Numerical consequences of representation and reduction

Page status: generated structural, linearized, symbolic KKT, and package-level solver-diagnostics crosswalk witnesses; solver-internal nonlinear exports remain future work.

A graph choice is also a numerical choice

Learning question. What can change numerically when the declared electrical relation is preserved? Follow three comparisons: physical residuals after scaling, dependency/fill after compilation, and coverage before a verdict.

Two representations can describe the same declared electrical behaviour while producing very different numerical problems. A quotient may reduce the number of variables but worsen scaling; a Kron reduction may preserve a boundary relation but introduce dense couplings; a factor model may expose physical structure while a bus–branch compiler produces a smaller, more familiar Jacobian. These are computational consequences, not afterthoughts.

The relevant comparison is therefore not only

\[\text{same terminal map?}\]

but also

\[\text{same feasible observations, with what conditioning, sparsity, and recovery cost?}\]

The statements below concern a declared operating point, coordinate system, and solver tolerance. They do not claim that one representation is uniformly better.

The federated PSK-000011 guard makes one practical boundary executable: reference_singularity_validation compares mapped connected-island reference incidence and rank declarations before accepting a transformed model. A successful import or solver termination is not an independent certificate of full rank. The package pass remains declaration-level evidence; equation, Jacobian, conditioning, and solver-specific checks still belong to the study.

Transfer task. Choose one proposed reduction. Name the electrical relation it preserves and one numerical quantity that still requires a separate check.

Scaling and conditioning

Learning question. Which error does a small scaled residual actually bound?

Let a linearized or linear subproblem be

\[\mathbf A x = b.\]

Changing units or terminal coordinates is represented by nonsingular maps $x=\mathbf D_x\widetilde x$ and $b=\mathbf D_b\widetilde b$. The coefficient matrix becomes

\[\widetilde{\mathbf A} = \mathbf D_b^{-1}\mathbf A\mathbf D_x.\]

The coordinate map bijectively relates the two solution sets. For nonsingular $\mathbf A$ and an induced matrix norm, the condition number can change:

\[\kappa(\widetilde{\mathbf A}) = \|\mathbf D_b^{-1}\mathbf A\mathbf D_x\| \|\mathbf D_x^{-1}\mathbf A^{-1}\mathbf D_b\|.\]

This is why per-unit coordinates are a numerical convention as well as a reporting convention. A declared base scales voltage, current, power, and impedance together; it does not erase poor physical conditioning, singular topology, or an omitted conductor. A coordinate transformation is safe to call an exact rewrite only when its inverse and its power-dual action are retained, as in the coordinate-transformation chapter.

For a computed solution $\widehat x$ define the residual $r=b-\mathbf A\widehat x$. A backward-error measure is, for example,

\[\eta(\widehat x) = \frac{\|r\|}{\|\mathbf A\|\,\|\widehat x\|+\|b\|}.\]

For nonsingular $\mathbf A$, nonzero exact solution $x$, and the same vector norm and its induced matrix norm throughout, put $\theta=\kappa(\mathbf A)\eta(\widehat x)$. If $\theta<1$, an explicit bound is

\[\frac{\|\widehat x-x\|}{\|x\|} \le \frac{2\theta}{1-\theta}.\]

To see the factor of two, use $\widehat x-x=-\mathbf A^{-1}r$ and $\|b\|\le\|\mathbf A\|\|x\|$. The relative error $\delta$ then satisfies $\delta\le\theta(2+\delta)$. This bound concerns the stated linear system; it does not certify a nonlinear solution or the physical input data. The LAPACK error-analysis discussion also distinguishes normwise and componentwise backward error. Record both residual/backward-error information and conditioning, with their norms and scaling, rather than relying only on solver termination.

Transfer task. Rescale one state coordinate by ten. State the inverse map, the norm and the physical residual tolerance you would retain before comparing the two numerical residuals. Explain why equal coordinate solution sets do not imply equal condition numbers.

Choose a model against a constraint margin

Learning question. Is the modelling error small relative to the decision margin, rather than merely small relative to the reported voltage?

The parallel-member lesson derives an exact aggregate for fixed data. Now suppose the total nonnegative current $J$ is prescribed, both members are rated 100 A, the second conductance is 1 S, and the first is known only to lie in $[8,12]$ S. The nominal first conductance is 10 S. These are declared interval assumptions for a resistive teaching case, not measured equipment uncertainty or a probability model.

