The Joule Standard: A Unified Economics of Work, Trust and Intelligence

Research question

The synthesis asks:

Can work, machine intelligence, settlement assurance, and economic realization share one energy-normalized accounting architecture while retaining their different meanings, units, causal requirements, and institutional boundaries?

The answer is yes for a typed vector account and no for an undeclared scalar account. This answer follows from the foundations in Paper 1, the typed transduction graph in Paper 2, the task-service measure in Paper 3, the scenario-conditioned assurance account in Paper 4, the causal inference chain in Paper 5, the marginal allocation model in Paper 6, the rebound identity in Paper 7, the grid-market model in Paper 8, and the satellite-account structure in Paper 9 [1, 2, 3, 4, 5, 6, 7, 8, 9].

Contributions

The paper makes six synthesis contributions.

1. It defines a type system for energy flows, physical services, task services, money flows, money stocks, probabilities, rates, and indices.

2. It specifies a boundary manifest that is common across the nine application papers and separates physical scope from causal attribution.

3. It proves non-addition, non-aggregation, composition, and Pareto-selection results and states the assumptions that limit each result.

4. It publishes an equation and notation crosswalk so that a symbol used in one application cannot silently acquire another meaning in the synthesis.

5. It joins physical-lineage and national-account controls in one double-counting audit.

6. It supplies executable reference code and adversarial tests for the generic rules rather than hard-coding any paper-specific answer.

Precise nonclaims

The Joule Standard makes the following twenty nonclaims explicit.

1. Joules are not a currency, utility unit, welfare unit, or conserved quantity of value.

2. Energy alone does not determine price, intelligence, legitimacy, trust, or social order.

3. Exergy does not extend to semantic value without a task, observer, institution, and value functional.

4. Tokens are not useless operational measures. They remain useful under a fixed tokenizer and service contract, but they are not stable value units.

5. $\operatorname{QAIJ}$ is neither general intelligence nor realized economic value.

6. Aggregate proof-of-work energy does not identify security, attack cost, censorship resistance, or losses prevented.

7. Waiting for confirmations does not necessarily cause proportional network energy use.

8. Gross protected exposure is not settlement output or value added.

9. Observed revenue per joule does not identify causal productivity.

10. A causal ratio with a weak or zero energy effect need not have a stable finite interpretation.

11. A high average value per joule does not identify the best recipient of the next joule.

12. The marginal-allocation rule does not automatically apply under market power, nonconvexity, omitted constraints, or unknown response curves.

13. Efficient AI does not necessarily lower or raise total energy.

14. Flexible computation is not automatically beneficial to the grid.

15. Direct-current power flow or reserve headroom does not prove voltage, frequency, or dynamic stability.

16. Settlement closure does not prove price formation, truthful bidding, or social optimality.

17. A value-per-joule satellite account neither replaces nor enlarges GDP.

18. Primary, final, operational, and lifecycle energy are not additive stages.

19. Equal weights do not create a politically neutral composite.

20. Passing code tests does not upgrade empirical or causal claims to mathematical proof.

Energy productivity, exergy, and useful work

Energy productivity and energy intensity are established indicators. Patterson distinguishes thermodynamic, physical, economic-thermodynamic, and economic measures, showing that the label “efficiency” does not fix a numerator or boundary [10]. Ayres and Warr place useful work in historical production analysis, while Warr and coauthors construct long-run useful-work series [11, 12]. Exergy analysis asks how much useful work a system could obtain as it approaches equilibrium with a reference environment. It is indispensable for locating physical irreversibility. It does not determine whether the resulting service is wanted, correct, adopted, or welfare improving.

That boundary is important. Let $E_b^e$ denote an energy input, $X^{\mathrm{ex}}$ the exergy carried by that input relative to a reference environment, and $U^{\mathrm{work}}$ useful physical work. Then physical measures such as $U^{\mathrm{work}}/E_b^e$ or $U^{\mathrm{work}}/X^{\mathrm{ex}}$ compare commensurate physical quantities. A monetary realization $\Delta V/E_b^e$ requires prices or another value functional, a baseline, and an attribution design. Replacing $U^{\mathrm{work}}$ by $\Delta V$ is not a unit conversion. It changes the object being measured.

Information, computation, and physical limits

Shannon’s information measure concerns uncertainty in messages, not semantic usefulness [13]. Landauer identifies a thermodynamic cost for logically irreversible operations under stated physical conditions [14]. These results motivate a careful connection between energy and computation, but neither provides an economic valuation of an inference. A token may carry different semantic content under different tokenizers. A correct answer may be useless because it arrives late, cannot be verified, or is not adopted. A useful answer may have large economic effects in one context and none in another.

Machine-learning evaluation has accordingly moved beyond raw throughput. Benchmark suites report accuracy, robustness, calibration, latency, and system performance under declared tasks and hardware [15, 16]. Paper 3 uses those ideas to define a quality-adjusted task service. Paper 5 then adds adoption and causal realization. The synthesis preserves this separation: quality adjustment is still not a welfare functional.

Settlement, assurance, and adversarial systems

Nakamoto’s proof-of-work construction joins a chain-selection rule to an economic cost of rewriting history [17]. Formal analyses give security properties under explicit network and adversary assumptions [18]. Economic analyses show that attack incentives depend on rewards, rents, market structure, outside options, and the value at risk [19]. Energy expenditure is one input to a security mechanism. It is not a complete measure of censorship resistance, finality, decentralization, or expected loss.

Paper 4 therefore treats assurance as a vector and permits avoided loss per joule only for a matched intervention under a named scenario. This synthesis retains that constraint. The scalar $\operatorname{ALPJ}$ is a conditional monetary decision statistic, not a physical law and not a synonym for trust.

