02-coding-theorem-of-value.md

The Coding Theorem of Value

Thread A from 01-core-formalization.md §8: push the Kelly connection to a growth-rate limit and obtain the closest thing to a Shannon channel-capacity theorem for value. The static log law of 01 gave one allocation result; this gives the rate law — how fast value can be created under uncertainty, and what hard ceiling bounds it.

1. The repeated value game (compounding regime)

In 01 the budget E was fixed. Here it compounds: value realized this round becomes resource next round (an agent that achieves its goal earns more compute, trust, or budget). This is the regime that matters for a long-lived agent.

Each round t:

• the world settles into a state s ∈ S (the goal-relevant situation), drawn s ∼ q (the true law);
• the agent splits its budget into fractions b = (b₁,…,bₙ), bᵢ ≥ 0, Σ bᵢ = 1, across n channels;
• channel i returns a multiplier oᵢ(s) ≥ 0 per unit committed when the world is in state s (goal-progress per unit resource); resource on a channel that does not pay is dissipated.

Budget update: E_{t+1} = E_t · Σᵢ bᵢ oᵢ(s_t). Hence

$$\log E_T = \log E_0 + \sum_{t} \log!\Big(\sum_i b_i, o_i(s_t)\Big),$$

and the long-run value growth rate (per round) is

$$\boxed{,G(b) = \mathbb{E}_{s\sim q}\Big[\log \sum_i b_i, o_i(s)\Big].,}$$

G(b) is concave in b (log of a linear form), so maximizing it is a well-posed convex program — the value-Kelly allocation b* = argmax G(b).

Why the log is forced again

01 derived the log from diminishing marginal value (axiom A3). Here the log reappears from a different premise: the time-average growth rate of a multiplicative process is the expected log of its multiplier (Peters, ergodicity economics — the log is the time-average, not a utility quirk). The logarithm of value is therefore over-determined — two independent routes, psychology and dynamics, force the same form. Over-determination is exactly the robustness signature of a fundamental quantity.

2. Value growth = informational advantage

Take the canonical "horse-race" structure (Kelly's fair-odds case): the states are the channels (n = |S|), committing to channel i pays only if state i occurs, and returns are set against a reference belief r over states: oᵢ = 1/rᵢ. (r is the market / prior / baseline expectation.) Then

$$G(b) = \sum_s q_s \log\frac{b_s}{r_s}.$$

Maximizing over the simplex (Gibbs' inequality): the optimum is to commit proportional to your model of the world, b_s = p_s, where p is the agent's belief about q. With a correct model (p = q):

$$\boxed{,G^{*} = \sum_s q_s \log\frac{q_s}{r_s} = D(q,|,r).,}$$

Maximal value growth equals the KL divergence of the agent's (correct) model from the baseline. An agent whose model is the baseline (p = r) grows value at rate zero. Value creation is informational edge over the prevailing expectation — nothing more, nothing less. This is the precise form of 00 §5's "value created by gradients": the gradient is D(q\|r).

3. The capacity theorem — value-throughput ≤ information-throughput

Now equip the agent with a perception channel: a signal Y, correlated with the world-state X, that it observes before allocating. Optimal play is Bayesian — bet the posterior, b_s = p(s \mid y). The growth rate with perception:

$$G_Y = \sum_{x,y} p(x,y),\log\frac{p(x\mid y)}{r(x)}.$$

The gain over acting on the prior alone (G_0 = D(p_X \| r)):

$$\boxed{,\Delta G = G_Y - G_0 = \sum_{x,y} p(x,y),\log\frac{p(x\mid y)}{p(x)} = I(X;Y).,}$$

Coding Theorem of Value. The incremental rate at which an agent can compound value, by perceiving the world through a channel Y, is at most the mutual information I(X;Y) between the world-state and the perception — and this bound is achieved by Bayes-optimal proportional allocation.

It has the full structure of a Shannon coding theorem:

• Achievability. Betting the posterior b_s = p(s\mid y) attains ΔG = I(X;Y).
• Converse. No allocation does better: any b ≠ p(\cdot\mid y) loses exactly D(p(\cdot\mid y)\,\|\,b) per round (Gibbs), so I(X;Y) is an upper bound.

The slogan from 00 §5 is now a theorem: an agent cannot create value faster than it can perceive the world. Value-throughput is bounded by information-throughput. Perception capacity is the hard ceiling on value-generation rate — the value analog of channel capacity C.

