 quantifies what re-admitting self-report costs: every false claim settles.
AttackSplit one outcome into many units to harvest divisible partial credit.
DefenseAtomic settlement as the default for outcomes whose value is holistic, and the owner's monopoly on unit creation: workers execute units; they do not define them.
The honest summary: quirq integrity reduces to environment integrity plus budget governance. The first is an engineering property, demonstrated. The second is an institutional property, stated as an open hypothesis, because that is what it is.
The program is hypothesis-first: hypotheses stated at full strength, each bound to numbered experiments on an open harness whose results ledger is content-addressed. A skeptic re-runs and compares hashes rather than trusting the prose.
False completion claims settle at rate zero when read from environment-captured state; any self-report channel settles them at the false-claim rate.
Falsifier · the environment arm settling any incomplete unit, or hardened checks failing to detect verification-surface attacks.
In persistent environments, cost per quirq falls with unit index toward a stable execution floor; fresh contexts pay cold start every time.
Falsifier · flat or rising persistent-arm trajectories, or equal decay in fresh contexts.
Audit-corrected QER trend predicts renewal, expansion, and P&L attribution better than token spend, task counts, or benchmark scores.
Falsifier · QER* failing to outperform those baselines in instrumented deployments.
Under payer-set budgets, benchmark anchoring, and sampled value audits, the ratio of budgets to audited value stays within tolerance under sustained pressure.
Falsifier · systematic budget drift despite the mitigations.
Mock-mode caveat, in full: the scripted agent implements the cost equation, so this demonstrates harness consistency, not agent behavior. Real-mode measurement (E4) is the program's current deliverable.
Every load-bearing claim in the paper is tiered, and the open ones say so on the tin. Where evidence is mechanical rather than empirical, the caption says so, not a footnote.
We publish adversarial validation receipts, including the ones that dispute us. The current draft carries a disputed verdict from a frontier-model audit, and its issues feed the next revision.
"If the program succeeds, 'is AI working here' becomes a number with a trend. If it fails, the ledger will say so, which is the point of having one."