Causal effects and accounting boundaries

An incremental economic numerator is a counterfactual contrast. Potential outcomes distinguish such effects from observed differences [20, 21, 22]. When the denominator is itself an estimated intervention effect, a ratio of effects can become weakly identified near zero. Paper 5 makes this issue explicit and recommends joint uncertainty sets or Fieller-type logic rather than mechanically dividing two noisy point estimates.

At economy scale, the SNA records production flows and the SEEA connects physical environmental flows to economic accounts [23, 24, 25]. Gross supply-use tables may show a primary carrier, its conversion output, and final use. Those appearances are useful for balancing accounts, but their sum is not a total energy input. Paper 9 contributes an energy-stage bridge and keeps AI outcomes and settlement assurance in memorandum panels. The present paper turns that architecture into a general typed audit.

Composite indicators and social choice

Composite indicators require normalization, weights, aggregation, and sensitivity analysis [26]. Equal weights are still weights. A data-derived weighting method encodes a statistical criterion rather than discovering a natural social objective. More broadly, heterogeneous welfare dimensions cannot be ordered without assumptions about standing, tradeoffs, and aggregation [27, 28]. Energy normalization does not remove those assumptions. It changes the resource basis of the comparison.

The boundary manifest

Paper 1 defines a measurement tuple $$\mathcal M=(b,c_0,W,\tau_H,e^{\mathrm{conv}}, a^{\mathrm{attr}},\Pi^{\mathrm{unc}}), \qquad \operatorname{VPJ}_{\mathcal M}(s) =\frac{\mathbb{E}_{\Pi^{\mathrm{unc}}}[\Delta W(s\mid c_0)]} {E_b^{e^{\mathrm{conv}}}(s;\tau_H)}. \label{eq:paper1}$$ The synthesis expands this tuple into an operational manifest. For a study $r$, define $$\mathfrak M_r= (f^{\mathrm{unit}},p^{\mathrm{system}},\tau_H,\ell^{\mathrm{geo}}, c_0,stage^E,a^{\mathrm{shared}},i^{\mathrm{failure}}, m^{\mathrm{embodied}},W,a^{\mathrm{attr}},\Pi^{\mathrm{unc}}, z^{\mathrm{exclude}}), \label{eq:manifest}$$ The entries are, in order, the functional unit, physical system, temporal interval, location, counterfactual, energy stage, shared-infrastructure rule, idle and failure rule, embodied-energy scope, outcome functional, attribution rule, uncertainty model, and exclusions.

Every denominator is accompanied by the record $$\begin{aligned} D_E=(\,&\text{stage},\text{physical boundary},\text{geography}, \text{interval},\\ &\text{allocation},\text{idle/failure treatment}, \text{embodied scope},\text{lineages}). \end{aligned} \label{eq:denominatorrecord}$$ An equation that uses a short symbol $E$ inherits this record from its manifest. The short symbol is typography, not permission to omit the fields.

The expansion does not change the theory in Paper 1. It makes recurrent implementation choices explicit. A direct comparison between two reported ratios passes the boundary gate only if the relevant fields match or a documented bridge maps one field to the other. A common string such as “facility energy” is not a bridge. The analyst must show that metering, allocation, time, location, and inclusion rules align.

Estimand-family tags

Identical printed units can conceal different mathematical objects. Every energy-normalized result in the synthesis therefore carries one of six estimand-family tags:

Level intensity.

$I=X/E_b^e$, where $E_b^e>0$ is a matched inventory. Paper 1’s basic $\operatorname{VPJ}$ is in this family when its expected numerator is divided by a fixed positive inventory.

Ratio of expectations.

$I_{\mathrm{RE}}=\mathbb{E}[X]/\mathbb{E}[E_b^e]$. Paper 3’s $\operatorname{QAIJ}$ is in this family.

Mean of unit ratios.

$I_{\mathrm{MR}}=\mathbb{E}[X/E_b^e]$. This is generally not $I_{\mathrm{RE}}$.

Causal ratio of effects.

$\theta^{\mathrm{causal}} =\Delta_V^{\mathrm{ATE}}/\Delta_E^{\mathrm{ATE},b}$, where both effects describe the same intervention. The denominator can be positive or negative and the ratio does not exist at zero.

Marginal derivative.

$m_{i,n,t}=\partial B_i/\partial e_{i,n,t}$, evaluated at an operating point under constraints. This is not an average intensity.

Descriptive national intensity.

A matched period production or outcome flow divided by a matched period energy flow. This is not causal absent a separate design.

Typed quantities

An amount alone is not enough for an accounting operation. The synthesis type of a quantity $\xi$ is $$\operatorname{type}(\xi) =(\sigma,u,k,b,\tau), \label{eq:type}$$ where $\sigma$ is its semantic name, $u$ its unit, $k$ its accounting kind, $b$ its boundary identifier, and $\tau$ its horizon identifier.

The accounting kinds used here are energy flow, physical service, task service, money flow, money stock, probability, rate, and index. “USD” does not erase the distinction between a flow of value added during a quarter and a protected asset stock at quarter end. “Count” does not erase the distinction between generated tokens and verified completed tasks. “Joule” does not erase the distinction between primary supply, conversion output, final energy, facility operation, and a lifecycle inventory.

Energy stages and physical lineage

For an energy component $j$, record $$\mathfrak e_j=(\text{id}_j,\chi_j^{\mathrm{lineage}}, stage_j^E,E_j),$$ where $\chi_j^{\mathrm{lineage}}$ is a physical-lineage identifier and $stage_j^E$ is an energy stage. The standard distinguishes device operational, system operational, facility operational, uniquely allocated embodied increment, lifecycle, primary, conversion output, final, and marginal-grid energy. A total is legal only over a non-overlapping set of lineages at a coherent stage.

