Frequently asked questions

What is value's nature — and is it value that drives a human or an AI agent to decide?

Value's nature: a relation, not a substance. Strip the semantic clothing and value is the rate at which an agent converts a scarce resource into progress toward a goal, relative to a frame the goal fixes. Three faces: its currency is substance (free energy / exergy — what's spent), its structure is information (organized by H, D, I), and its frame is the goal (what gives it direction). That is the precise sense of "value sits between substance and information": value is information given a direction by a goal, paid for in substrate. Remove the goal and there is no value — only energy and bits.

• It is relational, not fundamental. Energy exists with no observer; information needs only a distribution; value needs an agent with a goal. A derived, frame-relative quantity — real and lawful (the same type of thing as information), not a newly-discovered fundamental substance.
• It is not conserved. Value is created by an agent aligned with its world and destroyed by miscalibration (the Second Law of value). Non-conservation is part of its nature.
• Its reference is agent-supplied — the is/ought asymmetry: the world pins a Hamiltonian (substance) and a source distribution (information), but it does not pin the goal.

Does value drive decisions? Only in two honest senses — not as a literal cause.

• As a model — yes, almost tautologically: any agent that chooses consistently can be described as maximizing something. But that's a lens we impose, not an engine we find — and you cannot uniquely recover an agent's value function from its behavior (Armstrong & Mindermann, 2018). There is no fact of the matter about the one value driving it.
• For humans — no, not cleanly: real decisions come from habit, emotion, social context, and conflicting, time-inconsistent drives. "Value" is a useful first-order story, not the mechanism.
• For today's AI — no, literally: an LLM agent completes patterns shaped by training; it is not running a value function at decision time.
• For engineered / future AI — the project's bet, and it is prescriptive: if we build agents around explicit goal/value representations, then value is the right quantity to route, align, and govern them with.

Value does not drive decisions the way force drives motion (world-given, deterministic); it organizes and governs them as a frame the agent — or its designer — supplies. Because the goal is agent-supplied and unreadable from behavior, value can only be the lens that makes decisions legible and the control quantity used to steer agent populations — which is why this project's defensible contribution is a governance lens, not a discovered mechanism.

Value is, literally, already a number — even a price. So why measure it with mathematics at all? Is it necessary?

This is the sharpest challenge the whole project has to survive, so it deserves a serious answer — including where the honest answer is "no."

"Value is already a price" is the trap, not the refutation

This is exactly the situation Shannon faced. Before 1948 you could have said the same about information: "Information is just the message — the words, the telegram. It is already concrete. Why do I need mathematics?" Shannon's move was to notice that the visible thing (the message, its meaning) was not the lawful quantity — the lawful quantity (bits, reduction of uncertainty) was hiding underneath, and only once you stripped away the meaning did the laws appear (channel capacity, the coding theorems).

Price is the same. Price is not value — price is value seen from one particular frame: the market's. The theory's central distinction:

Value is frame-relative (it depends on the agent's goal k); price is frame-independent (the one scalar all traders agree on at equilibrium, the shadow price λ = K/E).

Price is what value collapses to when many agents trade the same resource in a market — an emergent coordination signal, not the underlying quantity. Saying "value is just a price" is like saying "information is just the message." You have mistaken one projection for the thing.

Three things a number/price cannot give you

1. Most AI agents have no price. An agent classifying intents, triaging papers, or routing tickets is creating value but producing no price, trading in no market. If your only instrument is price, you cannot measure it at all. You need a measure that works without a market — that is ΔG and I(X;Y), defined from the agent's outputs and the world, no trading required.
2. A price is a point; you need the laws. Price gives an exchange ratio now, at equilibrium. It says nothing about how much value an agent can possibly generate (the ceiling ΔG ≤ I(X;Y), set by its perception, not its price), how much it wastes through overconfidence (dissipation D(q‖p)), or how much a whole fleet can produce together (capped by H(X)). Those are limits and dynamics — a number cannot carry them, only a theory can. It is the difference between knowing the price of electricity and having thermodynamics.
3. Governing requires the substrate, not the exchange rate. To manage populations of separated AI agents you must connect value to the physical resource they burn (compute, energy) and ask "is this agent converting joules into goal-progress efficiently, or dissipating them?" Price does not touch the substrate; the value measure (V = Σ kᵢ ln eᵢ, resource → goal-progress) is built on it.

