Federated scientific knowledge: end-to-end trace

Page status: maintained cross-repository implementation trace and scope boundary.

This page traces eighteen scientific and negative-knowledge objects through book evidence, applicable BMOPFTools guardrails, and the book's existing LLM context packets. The stable ownership rules are defined in the repository-root ARCHITECTURE.md; this page is the worked implementation trace.

First trace: parallel member limits

The linked objects

The same identifiers appear in two exports, but their contents do not have two owners. generated/scientific_knowledge.jsonl describes the scientific object. BMOPFTools' generated/executable_knowledge.jsonl describes code, Findings, and fixtures. generated/federated-knowledge-pair-manifest.json pins the matching export hashes and checks that both sides agree on their shared IDs.

Scientific statement and evidence

PSK-000001 states that summing fixed linear parallel admittances preserves the unconstrained aggregate terminal relation, while giving the aggregate the sum of member ratings can relax the member-constrained feasible set. Its declared scope is fixed-linear members with common endpoints and voltage coordinates, retained admittances, and explicit current limits.

The record deliberately does not establish that all parallel aggregation is invalid, that every aggregate rating is an outer relaxation, or that a scalar witness proves the multiconductor or state-dependent case. The claims and generated book artifacts remain the evidence authority; the package does not independently rewrite that conclusion.

Executable decision

BMOPFTools receives an explicit mapping from source members to a target aggregate. In its current supported domain it requires scalar, finite, nonzero, series-only impedances, common terminal coordinates, and scalar ratings. It then performs two distinct checks:

1. It verifies that the target admittance equals the sum of the source-member admittances. Failure produces E.CONTRACT.PARALLEL_TERMINAL_RELATION_MISMATCH.
2. If terminal behavior is preserved, it compares the exact scalar voltage-drop region induced by every source member rating with the target aggregate-rating region. An inner restriction or outer relaxation produces W.CONTRACT.PARALLEL_MEMBER_LIMIT_LOSS.

Missing mapped data produces W.CONTRACT.INDETERMINATE. Multiconductor, shunted, singular, or otherwise out-of-domain cases produce I.CONTRACT.NOT_APPLICABLE. Those statuses are part of the scientific behavior: the implementation refuses to silently generalize beyond its declared domain.

The same check is available as the source-hash-bound BMOPFTools recipe parallel_member_limits. The pinned pair manifest carries its command, expected status, expected Finding codes, and recipe source hash. Context packets therefore expose a directly runnable implementation example without making the book execute a sibling checkout or copying the package implementation into the scientific corpus.

For the minimized fixture, the two source lines have resistances 0.1 Ω and 1 Ω and ratings of 100 A each. Their summed admittance is 11 S, so the target resistance is 1/11 Ω. A naïve 200 A aggregate rating permits a voltage drop of about 18.18 V, while the source members permit only 10 V. At the committed 15 V witness, aggregate current is 165 A, but member currents are 150 A and 15 A. The target accepts the point while the first source member is overloaded. The exact scalar target rating is 110 A, not 200 A.

Retrieval and answer contract

A query such as “Can a preprocessing pass merge parallel branches by summing admittance and capacity?” follows the existing misconception router. The context packet makes the three claims, three vocabulary concepts, and knowledge:PSK-000001 mandatory. It retains the existing qualified, under_retrieved, and unsupported statuses and source-hash checks.

The packet now also exposes explicit sections:

• scientific_basis for the scoped PSK statement and evidence status;
• known_misconceptions for the tempting shortcut and required qualification;
• counterexamples for book artifacts and stable fixture identities;
• negative_results, numerical_pathologies, scope_boundaries, and open_questions for distinct qualified negative-knowledge classes;
• executable_checks for the BMOPFTools repository, contract IDs, Finding codes, fixtures, and implementation status;
• implementation_examples for concrete executable fixtures and pinned recipes; and
• unresolved_boundaries for every claim the scientific object says it does not establish.

These are structured views over mandatory records, not a second retrieval index. The deterministic corpus, release identity, source hashes, retrieval evaluation, MCP route, HTTP route, and CLI route remain the book's established machinery.