Consider $100$ J of primary fuel entering a converter that produces $35$ J of electricity, of which $32$ J reaches final use. A physical supply-use table may display all three entries. The energy input is not $167$ J. The entries are linked stages of one lineage. A bridge may report conversion loss $65$ J and network loss $3$ J, but the stages are not independent resources.

Physical efficiency and economically realized value

The accepted program contains two local AI chains that use $O$, $U$, and $T$ differently. The synthesis resolves that collision with the canonical optional-node graph $$E_b^e \longrightarrow C_{\mathrm{named}} \longrightarrow O_{\mathrm{emit}} \longrightarrow U_{\mathcal P} \longrightarrow A^{\mathrm{adopt}} \longrightarrow \Delta\mathbf o_{c_0} \longrightarrow W(\Delta\mathbf o_{c_0}). \label{eq:canonicalchain}$$ Here $O_{\mathrm{emit}}$ is immediate output and $U_{\mathcal P}$ is a verified useful outcome under protocol $\mathcal P$. For a token-emitting system, the representation statistic $$T_{\kappa}^{\mathrm{tok}} =\left|\kappa(O_{\mathrm{emit}})\right| \label{eq:tokenattachment}$$ is attached to the immediate-output node. It is not substituted for that node and does not become the verified outcome. An adoption node can be bypassed only by a named bridge that explains why produced service equals institutionally used service in the application.

Not every application uses every node. A physical-work branch is $$E_b^e\longrightarrow X^{\mathrm{ex}}\longrightarrow U^{\mathrm{work}}\longrightarrow\text{delivered physical service} \longrightarrow\Delta\mathbf o_{c_0}.$$ A proof-of-work branch proceeds from hashing energy through hashrate, productive share, protocol rules, a named risk measure, economic deterrence, and a loss model. A grid branch proceeds from power integrated over an interval to feasible dispatch, reserve, and cash settlement. The point is not to force one causal graph on every system. It is to prevent a jump from an upstream observable to a downstream claim without a declared edge.

Interface composition

Paper 2 represents an economy as a typed transduction graph. For a path $$X_0^{\mathrm{int}}\xrightarrow{\mathcal T_1^{\mathrm{stage}}} X_1^{\mathrm{int}}\xrightarrow{\mathcal T_2^{\mathrm{stage}}}\cdots \xrightarrow{\mathcal T_n^{\mathrm{stage}}}X_n^{\mathrm{int}},$$ define a stage yield $y_k^{\mathrm{stage}}=X_k^{\mathrm{int}}/X_{k-1}^{\mathrm{int}}$ only after the output of $\mathcal T_{k-1}^{\mathrm{stage}}$ matches the input of $\mathcal T_k^{\mathrm{stage}}$ in amount and type. Then $$\frac{X_n^{\mathrm{int}}}{X_0^{\mathrm{int}}} =\prod_{k=1}^{n}y_k^{\mathrm{stage}}. \label{eq:composition}$$

Correlation, selection, missingness, and endogenous adoption therefore matter. The reference implementation composes exact observed or modeled stages. It does not infer stochastic independence.

Evidence-labeled edges

Each transduction edge carries one of four labels.

PROVED

A definition or algebraic consequence follows from stated assumptions.

COMPUTATIONAL

An exact fixture or tested algorithm produces the result.

CONDITIONAL

The interpretation depends on an empirical design, scenario, or behavioral assumption.

OPEN

The relevant edge is proposed but not identified by the available evidence.

An energy meter can support a computational statement about facility electricity. A benchmark can support a conditional statement about verified task service on its sampled task distribution. Neither alone identifies economic value. The service-to-adoption and adoption-to-outcome edges need their own evidence. This prevents evidence from flowing farther down a graph than the design supports.

For terminal claims, the synthesis publishes the evidence-status vector $$\mathbf s_{\mathrm{ev}} =(\text{algebra},\text{code},\text{measurement}, \text{identification},\text{transport},\text{normative aggregation}). \label{eq:evidencevector}$$ For example, a tested avoided-loss calculation may have $$\begin{aligned} \mathbf s_{\mathrm{ev}}=(&\mathrm{PROVED},\mathrm{COMPUTATIONAL}, \mathrm{CONDITIONAL},\\ &\mathrm{OPEN},\mathrm{OPEN},\mathrm{CONDITIONAL}). \end{aligned}$$ A proved algebra coordinate never upgrades an open identification coordinate.

For a composed transduction path, the reference implementation uses the conservative stage-status order $$\begin{aligned} \mathrm{PROVED} &\prec\mathrm{COMPUTATIONAL} \prec\mathrm{CONDITIONAL}\\ &\prec\mathrm{OPEN} \prec\mathrm{OBSTRUCTED}. \end{aligned} \label{eq:evidencejoin}$$ The path status is the join, which is the most limiting status present. Thus a COMPUTATIONAL meter edge followed by an OPEN service-to-value edge yields an OPEN terminal path. The multi-coordinate vector in the equation remains visible because stage status cannot substitute for measurement, identification, transport, or normative status.

The assurance branch

Settlement assurance is not naturally one scalar node. For protocol $\mathcal P^{\mathrm{settle}}$ and scenario $\omega$, define an assurance record $$\mathcal A(\mathcal P^{\mathrm{settle}},\omega) =\bigl(\alpha_A^{\mathrm{hash}},Pr^{\mathrm{reorg}}, Loss_P,Cost_A^{\mathrm{net}},M_A,\mathbf k\bigr), \label{eq:assurancevector}$$ where $\alpha_A^{\mathrm{hash}}$ is effective adversarial share, $Pr^{\mathrm{reorg}}$ a named reorganization risk, $Loss_P$ loss conditional on success, $Cost_A^{\mathrm{net}}$ modeled net private attack cost, $M_A$ attack margin, and $\mathbf k$ non-energy constraints. Each component keeps its own unit.