4. The Second Law of Value, made exact

Drop the assumption that the agent's model is correct. Let q be reality, p the agent's model, r the baseline. The agent bets b_s = p_s but the world is drawn from q. Then

$$G_{\text{actual}} = \sum_s q_s \log\frac{p_s}{r_s} = \underbrace{D(q,|,r)}_{\text{available potential}} ;-; \underbrace{D(q,|,p)}_{\text{dissipation: model error}}.$$

Read it term by term:

• D(q‖r) — the value a perfectly-informed agent could extract (the exergy of the situation; the gap between reality and the baseline).
• −D(q‖p) — value dissipated by the agent's misalignment with reality. Zero iff p = q.

This is 00 §4's Second Law of Value made quantitative: realized value = available potential − dissipation from model error, every term in nats. Three consequences:

1. Misalignment is measurable dissipation. Confident error (D(q‖p) > D(q‖r)) drives growth negative — a wrong, certain agent actively destroys value. Calibrated humility is value-preserving.
2. Learning is value-recovery. D(q‖p) is exactly the cross-entropy excess that supervised/predictive training minimizes. So the ML training objective IS the minimization of value dissipation. Training an agent and conserving value are the same operation viewed from two fields.
3. Alignment is a value law, not only an ethics norm. Reducing D(q‖p) — making the agent's model match the world — is mechanically identical to reducing wasted value. This is the cleanest bridge yet from the thesis to AI alignment.

5. Consequences for governing AI individuals

• Each agent's value-generation rate is capped by I(X;Y_a) — the mutual information of its perception channel. A weak model with a noisy channel has low I, hence a low value ceiling, no matter its effort.
• This is the formal reason the Macrokit pattern produces value. A precompiled macro hands the agent the goal-relevant bits directly, raising effective I(X;Y) per joule of compute — it lifts value-throughput without raising the deliberation cost. "Move reasoning to design-time" = "raise the perception channel's mutual information cheaply." The coding theorem is why cheap reflexes can match expensive deliberation on narrow tasks: on a narrow task the macro's I can equal the strong model's at a fraction of the joules.
• Fleet governance = waterfilling perception bandwidth. Total growth Σ_a G_a; raise it by routing scarce perception/information to the agents where its marginal value (goal-stakes × leverage) is highest, and by driving every agent's D(q‖p_a) down (keep models calibrated). Both are now optimization targets with units, not slogans.

6. Boundary conditions — where the theorem's assumptions bite

In Shannon's spirit of naming what the model assumes:

• Compounding is required. The clean rate law needs value reinvested as resource (multiplicative dynamics). For one-shot, non-reinvested value, use the static law of 01 instead.
• The KL form needs the horse-race structure. General returns oᵢ(s) give a concave program with no closed form; G* = D(q‖r) is the clean special case (mutually-exclusive channels, fair returns). The bound ΔG ≤ I(X;Y) is more robust but tightest in this structure.
• Stationarity / ergodicity. Growth rates are long-run averages; under non-stationary q (drifting world) the time-average and ensemble-average diverge and the limit must be handled carefully.
• No-leverage, no-ruin. Betting fractions sum to ≤ 1 (no borrowed resource). Leverage reintroduces ruin risk that the log objective is precisely designed to avoid — but only without leverage.

7. What threads A has now closed and opened

Closed: the value analog of channel capacity exists and is proven — ΔG ≤ I(X;Y), achievable, with a converse. Value-throughput is bounded by information-throughput; misalignment is quantified dissipation; learning is value-recovery. The thesis now has two hard theorems (the log law, the coding theorem) and a quantitative Second Law.

Open (next):

• Cross-frame transformation (the hard one). Everything here is single-frame: one agent's q, p, r, k. Fleet-level value still needs the transformation law between two agents' goal-weight/belief bases and the invariant beyond E. Thread B.
• Dynamics of the model p and goal-weights k. §4 says learning reduces D(q‖p); a dynamical theory of how fast (a value-flavored learning-rate / regret bound) would connect this to online learning and give a transient law to sit beside the steady-state one. Thread C.
• Multi-agent capacity region. Shannon has a multi-user capacity region (broadcast, MAC). The value analog — the achievable region of joint growth rates for interacting agents sharing a world — is the natural unification of A and B, and the real prize for governance.