BMOPFTools also provides a separate read-only MCP stdio adapter for the six settled package execution operations. Connecting both servers gives an agent a book-owned retrieval route and a package-owned execution route; it does not merge their authority or allow an execution result to silently upgrade a book claim. The federated recipe pins include the adapter source files, so changes to that transport invalidate the paired source hashes.

Second trace: neutral, ground, and reference

The second slice reuses the same cross-repository mechanism for a model-semantics failure. Its stable chain is:

PSK-000002 states that an explicit neutral conductor, a finite grounding relation, a physical earth-return model, and a mathematical voltage reference are distinct objects. The book owns that definition and the recorded comparison between floating, finite-impedance, and ideal customer-end grounding. The package fixture is executable evidence for one concrete transformation, not a promotion of that fixture into a theorem.

In the fixture source, a phase/neutral feeder connects the source and load buses. The source neutral is perfectly grounded, while the load neutral has a finite 0.1 S grounding shunt. The unsafe target retains both simple bus records and their n labels, but removes neutral continuity from the feeder and marks the load neutral as perfectly grounded. The contract therefore emits two findings: one for the changed neutral-continuity graph and one for replacing the finite grounding relation with a perfect local ground. An exact target may rename the feeder and shunt while preserving all three checked representation dimensions.

The contract explicitly does not establish electrical terminal equivalence, explicit-earth-conductor behavior, soil or electrode behavior, grounding-asset identity or state, fault current, touch voltage, or protection operation. It returns inapplicable for unsupported coupled grounding models and indeterminate when mapped evidence is unavailable.

The source-hash-bound neutral_ground_reference recipe makes this second transport mapping directly runnable. Its pinned command, expected failed status, two expected Finding codes, and source hash are carried by the federated pair manifest. Context packets therefore expose a concrete operation without copying BMOPFTools behavior into the book or inferring the bus mapping from matching names.

Queries from all three audience routes for ground-neutral-reference-are-one-node now make knowledge:PSK-000002 mandatory. Their context packets expose the scoped scientific basis, linked counterexample, implemented BMOPFTools contract, pinned recipe, dedicated Finding codes, and the unresolved dimensions without changing the deterministic router or its qualified, under_retrieved, and unsupported semantics.

Third trace: solver termination and solution validity

The third slice tests an invalid scientific inference about numerical evidence:

PSK-000003 states that a solver termination label is algorithm evidence, not an independent certificate of finite primal values, model equations, study limits, recovery, or local/global optimality. The related book diagnostics artifact already records solver-status and package-diagnostic layers separately; the package fixture now makes the invalid inference directly executable.

The fixture network declares a 200–260 V bus-voltage range. Its negative result is labelled LOCALLY_SOLVED but reports 180.5 V. The contract first requires complete numeric vr, vi, and vm data for every declared bus terminal, then reuses BMOPFTools' existing profile_solution implementation. It retains E.SOL.VOLT_VIOLATION as underlying evidence and emits the dedicated contract finding. A companion 230.5 V result passes only termination status, result-tree finiteness, and the declared bus voltage/angle dimensions.

That pass is intentionally not a general solution certificate. Network-equation residuals, thermal and device limits, load-model residuals, power balance, objective optimality, global guarantees, and solver derivative quality remain explicitly unassessed. A non-feasible solver status is inapplicable; missing status or bus-terminal evidence is indeterminate.

All three audience routes make knowledge:PSK-000003 mandatory and expose the counterexample, executable contract, dedicated Findings, and unresolved checks through the same deterministic, source-hash-bound context machinery.

Fourth trace: load connection voltage bases

The fourth slice turns an existing package plausibility diagnostic into an explicit scientific contract:

PSK-000004 states that nominal voltage is part of a voltage-dependent load's constitutive coordinate. WYE uses phase-to-neutral voltage; DELTA uses line-to-line voltage. On the declared nominal three-phase system those bases differ by sqrt(3). Copying one numeric value into both declarations therefore changes normalized constant-current, constant-impedance, ZIP, or exponential behavior rather than merely changing metadata.

The fixture uses a balanced source with a 230 V phase-to-neutral base. Its DELTA ZIP load incorrectly declares v_nom=230 V, while the corresponding line-to-line base is about 398.37 V. The executable contract reuses the same source/transformer voltage propagation, connection-coordinate conversion, and default 0.8–1.25 plausibility band as the existing W.LOAD.VNOM_MISMATCH diagnostic. The negative case produces the dedicated contract failure; the companion 398.37 V declaration passes.