Paper 4 locally declares both an energy-account interval and a post-confirmation attack window. The synthesis names them $\tau_E^{\mathrm{hash}}$ and $\tau_A^{\mathrm{attack}}$. Hashing energy is $$E_H(\tau_E^{\mathrm{hash}}) =H_H\epsilon_H^{\mathrm{J/TH}}\tau_E^{\mathrm{hash}}. \label{eq:hashenergy}$$ Risk is evaluated over $\tau_A^{\mathrm{attack}}$. The restriction $\tau_E^{\mathrm{hash}}=\tau_A^{\mathrm{attack}}$ may be imposed for a particular fixture, but it is never implicit.

Paper 4’s $Loss_P$ is loss if the modeled event succeeds. Paper 9’s local “gross protected exposure” is instead an exposure $X_P^{\mathrm{gross}}$. The synthesis relates them only through a declared loss severity $s_P^{\mathrm{loss}}\in[0,1]$: $$Loss_P=X_P^{\mathrm{gross}}s_P^{\mathrm{loss}}. \label{eq:lossseverity}$$ Treating the two as equal assumes $100$ percent severity.

A monetary avoided-loss ratio is permitted only for matched risks and a named incremental intervention: $$\operatorname{ALPJ}(\mathcal P^{\mathrm{settle}};\omega) =\frac{Loss_P\{Pr_0^{\mathrm{reorg}}(\omega) -Pr_1^{\mathrm{reorg}}(\omega)\}}{\Delta E_P}, \qquad \Delta E_P>0. \label{eq:alpj}$$ The risk probabilities are matched under the baseline and intervention.

The expression is algebraically well defined. Its interpretation is CONDITIONAL. Equal energy does not imply equal risk, attack cost, decentralization, censorship resistance, or loss prevention. The protocol, network topology, hardware market, reward schedule, adversary, time horizon, and asset exposure all enter the scenario.

Task service rather than token volume

Paper 3 defines a quality-adjusted service contribution for task $i$: $$s_i^{\mathrm{svc}} =\widetilde d_i^{\mathrm{task}}u_i^{\mathrm{svc}}q_i^{\mathrm{AI}} c_i^{\mathrm{correct}}\alpha_i^{\mathrm{svc}}\ell_i^{\mathrm{lat}}, \label{eq:service}$$ where $\widetilde d_i^{\mathrm{task}}$ is a normalized difficulty weight, $u_i^{\mathrm{svc}}$ usefulness, $q_i^{\mathrm{AI}}$ quality, $c_i^{\mathrm{correct}}$ correctness, $\alpha_i^{\mathrm{svc}}$ reliability, and $\ell_i^{\mathrm{lat}}$ latency or timeliness. The factors and their scales must be declared. For evaluation protocol $\mathcal P^{\mathrm{eval}}$, $$\operatorname{QAIJ}_{\mathcal P^{\mathrm{eval}}} =\frac{\mathbb{E}[\sum_{i\in\mathcal P^{\mathrm{eval}}}s_i^{\mathrm{svc}}]} {\mathbb{E}[E_b^e]}. \label{eq:qaij}$$

The multiplicative form is a preregistered grading aggregation. It is not a physical conservation law and not the interface-cancellation identity in the equation. Raw task difficulty does not enter the dimensionless product without the declared normalization that produces $\widetilde d_i^{\mathrm{task}}$.

This measure is not “intelligence” in the unrestricted psychological or philosophical sense. It is quality-adjusted service on a named task distribution, evaluated with declared scoring and energy boundaries. A tokenizer change can alter token count while leaving user-visible content unchanged. Verbosity can increase tokens without improving correctness or usefulness. Task mix can reverse system rankings. The task distribution is therefore part of the estimand.

Quality is not realized value

A high $\operatorname{QAIJ}$ can coexist with low realized value. The system may not be adopted, may be used for a low-stakes task, or may displace an equally good baseline. Paper 5’s local inference chain is translated into the canonical graph in the equation. Its token count is $T_\kappa^{\mathrm{tok}}$, its verified outcome is $U_{\mathcal P}$, and its adoption node is $A^{\mathrm{adopt}}$. None of those nodes is silently renamed in the synthesis.

Paper 5 locally calls a component sum $E_i^{\mathrm{wall}}$ while permitting a separate facility-overhead component. The synthesis does not inherit that label. It uses $$\begin{aligned} E_i^{\mathrm{system-wall}} &=E_i^{\mathrm{device}}+E_i^{\mathrm{host}} +E_i^{\mathrm{memory}}+E_i^{\mathrm{network,inside}},\\ E_i^{\mathrm{facility-operational}} &=E_i^{\mathrm{system-wall}}+E_i^{\mathrm{cooling}} +E_i^{\mathrm{power-conversion}}+E_i^{\mathrm{shared,facility}}, \end{aligned} \label{eq:wallfacility}$$ with component lineages checked for overlap. A general component sum is named $E_i^{\mathrm{operational},b}$, where $b$ states which of these boundaries applies. Facility overhead is never hidden inside a quantity still labeled system-wall energy.

Let $\Delta_V^{\mathrm{ATE}}$ be an intervention effect on realized value and $\Delta_E^{\mathrm{ATE},b}$ its effect on operational energy at boundary $b$. The causal value-per-energy estimand is $$\theta^{\mathrm{causal}} =\frac{\Delta_V^{\mathrm{ATE}}}{\Delta_E^{\mathrm{ATE},b}}, \qquad \Delta_E^{\mathrm{ATE},b}\ne0. \label{eq:ratioeffects}$$ Both effects must refer to the same intervention, boundary, counterfactual, horizon, and population. Dividing an observational revenue difference by a metered energy effect does not meet that condition.

The executable identification key is $$(\text{intervention},\text{boundary},\text{counterfactual}, \text{horizon},\text{population}). \label{eq:identificationkey}$$ Population is not inferred from the other four fields.