For conductance $g$ expressed numerically in siemens, current division is

\[I_1(J,g)=\frac{g}{g+1}J,\qquad I_2(J,g)=\frac{1}{g+1}J.\]

Compare three choices: solve the two nominal member relations; use their exact nominal aggregate cap of 110 A with recovery; or require both member limits for every conductance in the interval. The first two describe the same nominal feasible set. Neither settles the uncertain-data question.

Because $I_1$ increases with $g$ and $I_2$ decreases, their worst values occur at opposite interval endpoints. Thus the exact robust cap in this local model is

\[J\le\min\{100(13/12),100(9)\}=325/3\ \mathrm A.\]

At 99 A, the nominal first-member current is 90 A; the worst member current is $1188/13$ A, leaving $112/13$ A of margin. At 109 A, the nominal model still accepts, but the worst member current is $1308/13$ A and exceeds its rating by $8/13$ A. The same uncertainty that was harmless for the first decision changes the second decision's feasibility classification.

A precise nominal computation cannot remove an uncertain input. Compare the possible observation error with the source constraint margin. Here the nominal aggregate is defensible for a stated nominal question; the interval requirement needs the robust cap. This is a proof over the full declared interval using monotonicity, not a claim based only on sampled scenarios. Other coupled or nonlinear models need their own worst-case argument.

python3 experiments/lessons/model_choice.py
python3 experiments/lessons/model_choice.py --check
python3 experiments/lessons/model_choice.py --benchmark

The first two commands use rational arithmetic. The optional local timing comparison uses floating-point arithmetic for nominal explicit, nominal aggregate, and interval-robust evaluations. Each measured call includes its cap/current calculation, limit checks, and current recovery; imports and I/O are excluded. It reports the median of seven batches of 10,000 calls. This is a tiny algebraic microbenchmark with no PF solver: use it to learn to count recovery work, not to rank network formulations. Timings are printed for the actual machine and are not stored as universal performance evidence.

Transfer exercise. Narrow the interval to $[9,11]$ S. Derive the new robust cap and decide whether 109 A is admissible. The cap is $1200/11$ A, so it is admissible under this interval. State which changed assumption made that conclusion possible. Do not report it as evidence for the original $[8,12]$ S interval.

Jacobians: physical topology is not matrix sparsity

Learning question. Does a new matrix edge denote new equipment, a shared equation dependency, or elimination fill?

For a nonlinear decision model $F(y,u)=0$, Newton or interior-point methods use a Jacobian such as

\[\mathbf J = \frac{\partial F}{\partial y}.\]

The physical graph predicts some nonzeros, but the Jacobian graph is a graph of variable–equation dependence. A single multi-terminal factor can create a dense block between all of its ports. A nominal-$\pi$ element can couple endpoint voltages through series terms and endpoint currents through shunts. Constraint rows for limits, controls, and recovery variables add edges that have no corresponding physical line. Conversely, a physical asset may be absent from a particular Jacobian block when its state is fixed or its equation has been eliminated.

Write $G_phys$ for a chosen physical incidence graph and $G_J$ for the bipartite Jacobian graph. In general there is no equality $G_phys=G_J$. The useful statement is a declared dependency map

\[\delta:\; (\text{factor},\text{terminal},\text{constraint}) \longmapsto (\text{equation rows},\text{variables}),\]

from which $G_J$ is generated. This map is the numerical analogue of source provenance: it explains why a fill-in or a dense block exists.

The five-bus structural witness makes the distinction concrete. Its source has seven line identities but only six edges after simple projection because $q$ and $r$ are parallel. Eliminating $j$ then adds the $i$–$l$ fill edge to the projected pattern. The fill argument and the dependency argument are shown separately so that a reader does not confuse a Schur-complement edge with a numerical derivative. The conceptual pair now appears in the formulation and lowering chapter; this chapter retains the executed witnesses and their fixture-specific counts.

The data and checks are recorded in experiments/generated/numerical-structure-witness.json; the renderer is experiments/generate_numerical_structure_views.py. This separation is intentional: both structural views can be checked without pretending that either is a solver-exported Jacobian.

Fill is not inevitable. The two-level topology chapter proves a positive structural case: dense two-terminal conductor stamps over a bus-level tree form a chordal scalar-support graph, and eliminating leaf-bus coordinate blocks inward is a perfect elimination order with zero symbolic fill. The warning here concerns arbitrary orderings, meshes, reductions, and factor libraries; it should not hide structure that a declared model makes exploitable.

