Total Structure v3: Law V — THE MACHINE PLANE

The Valuable Output

Reasoning has physical cost — compute, tokens, retrieval, context, verification, red-team, repair, replay, latency, human review, privacy risk, failure risk. Once reasoning is measured, an economic structure becomes visible that the economy has not yet priced, and the first finding is: the valuable output of reasoning is the proof artifact, not the answer.

A proof artifact is the replayable, ledgered record of a reasoning event: prompt, inputs, definitions, scope rules, the logical-unit graph, dependencies, evidence references, red-team attacks, repairs, unresolved nodes, conclusion, and the hash that lets any verifier replay every step. An answer is disposable; a proof artifact is a durable asset — verifiable, reusable, transferable, challengeable, repairable, amortizable. (A logical unit is the smallest auditable inference step: true, false, unknown, conflicted, insufficient, or out of scope.) Under A₁₁, the proof artifact is simply the receipt of a reasoning event — one species of the universal primitive.

The Economic Primitive

Surety is not confidence. Confidence is the model's report on itself; surety is what survives external test:

Surety = Correctness × Auditability × Reproducibility × Adversarial Survival

The multiplicative form is load-bearing: any factor at zero collapses the score. Correct but unauditable → zero. Reproducible but unable to survive attack → zero. (A₁ operationalized: a claim untested by its adversary has no structural definition.)

Logical energy is the total physical and symbolic cost of producing and sustaining the proof across its lifecycle. The compressed primitive:

Logical Density = Surety / Logical Energy

— how much survives audit per unit of cost. The rigorous form, which real decisions price:

Task-Adjusted Logical Density = Expected Verified Decision Value / Total Lifecycle Logical Cost

where Expected Verified Decision Value = Task Stakes × Correctness × Auditability × Reproducibility × Adversarial Survival × Actionability × Freshness, and Total Lifecycle Logical Cost = generation + retrieval + context + tool use + verification + red-team + repair + human review + privacy risk + failure risk + replay/adaptation + latency/opportunity cost. At unit stakes, unit actionability, unit freshness, and zero latency cost, the rigorous form reduces to the compressed form.

Three disciplines keep the primitive honest. Latency lives in the denominator — a perfect proof delivered after the deadline has zero verified decision value; a lower-surety answer can dominate when delay destroys the opportunity. Validity is bounded — a proof artifact has value only within its declared scope, freshness window, and similarity class; reuse requires contextual similarity, dependency freshness, and verification cheaper than regeneration; an artifact that does not declare its bounds is half-built. Amortization is the mechanism — average cost falls as valid reuses rise; where reuse occurs, logic-dependent remedy cost falls toward the cost of an API call or a local model run; where it does not, the economic argument weakens. The amortization rate is empirical and remains a declared falsification surface.

Alpha as Energy Competition

Alpha — novelty that dominates an existing field — has a calculable energy price: the total expenditure across probabilistic search required to discover a dominating pattern. The transition from the stochastic era to the deterministic era is therefore an energy-cost competition, not a philosophical phase change: a deterministic revision producing equivalent or superior alpha at lower recurring cost reveals the probabilistic assumption as more expensive than necessary. Probability is not disproven; it is outcompeted on the recurring cost of expression wherever a deterministic path is discoverable. Many tasks have no such path. Many do. The arbitrage is in finding the second kind and converting them. (A₂'s operational content: order along the grain is cheaper, and cheaper eventually wins.)

LLM-as-OS

Stochastic models do not become deterministic. The determinism lives one layer up.

Stochastic weights become dynamic reasoning plumbing — used where probabilistic generation is genuinely needed, replaced where it is not — and a deterministic command plane sits above them, electing per task: which model(s); which scaffold depth; which context package and sources; which tools; which red-team depth and adversarial budget; which privacy mode and data custody; which ledgering standard; which human-escalation threshold; which cost ceiling and surety target. The command plane's objective function is task-adjusted logical density under the task's constraints.

This is glass box over black box: weights opaque inside, every routing decision, context slice, tool call, claim, attack, repair, and unresolved node visible on the surface and replayable from the ledger. Three constraints make it honest rather than theatrical:

The glass box must itself be glass. A command plane that records what the models did but hides what the plane decided is a one-way mirror with the operator behind it. The meta-decisions — routing, admission, reuse, red-team allocation — must themselves be typed, logged, replayable, challengeable, expiry-limited, and revocable. If the meta-decisions are not auditable, the audit is theater.

The admission invariant. All context, tools, model outputs, router decisions, proof artifacts, and reuse events are untrusted until admitted: declared type, declared scope, verified provenance, permission check, adversarial check, expiry, entry into the replayable proof graph. Anything entering the proof without admission is contamination. The default state of an input is hostile; verification is what makes it usable.

Structural isolation where stakes warrant. The instance that reads raw context must not be the instance that architects the proof graph for high-stakes decisions — an instance that both reads adversarial input and decides what counts as evidence is one injection away from a captured proof. Ingestion, construction, verification, red-team, repair, and ledgering separate into distinct instances as risk requires. Separation makes capture expensive. (Law III's checking network, rebuilt in silicon: same capture mechanics, same remedy.)

The claim is not that any one configuration dominates. The claim is that determinism-at-the-command-plane dominates unscaffolded probabilistic generation for any task whose output must survive audit — and that LLMs become reliable agency infrastructure precisely when wrapped this way, and not before. Unauditable AGI, if it arrived, would be generalized opacity, not agency infrastructure.

The Floor and the Ceiling

The floor is remedy. Where actors are trapped by logic-cost — the tenant who cannot decode the lease, the worker who cannot prove the harm, the small business that cannot afford compliance review — lowering the cost of proof lowers the cost of agency. As proof cheapens, the floor rises; the anti-subjugation function lives here. (Where the trapping condition is material scarcity, cash transfer remains the correct instrument; the two are complementary.)

The ceiling is ascent. The same protocol applied to capable actors and functioning systems compounds decision quality, ratios, learning, contracts, governance, execution.

One protocol, two positions on one gradient — because subjugation and inefficiency are the same deviation from full-scope optimality: a system burning energy to hold itself below what it could produce. The friction that traps the powerless is the friction that drags the powerful. The same invariant unwinds both.

The Two Eras

The deterministic era does not require AGI. It requires deterministic scaffolds, role-separated verification, replayable ledgers, revocable trust, and per-task routing. In v1.1 this paragraph ended: these exist now; they are deployable now. In v3.0 it ends differently: they are deployed. Law VI is the record.

---

---

Corpus map

• Canonical shelf: Total Structure root