Identification ladder

The synthesis uses an identification ladder rather than one evidence label for an entire system.

1. Metering: identify operational energy for the declared physical boundary.

2. Service evaluation: estimate correctness, usefulness, reliability, latency, and task difficulty on a declared task distribution.

3. Adoption: observe whether the service enters the relevant workflow and whether failed attempts or retries are retained.

4. Outcome effect: identify the counterfactual change in a physical, monetary, or welfare outcome.

5. Attribution: allocate joint effects and shared burdens without adding the same outcome twice.

6. External validity: state the population, time, and institutions to which the result may travel.

Failure at a lower rung does not invalidate observations above it, but it limits their interpretation. For example, a well-measured energy denominator and benchmark accuracy remain useful engineering evidence even when economic realization is OPEN.

Definition

For a feasible system or allocation $s$ under manifest $\mathcal M$, define $$\mathbf J_{\mathcal M}(s) = \left( E_b^e(s); \frac{U^{\mathrm{work}}(s)}{E_b^e(s)}, \frac{S_{\mathrm{task}}(s)}{E_b^e(s)}, \frac{\Delta V(s)}{E_b^e(s)}, \mathcal A(\mathcal P^{\mathrm{settle}},\omega), \frac{\Delta G_{c_0}^{\mathrm{grid}}(s)}{E_b^e(s)} \right). \label{eq:joulevector}$$ The displayed coordinates are illustrative, not mandatory. They represent useful work, quality-adjusted task service, causal monetary realization, scenario-conditioned assurance, and a declared environmental or grid effect. Every coordinate names its native numerator unit. The semicolon after energy is deliberate: energy is also a resource to minimize, not merely a denominator that disappears after ratios are formed.

The vector representation can report “MJ of useful work per J”, “verified task-equivalent per J”, and “2026 USD of causally attributed avoided loss per J” in adjacent columns. It does not add the columns. If assurance is expressed as a probability or attack margin, that coordinate remains a probability or a scenario-specific economic quantity. If the causal value effect is not identified, the vector may report the upstream service coordinate while marking the value coordinate OPEN.

Dominance and the Pareto frontier

Let $\mathcal F$ be a feasible set of allocations. Energy is weakly preferred lower. For every beneficial outcome axis $j$, $z_j$ is weakly preferred higher; cost or risk axes are oriented lower. An allocation $s$ dominates $r$ when it is weakly better on energy and every outcome and strictly better on at least one coordinate. The Pareto frontier is $$\mathcal P = \{s\in\mathcal F: \nexists r\in\mathcal F\text{ such that }r\succ s\}. \label{eq:pareto}$$

Dominance is a partial order. It can reject a system that uses more energy and produces no better outcome. It cannot rank one system that provides more verified task service against another that provides more useful work or assurance.

Conditional scalarization

For a decision maker $d$, let $n_j(o_j)\in[0,1]$ be a published monotone normalization rule and let $\omega_{dj}^{\mathrm{index}}\ge0$, $\sum_j\omega_{dj}^{\mathrm{index}}=1$, be published weights. Include energy as an oriented axis, for example $n_E(E_b^e)=(E_{\max}-E_b^e)/(E_{\max}-E_{\min})$. A conditional view is $$S_d(s)=\sum_j\omega_{dj}^{\mathrm{index}}n_j(o_j(s)). \label{eq:scalar}$$ The subscript $d$ matters. The score belongs to a decision rule, not to physics.

No natural common scalar

The theorem is constructive. It does not say that all weights are equally reasonable. Evidence, rights, budgets, law, or institutional mandates can constrain admissible weights. It says that the ordering does not follow from joules alone. A claim of one “Joule Standard score” must therefore identify the decision maker, normalization, weights, and robustness region.

The next joule

Paper 6 studies the marginal value of energy in an automated economy. Let $e_{i,n,t}$ be MWh assigned to service $i$, node $n$, and time $t$. Let $B_i^{\mathrm{work}}(\mathbf e_i)$ be a declared benefit function and let grid and operational constraints define $\mathcal F$. A scalar social-planner problem can be written $$\max_{\mathbf e\in\mathcal F} \sum_i B_i^{\mathrm{work}}(\mathbf e_i) -\sum_{i,n,t}k_{n,t}^{\mathrm{grid}}e_{i,n,t} -\sum_i\mathcal R_i^{\mathrm{risk}}(\mathbf e_i), \label{eq:allocation}$$ where $\mathcal R_i^{\mathrm{risk}}$ is a declared risk or external-cost functional. Under convexity and regularity, an interior allocation satisfies a marginal condition $$\frac{\partial B_i^{\mathrm{work}}}{\partial e_{i,n,t}} = k_{n,t}^{\mathrm{grid}} +\lambda_{i,n,t}^{\mathrm{KKT}} +\frac{\partial\mathcal R_i^{\mathrm{risk}}}{\partial e_{i,n,t}}, \label{eq:kkt}$$ where $\lambda_{i,n,t}^{\mathrm{KKT}}$ collects the opportunity costs of binding service and feasibility constraints. A declared nodal market price, if used, is $p_{n,t}^{\mathrm{nodal}}$, not this multiplier.

This condition does not produce a universal sector ranking. The marginal benefit function depends on the service, state, counterfactual, and decision objective. The next joule may be most valuable for a delayed manufacturing operation at one node, a latency-sensitive inference at another, or a settlement workload under an elevated threat scenario. Scarcity prices and shadow costs are local and time dependent.