The numerical witness now carries a crosswalk to experiments/generated/five-bus-typed-kron-witness.json. That crosswalk uses the non-pendant $\ell$ elimination, whose fill edges are $j-m$ and $k-m$, and records the small boundary residual together with the recovered $u$-branch limit observation. This ties the Schur-complement example to the numerical-structure discussion without conflating fill edges with physical assets or solver-private factorization output.

A pinned numerical export

The running fixture provides a second, numerical witness through BMOPFTools' passive and constant-$Z$ linearized $Y$-bus exporters. In the exporter's node order, the passive matrix has 20 rows and 166 nonzeros. The constant-$Z$ linearized matrix agrees with that passive matrix for this fixture. Realifying the complex current relation gives a 40-by-40 real matrix with 664 nonzeros. This is a coordinate embedding, not a promise that every complex invariant is preserved: the complex Ybus is transpose-symmetric to numerical precision, whereas the realified current-Jacobian embedding has a large transpose residual. That residual is expected for a nonzero imaginary part: the block matrix $\mathcal R(Y)$ with diagonal blocks $\Re Y$ and off-diagonal blocks $-\Im Y$ and $\Im Y$ is generally not symmetric. This is not a round-off failure. Realification preserves the coordinate support and the current relation, but symmetry claims must be stated in the representation in which they are checked (the witness residual is approximately $448.79$):

\[\begin{bmatrix}\Re I\\ \Im I\end{bmatrix} = \begin{bmatrix} \Re Y&-\Im Y\\ \Im Y& \Re Y \end{bmatrix} \begin{bmatrix}\Re V\\ \Im V\end{bmatrix}.\]

The ordinary 2-norm condition estimates are approximately $2.49\times10^{17}$ (unscaled) and $2.36\times10^{16}$ (after simple row/column equilibration). Because both matrices are numerically rank deficient (rank 18 of 20 at the witness tolerance), those finite values should not be read as meaningful condition numbers: they are dominated by singular values below the modelling tolerance. The witness therefore also reports the rank-aware effective estimates $\kappa_{\mathrm{eff}}=\sigma_1/\sigma_{18}$ and its equilibrated counterpart. These are diagnostic, not universal constants: they depend on units, scaling, norm, rank tolerance, and the chosen fixture. Pure row and column permutations preserve singular values, so they preserve the exact 2-norm condition number and a ratio of the same selected singular values. Ordering can change sparse-factorization fill, pivoting behavior, runtime, and floating-point estimation details; these are algorithmic effects, not a change in the exact spectral condition number. For this export the effective estimates are approximately $6.50\times10^8$ and $1.13\times10^7$ after equilibration. A solver must account for reference and grounding structure rather than treating any of these numbers as a graph invariant.

The complete export, including node order, nonzero entries, checks, and the linearization convention, is recorded in experiments/generated/ybus-jacobian-witness.json. Its generator and renderer are:

experiments/run_ybus_jacobian_witness.jl
experiments/render_ybus_jacobian_view.py

This is a pinned linear current/Jacobian witness; it is not yet a nonlinear OPF KKT Jacobian or an independent-solver comparison.

Nonlinear decision Jacobian and KKT fill

The next witness adds the nonlinear term that makes the parallel-line warning decision relevant. For a fixed complex endpoint voltage $U$ and member currents $I_1,I_2$, the residual includes

\[S(U,I_1,I_2)-\alpha S_0 = U(I_1+I_2)^*-\alpha S_0.\]

The source formulation retains two explicit current laws and two current vectors. The aggregate formulation replaces them by one summed current law. At the recorded operating point both residuals are exactly zero, but their finite-difference Jacobians and KKT patterns differ:

For avoidance of a common counting error, the source witness has two complex member-current coordinates, represented as four real variables, in addition to the three retained real coordinates. It therefore has seven real source variables; the aggregate witness has five real variables. The reduction removes two real variables (one complex current), not eight. The source KKT system has 13 rows and columns because it adds six residual/constraint multipliers; that KKT dimension is a different count from the number of primal variables.

The fill counts are symbolic graph counts, not floating-point factorization times. They show why an ordering is part of a numerical certificate: the same KKT graph has different fill under different elimination orders. The source and aggregate models also have different graphs, so a smaller KKT system does not by itself establish decision equivalence.

