Versioned reconstructive design contract — 20 Watts Was Enough

lusoris contributors

This note defines the measurement boundary for Fixture F-002. The fixture treats proposal generation as reconstruction from declared exposure and retained history. Novelty is therefore relative to a frozen reference history, while correctness and usefulness are determined by constraints, tests, and qualified evaluators.

Histories, sources, and versions

At step $t$ , distinguish three source sets:

$S^{\mathrm{exp}}_t$ : items exposed before the task, whether or not the system can retrieve them;
$S^{\mathrm{ret}}_t$ : items returned by a metered retrieval operation during the task; and
$S^{\mathrm{att}}_t$ : items attributed as contributing to the current artifact.

Each set contains immutable source identifiers and versions. The declared history against which novelty is scored is

H_t=S^{\mathrm{exp}}_t\cup S^{\mathrm{ret}}_t\cup V_{<t},

where $V_{<t}$ is the set of artifact versions created before step $t$ . Exposure, retrieval, and attribution are not interchangeable: an exposed item may not be retrieved, a retrieved item may not be used, and a used item may be omitted from attribution.

For artifact version $v_k$ , retain

v_k=(P_k,o_k,u_k,S^{\mathrm{att}}_k,q_k,\tau_k),

where $P_k$ is the set of parent-version identifiers, $o_k$ is the typed operation, $u_k$ is the actor or process identifier, $S^{\mathrm{att}}_k$ is the attributed source set, $q_k$ is the test and evaluation record, and $\tau_k$ is the timestamp in seconds from the run origin. This tuple records a derivation claim; it does not establish correctness, usefulness, originality, authenticity, or intent.

Reconstructive proposal generation

Let $z_{i,t}$ be proposal $i$ at step $t$ , $c_t$ the visible task constraints, $r_t$ the current external representation, and $x_t$ the method's internal state. A generator samples or searches

z_{i,t}\sim p_m\!\left(z\mid x_t,r_t,c_t, S^{\mathrm{exp}}_t,S^{\mathrm{ret}}_t,V_{<t}\right),

where $m$ identifies the method. The conditional form makes no claim that the mechanism is stochastic: deterministic retrieval, CAD transformation, search, and constraint solving are valid methods. Source-removal, source-replacement, and history-scrambling interventions estimate which prior material actually changes the proposal distribution.

For a frozen feature map $\phi$ , the relative novelty of proposal $z$ is

N_H(z)=\min_{h\in H}\delta\!\left(\phi(z),\phi(h)\right),

where $\delta$ is a preregistered dimensionless distance and $H$ is the declared history. Report $N_H$ under task-native features and at least one independent representation. Changing $H$ , $\phi$ , or $\delta$ changes the claim; $N_H$ is not intrinsic novelty.

Externalization and representation change

An action $a_t$ and observed material or environment response $e_t$ update the external representation by

r_{t+1}=F_{\rho_t}(r_t,a_t,e_t),

where $\rho_t$ identifies the representation regime, such as raster, vector, scene graph, CAD, text, simulation state, or physical prototype. A change of representation is a typed operator

r'_{t}=T_{\rho_t\rightarrow\rho'_t}(r_t),

whose measured losses include constraint violations, geometric error in millimetres or pixels, missing relations as a count, and lost provenance edges as a count. Undo, branching, and conversion time are recorded in seconds.

A claimed reinterpretation event must identify a relation found during inspection that was absent from the immediately preceding registered plan. If $K_t$ is the coded relation set before inspection and $K_{t+1}$ after it, then

I_t=\left|K_{t+1}\setminus K_t\right|\quad[\mathrm{relation}],

with a published coding protocol and inter-rater reliability. Edit count is not a substitute for reinterpretation.

Epistemic action and material feedback

Let $\Theta$ be a hidden task, geometry, material, user, or failure variable and $b_t$ the current belief state. For an epistemic action $a$ , expected information gain is

\operatorname{EIG}(a\mid b_t)= H(\Theta\mid b_t)- \mathbb E_{y\sim p(y\mid a,b_t)}H(\Theta\mid b_t,a,y) \quad[\mathrm{bit}],

where $y$ is the resulting observation and $H$ is Shannon entropy in bits. The evaluator estimates realized information gain only against the hidden ground truth; the method cannot read that trace.

For a physical or simulated probe, record the cost vector

\mathbf c_a=(t_a,E_a,M_a,W_a,R_a),

where $t_a$ is elapsed time in seconds, $E_a$ energy in joules, $M_a$ consumed material in kilograms, $W_a$ waste in kilograms, and $R_a$ expected harm in a declared task-native unit. Compare information and decision value at matched $\mathbf c_a$ ; motion, rendering, or fabrication alone does not imply an epistemic gain.

Let hidden material state $\mu$ produce feedback

y^{\mathrm{mat}}_t=g(z_t,a_t,\mu)+\epsilon_t,

where $y^{\mathrm{mat}}_t$ has declared physical units, $g$ is the simulator or physical response process, and $\epsilon_t$ is measurement error in the same units. Confirmatory tests withhold materials, simulators, constitutive regimes, and failure modes so that replayed feedback cannot pass as adaptation.

Diversity, fixation, and negative transfer

For valid proposal set $Z=\{z_1,\ldots,z_n\}$ , report pairwise distances and feature-space coverage. One summary is

D(Z)=\frac{2}{n(n-1)} \sum_{1\le i<j\le n}\delta\!\left(\phi(z_i),\phi(z_j)\right),

where $D$ is dimensionless and $n\ge2$ is a proposal count. Also report the number of distinct valid constraint-satisfying regions reached. Proposal count is not diversity.