The honest part — where the answer is "no"

This is the half most people will not tell you, and it is the project's credibility: the math is necessary for some claims and not for others.

• Necessary for: the value-generation ceiling (I(X;Y) — pure information theory, no price equivalent), the dissipation accounting, and the is/ought asymmetry (beliefs have a world-given target, goals do not → alignment as a control problem). Price and standard economics genuinely do not give you these.
• Not necessary for: the cross-frame flow predictions — there the math just re-derives arbitrage / law-of-one-price / general equilibrium. Economists already have those; there the equations are "price, repainted." The project itself concluded exactly this — the cross-frame real-data test came back not resolvable: it collapses to Kelly / general equilibrium.

For trading goods between humans in a market — no. Price already does it; you do not need this. For measuring and governing AI agents that mostly do not trade — where you need ceilings, dissipation, and alignment — yes. Price cannot reach those, and that is the only place the theory earns its keep.

The mathematics is necessary exactly where it tells you something a price cannot — and the discipline of the work is admitting it is redundant everywhere else. That honesty is the point.

Can it estimate the price of things that do not currently have one?

Yes for one meaning of "price," no for the other.

• ✓ Internal / shadow prices — yes. For things with no market, the theory assigns a principled valuation relative to a stated goal: the shadow price of the budget λ = K/E, and the value-per-cost of each agent I_a / cost_a. This is exactly what the Value-Meter does: take agents with no prices, return a defensible price for each, and decide where to spend a budget. Pricing the unpriced within a frame you choose is a core use.
• ✗ External market price — no. A market price is frame-independent because the market makes it so — it is the equilibrium of many frames plus supply, liquidity, and expectations. You cannot conjure it from one agent's value function; you need the actual market. And where a market exists, the theory reproduces general equilibrium and adds no pricing edge. For a genuinely unpriced external good no frame-free "true price" even exists — interpersonal value comparison is provably non-canonical (Arrow).

It gives a frame-relative valuation and an internal shadow price for the unpriced — useful for allocating resources inside agent systems — but it does not discover the objective external price, which by definition only exists once a market produces it.

Can value be stored and transferred like money?

The storable, transferable thing is price and resource — not value itself — and the difference is the reason money exists.

Value in this theory is exergy (available useful work), not conserved energy. So unlike money it is:

1. Not conserved — it dissipates. The Second Law of Value: realized value = available potential minus dissipation (G = D(q‖r) − D(q‖p)), and a moving world floors dissipation above zero. Value leaks; a dollar does not.
2. Frame-relative — no single magnitude to move. The same item is worth different amounts to agents with different goals; there is no one number to transfer.
3. Lossy to transfer. Cross-frame transfer is friction-limited and never lossless: move i→j only when V_j − V_i > f, with f > 0.

Money is the transferable shadow of value, not value itself. Because value is frame-relative and lossy to move, a frame-independent, conserved, storable token (money, or the shadow price λ) is invented to stand in for value during exchange. The physics analogy is energy: you can store it in a battery and send it down a wire, but the useful part (exergy) degrades on every store and round-trip. So you can store the substrate (free energy, compute, a budget E) — but it is exergy, it dissipates, not a frozen stock; and you transfer via price, because raw value cannot cross frames intact. This is why the theory coordinates agents through a price on shared resources rather than by moving value between them.

Isn't this just generalized Kelly with new terminology?

At the single-agent core, yes — and the project concedes it up front. The capacity theorem ΔG ≤ I(X;Y) and the fleet's Kelly-portfolio operating point are Kelly/Cover results re-read for an arbitrary scarce resource. None of the underlying mechanisms is individually new.

What is added — and what Kelly does not cover — is their unification under one substrate-grounded quantity, the cross-frame / price layer, the pooled-fleet ceiling (G_fleet ≤ H(X) for a fleet that fuses its perception — a corollary of the single-agent theorem; an earlier sum-form claim was wrong and is corrected in doc 04's erratum), and the governance mapping — reading goal-drift-under-selection as a control problem yields the ordering incentive design beats oversight. To be precise about that last one: its two halves are borrowed and conceded — the is/ought asymmetry's value-side is Hume / Armstrong & Mindermann (2018), and the control algebra (‖Vg‖/γ = the textbook velocity error 1/Kv; γ>λ_max = high-gain stabilization) is elementary — the contribution is the application, not a new theorem. The full prior-art comparison and contribution statement are in docs/related-work.md.