The complete finite-difference and symbolic-fill artifact is experiments/generated/nonlinear-kkt-witness.json, regenerated by experiments/run_nonlinear_kkt_witness.jl and experiments/render_nonlinear_kkt_view.py. It is deliberately labelled a nonlinear decision witness rather than a solver-internal Ipopt KKT export. A package-level crosswalk now binds it to the BMOPFTools Ybus witness, while actual solver-internal factorization diagnostics remain a separate boundary.

Package-level diagnostics crosswalk

The generated solver-diagnostics-crosswalk.json composes the two witnesses without pretending that either is a production solver export. It retains the physical node/terminal order, the BMOPFTools passive and constant-$Z$ Ybus, the realified current Jacobian, the finite-difference nonlinear residual Jacobian, and the symbolic KKT graph under two declared orders. In the source formulation, for example, the natural and constraints-first orders produce different fill counts, so the crosswalk records ordering as part of the diagnostic identity.

The crosswalk also exercises the public BMOPFTools.opf_checked_kkt_factorization callback on a staged OPF context. The probe accepts a regular matrix and rejects a near-singular one. This is a checked callback boundary. A minimal parameterized OPF now passes that callback through DiffOpt: the forward sensitivity agrees with a central finite difference to the recorded tolerance, and the KKT diagnostic is accepted. The adapter also captures the matrix passed to the callback, recording its dimension and nonzero count, alongside the JuMP variable and constraint order used to account for its rows and columns. Four additional callback rows remain outside that declared JuMP block, so they are retained as an explicit DiffOpt/solver-internal boundary rather than silently assigned a physical label. The crosswalk still records solver_internal_kkt_export = false because solver-provided row labels, scaling, linear-solver choice, pivot or inertia data, and factorization statistics are not yet exported as a source/target comparison.

The same witness records BMOPFTools' differentiability report: inequality counts, active, near-active, weakly-active, and violated labels, minimum inactive slack, and qualifications. In this deliberately unconstrained fixture the inequality count is zero. The report's ready flag is state provenance for the selected local solve, not a proof of LICQ, strict complementarity, second-order sufficiency, global optimality, or branch stability.

The crosswalk now also builds a regular JuMP mirror and records its native affine/quadratic variable-support count alongside a separate JuMP.NLPEvaluator view. This gives an executable constraint-row/nonzero summary for the model-level derivative graph (23 supported variable entries in the current 19-row fixture), while leaving native_nlp_export_is_solver_internal = false: it is not Ipopt's private KKT ordering, pivoting, inertia, or factorization export.

For a deliberately narrow parallel-line fixture, the crosswalk also compares a two-member source against a single 0.25 Ω equivalent member. The tested voltage sensitivity is preserved, while the captured KKT matrix grows for the explicit source. The native JuMP support count changes as well, so the distinction is visible before any solver-private factorization. This is evidence that an exact electrical scalar equivalent can still change solver structure; it is not a general AC or multiconductor reduction certificate.

Numerical evidence boundary

These diagnostic artifacts answer different questions and should not be collapsed into one solver claim:

An actual solver diagnostic would need to bind the exported derivative rows and columns to the source/target variable order, record scaling and tolerances, identify the linear solver and ordering, and retain the factorization or inertia report. Until that export exists, the chapter's KKT language is a symbolic comparison, not an implementation claim about a production solver.

Transfer task. Label one physical edge, one Jacobian dependency and one fill edge in the five-bus witness. For each, name its generating map. Identify which reported KKT quantity is symbolic and which comes from an actual callback; do not assign unlabelled solver rows a physical identity.

Elimination and fill-in

Learning question. Which costs can change when an internal variable is eliminated exactly?

Partition a sparse matrix by retained variables $B$ and eliminated variables $I$. The Schur complement is

\[\mathbf S = \mathbf A_{BB} -\mathbf A_{BI}\mathbf A_{II}^{-1}\mathbf A_{IB}.\]

Even when $\mathbf A$ is sparse, $\mathbf S$ can be dense on the neighbours of $I$. In the symbolic graph, eliminating a vertex connects its remaining neighbours into a clique. The extra nonzeros are fill-in. The same mechanism appears in Kron reduction: preserving a boundary relation does not preserve sparsity, asset identity, or the cost of a subsequent solve.

