Research & Ideasquirqquirq Foundations

Two meters

The token measures compute, energy, and carbon well. It structurally cannot price work.

Every consequential technology gets two meters: one for what it consumes, one for what it produces. Electricity has the watt and the ton-mile. Labor has the hour and the deliverable. AI has one meter.

Token: the input meter
Compute consumed
Energy and carbon
Inference cost
What the machine drew from the world. Real accounting, wrong side of the ledger for value.
quirq: the output meter
Outcomes verified
Owner-valued (B)
Minted, never reported
What the machine changed in the world, at the value a human set.
divide them → quirqs per dollar (QER) · quirqs per kWh · quirqs per tonne CO₂ · track over time → the AI effectiveness trajectory

What the token measures, exactly

Give the token its due. Inference cost is, to first order, proportional to tokens processed: a model with P active parameters spends roughly 2P floating-point operations per token, so a workload of N tokens implies about 2PN FLOPs, which hardware converts to energy at its achieved efficiency:

E    ≈  2·P·N / η          energy, with η = FLOPs per joule achieved
CO₂  ≈  E · c_grid          carbon, with c_grid = grid carbon intensity

The constants move with architecture (mixture-of-experts activates a fraction of P), batching, caching, and datacenter overhead, and measured deployments vary widely, but the structure is right. An operator who knows their token volume, fleet efficiency, and grid can estimate energy and carbon from the same meter that prices the API bill. Token accounting is real accounting.

Where it stops, structurally

Non-monotonic
A verbose failure costs more than a terse success; a retry loop costs most of all. A metric that rises when work goes badly cannot denominate the work.
Model-relative
The same outcome costs 10x different tokens across models and prompts: a ruler made of rubber.
Value-blind
Tokens resolving a $4 ticket and tokens reviewing a $25,000 contract look identical in the bill.

This is why enterprise AI budgets climb while ROI stays undemonstrable: not because the systems do nothing, but because a token bill is incommensurable with anything a business already measures. It swings with verbosity and retries, pools unrelated work into one line item, and prices the machinery rather than any result. The measurement crisis of enterprise AI is a units problem.

The division of labor

Tokens meter the machine side completely: cost, energy, carbon. The quirq meters the work side: outcomes delivered at human-assigned value, minted by verification. They are duals, coupled only through the world: no sequence of inference operations changes minted quirqs except by changing the state the checks evaluate. Divide them and the missing numbers appear:

quirqs / $ all-in     is the program paying?          (QER)
quirqs / kWh          value per unit of energy        (bridge)
quirqs / tonne CO₂    value per unit of carbon        (bridge)

The work side needs a denominator: something to count value in. Next: the denominator, and why the human-set budget is the only one that survives.