Vector allocation

When the benefit components are not commensurate, solve a vector optimization problem first: $$\max_{\mathbf e\in\mathcal F} \left( U^{\mathrm{work}}(\mathbf e),S_{\mathrm{task}}(\mathbf e), -Pr^{\mathrm{reorg}}(\mathbf e),\Delta\mathbf o_{c_0}(\mathbf e), -\mathrm{CO_2e}(\mathbf e) \right), \quad \min E_b^e(\mathbf e). \label{eq:vectorallocation}$$ The output is a Pareto frontier. A decision maker may then apply the equation, an epsilon-constraint method, lexicographic priorities, or legal constraints. The choice must remain visible. In particular, monetizing every effect is one optional model. It is not required by the Joule Standard.

Flexible workloads and grid settlement

Paper 8 represents AI and proof-of-work workloads as heterogeneous feasible sets in a nodal dispatch model. Both can be price-responsive, but their constraints differ. AI jobs may carry deadlines, data locality, checkpoint costs, latency targets, or quality degradation. Proof-of-work loads may have startup behavior, pool constraints, hardware efficiencies, and protocol revenue exposure.

Paper 8 uses $v_w$ locally for both a downward ramp limit and an audited true value. The synthesis imports neither collision. It writes $r_j^{\downarrow}$ for a load’s downward ramp limit and $\nu_j^{\mathrm{true}}$ for audited true private value. Submitted bid value is $b_j^{\mathrm{submit}}$, which remains distinct from $\nu_j^{\mathrm{true}}$ and from realized downstream social value.

Reserve capability is a power quantity. A reserve offer is stated here in currency per MW-hour of reserved capability, equivalently currency per MWh-reserved for a one-hour interval. For an interval of $\delta_t^{\mathrm{hrs}}$ hours, reserve payment multiplies the MW quantity by $\delta_t^{\mathrm{hrs}}$ before applying a currency-per-MWh-reserved price. This resolves Paper 8’s local currency-per-MW versus currency-per-MWh label inconsistency and prevents power from being treated as energy without integration.

For a direct-current dispatch approximation, node balance and line limits can be expressed as $$\begin{aligned} p_{n,t}^{\mathrm{gen}}-D_{n,t}^{\mathrm{fixed}} -\sum_jp_{j,n,t}^{\mathrm{load}} &=\sum_m b_{nm}^{\mathrm{sus}}(\phi_{n,t}-\phi_{m,t}),\\ -\bar F_{nm} &\le b_{nm}^{\mathrm{sus}}(\phi_{n,t}-\phi_{m,t})\le\bar F_{nm}. \end{aligned} \label{eq:dc}$$ A complete market settlement also needs revenue adequacy and an explicit reserve or uplift rule. The canonical system closure is $$Payment_D-Credit_G=K^{\mathrm{network}}, \label{eq:settlement}$$ where $Payment_D$, $Credit_G$, and residual network or constraint rents $K^{\mathrm{network}}$ balance under the modeled settlement.

The synthesis uses the grid model to locate marginal energy cost and constraints. It does not treat a dispatch solution as proof of alternating current feasibility, voltage security, transient stability, or deliverable reserves. Those require additional models.

Preserve the production core

The national-accounting core remains $$\operatorname{GDP}=\sum_j\operatorname{GVA}_j+ \text{taxes on products}-\text{subsidies on products}. \label{eq:gdp}$$ AI services, data centers, electricity generation, equipment manufacture, mining activity, and payment services can already enter gross output, intermediate consumption, and value added. A satellite account may expose their energy use and outcomes, but it must not add those same production flows to GDP a second time.

Let $G$ be an AI-related flow already included in an industry’s GVA. Let $G^{\mathrm{memo}}$ display the same source flow in a thematic memorandum. Then $$\operatorname{GDP}+G^{\mathrm{memo}}$$ double counts $G$. The memorandum can disaggregate or annotate GDP. It is not an additional production flow.

Stocks, flows, and assurance

Let $K_t$ be an asset position protected or transferred using a settlement system during period $t$. It is a stock or exposure measure, not a production flow. Adding $K_t$ to GDP violates both unit semantics and temporal type. Likewise, a modeled avoided expected loss $K_t\{R_0-R_1\}$ is not automatically value added. It is a scenario-conditioned benefit estimate. Depending on the application, some associated services may already be purchased and recorded in production.

The audit

Every additive account entry should carry:

• a unique entry identifier;

• a source-flow identifier;

• a typed quantity;

• a role, either integrated production, satellite memorandum, or physical memorandum; and

• a classification and bridge to the source account.

The deterministic audit emits a repeated-source finding whenever a source-flow identifier appears more than once. It emits a stronger GVA-satellite overlap finding when the same source appears in integrated production and a satellite memorandum. It also rejects an integrated total that mixes semantic, unit, stock-flow, boundary, or horizon types. This is a useful minimum control, not a complete national-accounts compiler. Source systems must still resolve revaluations, transfers, consolidation, residence, imports, taxes, subsidies, and classification changes.

Twenty-case cross-program screen

The lineage and type rules apply to the following audit cases. A conforming implementation either rejects the aggregation or supplies the named bridge.

1. Do not add primary, conversion-output, final, facility, and lifecycle energy as independent inputs.

2. Do not add manufacturing operation to embodied allocation and then add both again to the economy-wide operational total.

3. Do not add component estimates for host, network, facility, idle, retry, or rework when a covering meter already contains them.

4. Do not charge one shared inference’s full energy to several adopted uses and then sum the branches.

5. Do not add tokens, verified tasks, adopted tasks, and realized outcomes as independent outputs. They are typed stages or nested quantities.

6. Do not multiply quality factors that encode the same failure event.

7. Do not remove failed attempts from energy and then separately claim a reliability adjustment.

8. Do not add AI-enabled output to GVA when the source production flow is already recorded.

9. Do not add protected exposure, expected loss, avoided loss, settlement service, and insurance or risk-management production without an overlap bridge.

10. Do not sum overlapping protected exposures or repeated positions across assurance windows.

11. Do not add load payments, generator credits, congestion rent, reserve credits, and reserve charges without closing cash settlement and identifying transfers.