The order of elimination matters. For a fixed matrix, two orders can have the same exact Schur complement but different intermediate fill, memory use, and factorization time. A certificate that reports a reduction should therefore record at least:

These fields are properties of a compiled numerical problem, not of a simple graph in isolation.

Transfer task. Eliminate a leaf first and an internal junction first in the same declared support graph. Identify the added symbolic edges and the recovery work retained in each case. State which runtime or conditioning conclusions still require a measured numerical experiment.

A coverage table before a feasibility verdict

Learning question. What must be supplied and independently evaluated before a small residual can support the requested feasibility conclusion?

Declare the observations before calculating them. Keep a coverage row for each obligation, its required inputs, unit, tolerance and evidence class:

Predict. Two independent feeders each have a 10 V ideal source, a 1 Ω series resistor and a 1 S shunt to earth at the receiving terminal. In each feeder, the receiving voltage is 5 V and the line current is 5 A. If the two reported source currents change to 6 A and 4 A, does checking total power detect the corruption?

For either feeder, direct substitution gives

\[10-5=1\cdot5,\qquad I_{\mathrm{shunt}}=1\cdot5=5\ \mathrm A.\]

The source supplies 50 W; the line and shunt each absorb 25 W. With two feeders, the original source total is 100 W. The corrupted source currents still give $10\cdot6+10\cdot4=100$ W, so aggregate power agrees. But the signed KCL residuals at the two source terminals, taking outgoing current as positive, are $5-6=-1$ A and $5-4=1$ A. Each violates a $10^{-6}$ A acceptance tolerance. Updating the reported source powers to 60 W and 40 W does not repair those local equations.

This is a concrete use of claims NUMERICAL-004 and NUMERICAL-005: a supplied residual bundle and a solver status cannot replace recomputation with explicit coverage. The zero-voltage lesson gives a second failure mechanism: at an exactly grounded terminal a missing 1 A earth reaction can be invisible even to local power balance. A perfect ground fixes voltage; its reaction current must still be accounted for in a current audit.

What was evaluated, and what was reconstructed?

For an inline series-line model, an evaluator can read the authored matrix entries and check

\[r_V=V_f-V_t-ZI_f,\qquad r_{I,\mathrm{line}}=I_f+I_t, \qquad r_{I,b}=\sum_{\text{outgoing devices}}I-\sum_{\text{injections}}I.\]

These are distinct observations with distinct units and tolerances. The receiving terminal map determines where $I_t$ enters KCL, even when result rows use the sending conductor's label. A limit check must evaluate $|I_f|$, $|I_t|$ and any declared $|V_f\overline{I_f}|$ / $|V_t\overline{I_t}|$ bounds at their respective endpoints. Equality with a bound passes within the declared tolerance; a negative or malformed declared rating is not a zero limit violation. An absent optional rating means no such constraint was declared.

For a shunt with no independently supplied current, computing $I=YV$ constructs a contribution to KCL. Reporting $I-YV=0$ as an independent successful test would merely compare the calculation with itself. Label it reconstructed. A missing source current instead makes the affected audit unassessed; no residual has been measured. These categories should survive JSON serialization.

Package candidate and the running-network boundary

The package-owned audit_linear_solution candidate implements this small domain without calling the model builder, line_yprim, Ybus, or the result profiler. It shares the native SI input schema, result labels and arithmetic libraries. That is independent equation evaluation, not independent parsing, a second physical model, or measurement validation. The candidate's exact local revision is recorded in experiments/integration/linear-audit-candidate.json; it is not included in the published integration pin and is not exposed by MCP. The companion is experiments/lessons/verify_linear_coverage.jl, with execution instructions in the repository's experiments/integration/README.md.

The analytic tiny circuit and its native-coordinate solve pass; the cancelling-current mutation fails, and removing a required current returns an indeterminate result. The companion also records a solve using default per-unit scaling, evaluated at the same physical tolerances. In the local candidate run, that solver reported LOCALLY_SOLVED, but the two receiving-node KCL residuals were about $4.72\times10^{-4}$ A, exceeding the $10^{-6}$ A audit tolerance. The native-coordinate solve passed without relaxing the audit. This is a diagnostic observation for this tiny case and environment, not a claim that per-unit scaling is generally worse. The companion records the comparison without requiring every future solver run to fail the same way. On the running network, the older audit_linear_solution API produces the following bounded coverage. It is one view of the local candidate, not the published integration lesson or the broader audit_solution verdict. The workbook's evidence-generation table identifies all three environments and commands.