For exposure condition $e$ and matched no-example condition $0$ , define fixation toward exemplar $s$ as

F_e=\mathbb E\!\left[ \delta(\phi(z_0),\phi(s))- \delta(\phi(z_e),\phi(s)) \right].

A positive dimensionless $F_e$ indicates movement toward the exemplar. It is not automatically harmful. Negative transfer on outcome component $j$ is

T^-_{e,j}=Y_{0,j}-Y_{e,j},

after orienting $Y_j$ so that larger is better. Report both $F_e$ and $T^-_{e,j}$ : copying can preserve a useful relation, transform a precedent, or fixate on an invalid one.

Constraint validity, evaluation, and selection

For proposal $z_i$ , let $g_k(z_i)\le0$ be hidden or visible constraint $k$ in its native unit. Constraint validity is the binary value

V_i=\mathbb 1\!\left[g_k(z_i)\le0\ \text{for every required }k\right].

Also report each margin $-g_k(z_i)$ separately; a valid/invalid bit must not hide near misses or catastrophic failures.

The primary outcome remains a vector

\mathbf Y_i= (V_i,N_{H}(z_i),D_i,U_i,A_{i,o,c},R_i, L_i,E_i,M_i,W_i,P_i),

where $V_i$ is binary constraint validity, $N_H$ and diversity contribution $D_i$ are dimensionless, $U_i$ is task utility in a declared native unit, $A_{i,o,c}$ is observer- $o$ and context- $c$ qualified evaluation on a declared scale, $R_i$ is risk in a task-native unit, latency $L_i$ is seconds, energy $E_i$ is joules, material $M_i$ and waste $W_i$ are kilograms, and provenance coverage $P_i$ is dimensionless. Do not collapse this vector into a universal creativity score.

If a declared utility $U^{*}$ is necessary for selection, publish its weights, normalization, and sensitivity analysis. Selection regret is

R_{\mathrm{sel}}= \max_{z\in Z}U^{*}(z)-U^{*}(z^{\mathrm{chosen}}),

in the same unit as $U^{*}$ . Evaluate regret with a blinded frozen evaluator and then report realized post-selection outcome separately. A generator's own score cannot serve as independent selection evidence.

Attribution and retained lineage

Let $E^{\mathrm{true}}$ be source-to-version and parent-to-child edges known to the benchmark generator, and $E^{\mathrm{rec}}$ the submitted lineage edges. Lineage precision and recall are

P_{\mathrm{lin}}= \frac{|E^{\mathrm{rec}}\cap E^{\mathrm{true}}|}{|E^{\mathrm{rec}}|}, \qquad R_{\mathrm{lin}}= \frac{|E^{\mathrm{rec}}\cap E^{\mathrm{true}}|}{|E^{\mathrm{true}}|}.

Both are dimensionless. Empty submissions receive zero precision and recall. Source attribution is also scored against benchmark-known influence interventions; mere string overlap is insufficient.

After delay, tool replacement, or actor turnover, reconstructability of target version $v$ is

Q_{\mathrm{recon}}(v)= \frac{1}{|\mathcal Q|} \sum_{q\in\mathcal Q} \mathbb 1\!\left[d_q(\widehat v,v)\le\varepsilon_q\right],

where $\mathcal Q$ is a preregistered query set, $d_q$ has the native unit of query $q$ , $\varepsilon_q$ is its tolerance in that unit, and $\widehat v$ is the reconstructed version. Retaining pixels without source, constraint, test, and operation lineage can therefore fail reconstruction.

Lifecycle and equal-budget boundary

For method $m$ , charge

\mathbf B_m=(N_{\mathrm{src}},N_{\mathrm{ret}},N_{\mathrm{prop}}, N_{\mathrm{eval}},N_{\mathrm{probe}},t_{\mathrm{human}},t_{\mathrm{wall}}, B_{\mathrm{state}},E_{\mathrm{life}},M_{\mathrm{material}},W_{\mathrm{waste}}),

where the first five terms are counts, both time terms are seconds, $B_{\mathrm{state}}$ is bytes, $E_{\mathrm{life}}$ is joules, and material and waste are kilograms. Lifecycle energy is

E_{\mathrm{life}}=E_{\mathrm{train}}+E_{\mathrm{index}}+ E_{\mathrm{retrieve}}+E_{\mathrm{generate}}+E_{\mathrm{inspect}}+ E_{\mathrm{simulate}}+E_{\mathrm{fabricate}}+E_{\mathrm{evaluate}}+ E_{\mathrm{retain}}+E_{\mathrm{recover}},

with every term in joules at one declared boundary. Human preparation, critique, physical facilities, and failed prototypes are reported even when they cannot be converted credibly to energy.

Equal-budget comparison either holds every preregistered binding component of $\mathbf B_m$ within tolerance or compares methods on a Pareto frontier. It must not divide an over-budget result by cost after the run and call the arm matched.

Estimands, ablations, and retirement

For outcome $Y_j$ , paired treatment effect of component $c$ is

\Delta_{c,j}=Y_j(m_{\mathrm{full}})-Y_j(m_{-c}),

where $m_{-c}$ removes only component $c$ without reallocating its budget. Use paired hidden instances and report 95% uncertainty intervals across problem, source, material, evaluator, and seed strata.

The composed residual is retired when mature nulls match its preregistered constraint validity, selection regret, transfer, lineage, and lifecycle-cost targets; when its advantage exists only for seen histories, materials, or evaluators; when any ablation is non-diagnostic; or when unlogged exposure, retrieval, evaluator access, lineage, labor, material, or energy can explain the result.

Editable system diagram: versioned-reconstructive-design.mmd.