12. Do not add explicit congestion, loss, or energy components when the locational price already embeds them.

13. Do not charge reserve once as headroom and again as the same reserve requirement.

14. Do not sell one curtailment, battery, backup generator, or interconnection capability into overlapping reserve products without coupled constraints.

15. Do not count a demand-response baseline payment as grid surplus without verified reduction and a production-cost counterfactual.

16. Do not count direct AI energy again inside indirect digital or economy-wide energy.

17. Do not call displaced or shifted energy eliminated when it reappears at another site or time.

18. Do not add private attack cost, electricity-producer revenue, and social resource cost without an incidence account.

19. Do not sum ratios across units. Aggregate matched numerators and denominators, then form the declared ratio.

20. Do not add a normalized composite index to its underlying physical or monetary components.

Imported conflicts and explicit repairs

The independent cross-program audit identified ten actual or status-granularity conflicts. The synthesis disposition is explicit so that no local notation is silently inherited.

Symbol discipline

Several letters recur across economics and engineering. The crosswalk avoids a false unification by qualifying them.

• $E_b^e$ is an energy quantity under boundary $b$ and convention $e$. It is not expected value, for which the paper uses $\mathbb{E}$.

• $U^{\mathrm{work}}$ is useful physical work. $W(\cdot)$ is a declared economic or welfare functional. They are not interchangeable.

• $S_{\mathrm{task}}$ is verified task service. $S_d$ is a decision-maker-specific scalarization. Neither denotes GDP.

• $\mathcal A(\mathcal P^{\mathrm{settle}},\omega)$ is an assurance vector. $A^{\mathrm{adopt}}$ is adopted use. Neither is an empirical treatment assignment, for which the namespace uses $Z^{\mathrm{assign}}$.

• $Pr_0^{\mathrm{reorg}},Pr_1^{\mathrm{reorg}}$ in the equation are matched scenario risks. $\mathcal R_i^{\mathrm{risk}}(x)$ in the equation is a general risk or external-cost functional.

• $\lambda_k^{\mathrm{KKT}}$ is a constraint multiplier and $p_{n,t}^{\mathrm{nodal}}$ a declared nodal price. Neither is an intrinsic value of energy at all locations or dates.

The software function canonical_equation_crosswalk() fixes Papers 1 through 9 in order, assigns each an evidence label, and validates uniqueness and completeness. That executable list is a regression guard against losing a paper or silently changing its role.

Boundary-reversal counterexample

Let systems $a$ and $b$ produce $100$ and $120$ matched service units. Their device energy is $10$ J and $20$ J, so the device-boundary intensities are $10$ and $6$ service units per J and $a$ ranks first. Suppose the matched facility boundary adds $40$ J of non-overlapping cooling and shared infrastructure to $a$ and none to $b$. Facility intensities are then $2$ and $6$, and $b$ ranks first. Neither ranking is an error. They answer different boundary questions. Comparing one device ratio directly with the other facility ratio would fail the manifest gate.

Why avoided loss does not solve every incompatibility

Analysts sometimes monetize every outcome as avoided loss or willingness to pay. That can support a decision model when exposure, counterfactual risk, standing, prices, distribution, and uncertainty are defensible. It does not produce a natural scalar. A service may protect rights or institutional features that are deliberately not traded off at a market price. The same loss event may already enter insurance payments, business interruption, asset revaluation, and welfare estimates. Monetization therefore creates additional overlap and identification duties.

Why exergy does not solve every incompatibility

Exergy offers a common physical metric for the maximum useful work obtainable relative to an environment. It can improve the description of the energy-to-work portion of a transduction chain. It cannot determine the correctness of an answer, the social usefulness of a task, a protocol’s adversarial robustness, or a counterfactual welfare effect. The synthesis uses exergy as a typed physical node rather than a universal economic numerator.

Why price does not solve every incompatibility

Market prices encode marginal exchange under particular institutions, property rights, distributions, and market structures. They can value traded flows in a specified accounting context. They may omit externalities, nonmarket production, consumer surplus, distributional standing, tail risks, and rights constraints. A price-weighted scalar is legitimate when its purpose and limitations are stated. Calling it the Joule Standard without qualification would conceal the valuation choice.

Independent national closures

A small national fixture closes five ledgers independently.

1. Product energy. Primary supply is $1000$ J. Its disposition is $200$ J of direct final primary use, $100$ J of primary exports, $400$ J of conversion loss, $250$ J of final converted energy, and $50$ J of converted-energy exports. These disjoint dispositions sum to $1000$ J. The $300$ J conversion output is displayed as a stage, not added to the primary total.

2. Component lineage. A facility meter carries lineage op-2026-q3 and reports $100$ J. A server-manufacture allocation carries lineage server-manufacture and reports a $20$ J embodied increment. Their scoped $120$ J total is permitted because the lineages do not overlap. The fixture does not also add device and cooling components already inside the facility meter.

3. Production. GVA entries of $400$ and $300$ currency units, product taxes of $40$, and subsidies of $10$ give GDP of $730$.

4. Outcome chain. The memorandum reports $100$ produced, $70$ adopted, and $50$ realized task units. The quantities reconcile as nested stages and are not summed to $220$.

5. Settlement cash. Load payments of $1000$, generator credits of $850$, and network or constraint rent $K=150$ satisfy $Payment_D-Credit_G=K$. These transfers are not added as three welfare benefits.

Passing one closure does not repair another. A balanced cash settlement can coexist with a duplicated energy lineage, and a balanced physical account can coexist with a repeated production flow.