This linear audit contains 42 unassessed rows. That number counts obligations, not 42 missing devices or 42 independent experiments. The implementation conservatively suppresses every KCL verdict when any device current contribution is unassessed; local line checks remain available. A later implementation may propagate missing coverage by incidence, but must first demonstrate that its incidence inventory is complete. A failed evaluated check must also remain visible when other rows are unassessed.

Extending coverage without changing the older API

The separate candidate audit_solution adds default-role source neutral returns and signed per-phase source P/Q limits, implicit source grounding, WYE/SINGLE_PHASE/DELTA constant-power load equations, fixed ideal switches, bounded WYE/SINGLE_PHASE and cyclic DELTA generator injections/coil limits, bus and branch angle/magnitude bounds, three-phase sequence/VUF limits and bounded transformer equations. It retains recognized generator/source prices and transformer s_rating as declarations, without turning them into physical limits or objective certificates. The device/bound reference retains the complete applicability traces and their intermediate coverage counts.

The same default running-network result now has zero unassessed rows but three failed generator active-power upper bounds. Their excess is about 0.00997 W each, above the unchanged 1e-6 W audit tolerance. All 26 terminal KCL rows and aggregate power pass. A separate tighter native solve of the unchanged input passes the scoped electrical audit. Both results and solver settings are retained by the companion. The cost-scope lesson explains why the recognized price declaration opens these checks without certifying the selected objective, objective value, optimality or custom model. The 42-row table above remains the older audit_linear_solution coverage.

Source-return and device-terminal currents are reconstructed from explicit connection maps; each earth reaction is consumed once. The switch has a checked closed-voltage relation and four conductor current caps. The winding-list transformer has checked coil equations, leakage and current caps; s_rating remains metadata. The running voltage declarations have 28 phase-earth rows, four neutral-earth caps and six phase-pair bounds. No input fields or nonzero residuals are removed to obtain coverage. The delta coil lesson explains why successful KCL still does not replace individual device equations.

A switch needs a state law and a conductor limit

A fixed closed ideal switch requires equal voltages at its ordered endpoint pairs; its nonzero current is determined by the surrounding circuit. A fixed open switch instead requires zero current, with no voltage equality. The package checks both domains against small native solves, including an open switch between separately supplied 10 V and 7 V sides. Missing reported open-switch current remains missing evidence; the checker does not fill in a passing zero.

The companion's closed two-wire fixture uses different endpoint labels and a 10 V, $10+2\mathrm j$ VA load. The phase current is $1-0.2\mathrm j$ A, with its opposite returning through the neutral. At zero neutral voltage, neutral apparent power is zero. Changing only the neutral current rating from 1.1 A to 0.5 A therefore creates a violation of approximately 0.5198 A while all KCL rows, aggregate power and the neutral's zero apparent-power limit still pass. This is the zero-voltage current/power distinction applied to a conductor limit. Changing the limit does not change the supplied operating point; it changes whether that point is admissible.

The switch result reports one current per from-side conductor. The to-side current is its reconstructed opposite, with the to-side terminal map retained; this is not a second independent current measurement. The candidate checks finite nonnegative rating arrays in that order. Malformed or infinite ratings, unknown controls and unrepresented device semantics remain unassessed. The scope is static supplied-state verification, not switching optimization, interruption physics or protection coordination.

Transfer exercise. Keep the two equal-voltage source currents at 6 A and 4 A, but remove the second source's current field. Can the evaluator report zero aggregate error because the first source is present? Check your reasoning: no. A required term is unknown. The aggregate residual is null and its coverage is unassessed; it cannot be computed by substituting zero. Finish the supported model inventory before interpreting a small norm as feasibility evidence.

The solution-multiplicity lesson separates additional admissible circuit roots, continuous current modes and spurious roots introduced by a lost equation. It also distinguishes local optima from the globally preferred branch.

Rate the observation that the constraint names

The transformer reference provides a second representative failure. Its 4 V secondary coil has a reconstructed excitation current $0.2-0.1\mathrm j$ A and supplied series current $-1+0.25\mathrm j$ A. The terminal current is their sum, $-0.8+0.15\mathrm j$ A. The series magnitude is about 1.031 A, but the 0.9 A limit applies to the terminal magnitude, about 0.814 A, and passes. Comparing the series report directly with that terminal limit would reject the wrong observation. Neither current is a universal transformer thermal model.

Transfer task. Tighten only the terminal limit to 0.8 A. Explain why the limit now fails without changing KCL or the operating point. Then remove the series-current observation: distinguish the missing equation evidence from the reconstructed excitation contribution. Use the detailed reference when checking other winding types or declared rating fields.

Solver behaviour and decision margins

Learning question. Does the returned point satisfy the relevant constraints, and is its decision margin larger than the supported error bound?

For a constrained decision problem, solver behaviour is part of the evidence but not the preservation theorem. A reduced model may converge faster because it has fewer variables, or slower because it has denser and more ill-conditioned blocks. A feasible point can also be numerically ambiguous when a constraint margin is comparable with the estimated forward error.

A termination label such as LOCALLY_SOLVED or OPTIMAL records what the algorithm reports under a particular solver interface and tolerance. It is not an independent recomputation of the returned primal point. A scientific use of that result must separately name and check the relevant evidence: finite numeric values, model equations, declared bounds, device and network residuals, recovery obligations, and the level of optimality actually supported. Missing checks remain missing even when the status string sounds successful.

Start with the zero-voltage KCL lesson before using a small power-balance residual as feasibility evidence. At each terminal, $r_S=V\overline{r_I}$ loses its current information at zero voltage and scales it by $|V|$ elsewhere; a total-network power sum can additionally hide opposite nodal errors. The running-network verification script checks line recovery and package profile quantities, but does not thereby supply an independent all-device nodal KCL audit.

This separation also works in the other direction. A status that does not claim feasibility is not evidence that a candidate primal point passed an independent validator, and a narrowly validated primal point does not upgrade a local solver result into a global optimum. Status, primal validation, and optimality evidence are distinct fields of a solution certificate.

For a scalar inequality $g_k(y,u)\le 0$, define the signed margin

\[m_k = -g_k(\widehat y,\widehat u).\]

If $e_k$ is a declared bound or estimate for the error in that constraint, classify the result as:

\[\begin{array}{ll} \text{certified feasible} & m_k>e_k,\\ \text{numerically ambiguous} & |m_k|\le e_k,\\ \text{certified violated} & m_k<-e_k. \end{array}\]

This is deliberately stronger than reporting $g_k\le 0$ at printed precision. The same classification applies to decision comparisons: if two objectives or tap choices differ by less than the propagated numerical error, the certificate should not claim a unique optimum.

Transfer task. Give a residual, tolerance and propagated error bound for one constraint. If the margin lies inside that error bound, report an ambiguous decision even if the printed solver status claims success.

What a transformation certificate should expose

Learning question. Can another reader reconstruct both the electrical comparison and its numerical acceptance criteria?

The numerical fields extend the general preservation contract:

\[\mathcal N = ( \text{coordinates},\text{scaling},\kappa,\eta,\text{dependency graph}, \text{ordering},\text{fill},\text{recovery cost},\text{decision margins} ).\]

An exact electrical map may therefore be computationally unattractive, while an approximate map may be useful if its error and decision domain are explicit. The running network's six generated views should eventually carry this record alongside their source maps; the current fixture records the source maps and solver outcomes but does not yet claim a complete cross-solver conditioning study.

Transfer task. Complete this record for the scalar interval example and one running-case generation. Mark each unavailable field unknown. State which additional observation or bound would change your conclusion.

Find a detailed device or bound check

The audit reference retains every detailed applicability trace in HTML and retrieval. Choose a check by its observation, not by a shared field name. The PDF includes the workflow above; exhaustive array orders, native comparison paths and partial-coverage cases live in that reference. Existing links into this chapter land here and continue to the corresponding reference section.

Transfer task. Pick one reference case, name the supplied observation and its unit, then remove a required input. Identify which known residuals survive and which coverage becomes unassessed. Explain why neither outcome changes the meaning of an absent rating or certifies the selected objective.

Scope and open work

This chapter gives definitions and reporting requirements, not a universal solver benchmark. The next executable tranche should compare the running multiconductor network and its five-bus quotient under a pinned ordering and coordinate convention, report $Y$/Jacobian sparsity and fill-in, and test whether recovered current and rating margins remain decisive after reduction.