Executable objects

The module provides:

• QuantityType and TypedQuantity, with exact semantic, unit, accounting-kind, boundary, and horizon checks;

• BoundaryManifest, with a field-by-field direct-comparability gate;

• TransductionStage, typed exact path composition, retained stage statuses, and the conservative evidence join in the equation;

• EnergyInventory, physical-lineage identifiers, energy-stage controls, and separate operational plus embodied increments;

• source-tagged account entries and a minimum double-counting audit;

• SystemVector, direction-aware dominance, and deterministic Pareto filtering;

• published normalization rules and explicit scalarization weights;

• matched causal effect ratios and assurance contrasts;

• direct and finite rebound identities; and

• a canonical equation crosswalk for Papers 1 through 9.

Refusal behavior

Refusal is part of the implementation. The module raises a typed error when an operation would:

• add a money stock to a money flow;

• add tokens to verified task service;

• compose stages whose interface semantics or amounts differ;

• compare manifests with unmatched energy stages or other required fields;

• repeat a component identifier or physical lineage;

• total primary and final energy as independent stages;

• compare Pareto vectors with different axes or units;

• accept scalarization weights that do not sum exactly to one;

• form a causal ratio from unmatched interventions or a zero energy effect; or

• accept a probability outside $[0,1]$.

These rules do not decide contested empirical questions. They prevent invalid operations after a researcher has supplied the relevant records.

Adversarial test contract

The dedicated test file contains twenty-four tests. It verifies a valid energy-to-computation-to-service-to-value path, then substitutes an identically displayed but semantically different count at an interface. It attacks stock-flow addition, task-token addition, boundary matching, duplicate energy components, primary-final mixing, repeated account sources, GVA-memorandum overlap, outcome-axis substitution, invalid weight sums, zero-denominator causal ratios, invalid assurance risks, crosswalk omissions, crosswalk misordering, and binary floats. It also checks positive cases: separate operational and embodied increments, Pareto removal of a dominated point, conditional rank reversal, exact avoided loss, and the rebound identities.

The test suite is evidence for program behavior, not evidence that any real-world parameter has been measured correctly.

Boundary falsification

A comparison fails its direct-comparability claim if a line-item audit reveals different functional units, time intervals, locations, energy stages, failure treatment, embodied scope, counterfactuals, outcome functionals, attribution rules, uncertainty models, or exclusions. Researchers should publish manifests before examining rankings. A blinded boundary audit can then test whether the claimed comparison survives.

Facility energy should be reconciled against device telemetry, power distribution data, and site meters. Discrepancies need an allocation bridge, not a generic overhead factor. Lifecycle claims should publish inventory processes and show that operational energy is not added after already entering the lifecycle total.

Transduction falsification

Each edge needs an observable or model that could contradict it. Energy-to-work edges can be checked by calibrated physical measurement. Computation-to-service edges can be tested on held-out tasks with blinded scoring and failure retention. Service-to-adoption edges can be tested using workflow logs. Adoption-to-value edges require randomized assignment, a credible quasi-experiment, or explicit partial-identification bounds.

A transduction claim fails if the output population differs from the next stage’s input population, if attrition is silently removed, or if a semantic substitution occurs. A product of average stage yields should be compared with the average pathwise product. Their divergence is evidence against an independence simplification.

Quality-adjusted intelligence

Quality scores should be challenged by:

• tokenizer substitutions that preserve user-visible content;

• verbosity perturbations that add tokens without adding useful content;

• adversarial task-mix shifts;

• calibration and abstention tests;

• repeated trials that expose reliability and tail failures;

• latency thresholds tied to task usefulness; and

• human or system baselines evaluated under the same functional unit.

A claimed $\operatorname{QAIJ}$ advantage is falsified for a target population when it disappears under the prespecified task distribution or fails the boundary gate. A robust result should report a frontier over quality, latency, reliability, and energy rather than one chosen score alone.

Settlement assurance

An assurance claim should vary adversary resources, network delay, hardware availability, concentration, bribery, reward changes, censorship objectives, confirmation horizon, and asset exposure. Empirical event studies can examine reorganizations, fee shocks, hashrate migration, outages, and concentration, but observed survival alone does not identify counterfactual loss prevention.

The claim “more energy caused more assurance” is falsified if a matched intervention changes energy without improving the prespecified risk measure, or if the improvement is explained by another protocol or market change. Equal energy systems should be used as negative controls against any energy-only security mapping.

Causal realization

For AI and automation, prespecify numerator and denominator effects. Report first-stage energy effects, outcome effects, compliance, spillovers, interference, and missingness. If the energy effect is near zero, report a ratio-aware confidence set rather than a conventional bounded interval. Compare results with and without low-value, failed, repeated, and abandoned inferences. A value-per-joule estimate that improves only by deleting failures does not represent the deployment.

Marginal allocation and rebound

Allocation models should be back-tested against realized nodal prices, congestion, curtailment, deadlines, quality, and service delivery. Shadow-price interpretations fail when constraints or nonconvexities omitted from the model bind in operation. Flexible-load claims require measurement of response speed, recovery costs, and foregone service.

Rebound estimation needs a credible source of efficiency variation and a stable service definition. Researchers should report direct service demand, prices, complementary inputs, displaced work, and economy-wide effects over multiple horizons. A decline in device joules per task does not by itself test $\epsilon_{Q^{\mathrm{cog}},\eta_Q}$.

National-account audit

Statistical compilers should trace every monetary entry to a source transaction or imputation and every physical entry to a lineage. The following adversarial questions should be answered:

1. Does an AI outcome memorandum repeat sales or output already included in GVA?

2. Is a protected asset position being added to a production flow?

3. Does avoided loss overlap an insurance payment, revaluation, or service fee already recorded?

4. Are primary energy, conversion output, and final use summed?

5. Is embodied energy already contained in a lifecycle total?

6. Do imports, residence, and time-of-recording rules match the monetary core?

7. Does a composite ranking survive plausible normalizations and weights?