That pass remains declaration-relative. Source and transformer nominal values, terminal maps, coefficients, units, operating voltage, network equations, and equipment limits are unassessed. All three audience routes make knowledge:PSK-000004 mandatory and expose those boundaries through the same deterministic context-packet machinery.

Fifth trace: adjustable transformer tap domains

The fifth slice preserves a decision domain rather than one transformer snapshot:

PSK-000005 states that exact parameterized preservation retains the mapped tap identity and complete admissible domain. A tap start is one admissible initial point, not a substitute for that set. The existing TR-XFMR-005 certificate establishes the pointwise parameterized factor and a discrete decision witness; TR-XFMR-006 embeds the retained finite tap factor in its declared AC network and records the consequence of freezing the 1.00 start. Those book artifacts remain the scientific authority.

The package's initial check is deliberately smaller. It accepts one mapped two-winding isolating transformer, requires a finite positive continuous source interval and unchanged subtype/non-tap declarations, and compares the source and target intervals. Omitting target bounds has its documented package meaning: a fixed singleton at tap. The fixture therefore fails as an inner_restriction and carries a source-admissible, target-inadmissible tap witness. Retaining [0.95,1.05] passes even if the target's admissible start differs, proving that the implementation does not confuse domain identity with initialization identity.

That pass does not establish transformer-equation, loss, discrete-position, coupling, automatic-control, network-feasible-set, objective, optimal-tap, or solver equivalence. All three audience routes make knowledge:PSK-000005 mandatory and expose the book certificates, executable interval check, counterexample, and unresolved boundaries through the existing deterministic retrieval path.

Sixth trace: transformer winding conventions

The sixth slice prevents an ordinary-edge shortcut from silently changing a typed transformer factor:

PSK-000006 states the broad preservation obligation. The winding-normalization certificate records the exact terminal-coordinate action and dual current map; the fixed-factor certificate records the connection incidence, voltage transfer, leakage, excitation, grounding, and recovery anatomy. Those book artifacts remain authoritative for the general scientific claim.

The package check covers only fixed-tap single_phase, wye_delta, and delta_wye records with the same subtype, explicit one-to-one bus mapping, and stable or explicitly mapped terminal labels. It compares winding-side identity, ordered coil incidence, winding reference voltages, and fixed effective coil ratio. Adjustable taps route to PSK-000005; subtype-changing or fully reversed encodings require a complete typed transformation outside this initial contract.

The negative fixture retains the unordered bus pair but swaps its endpoint fields while leaving the WYE/DELTA subtype and terminal maps untouched. It therefore fails on mapped coil incidence. The companion target retains the convention and passes narrowly. Neither result establishes equality of leakage, excitation, grounding, limits, complete terminal factors, controls, network feasible sets, objectives, or solver behavior. All three audience routes make knowledge:PSK-000006 mandatory and retain those limits through the deterministic, source-hash-bound context packet.

Seventh trace: terminal versus decision equivalence

The seventh slice makes the federation's umbrella preservation question executable at the declaration boundary:

The book's observation-indexed definition and parallel feasible-set witness remain the scientific basis. The package fixture does not dispute its terminal certificate; it shows that this one reference cannot close the admissible domain, observations, constraints, decision variables, objective, and recovery obligations of an exact decision-equivalence claim.

The evidence-complete companion manifest declares every required dimension and passes narrowly. That pass means the manifest contains supported dispositions, not that BMOPFTools has resolved or authenticated the evidence IDs. It does not compare equations, feasible sets, objectives, optimizers, or recovery maps. Case-specific contracts such as parallel member-limit preservation and transformer tap-domain preservation remain the sources of actual evidence for individual dimensions.

All three existing audience routes now make knowledge:PSK-000007 mandatory and expose the completeness contract and terminal-only counterexample through the same deterministic context packet. The existing unsupported, under-retrieved, qualification, and source-hash semantics are unchanged.

Eighth trace: Kron boundary exactness and internal recovery

The eighth slice carries the tutorial's grounding warning into a deliberately narrow boundary contract:

The book's existing running-network, explicit-earth, multiple-grounding-point, and nonlinear witnesses remain the richer scientific evidence. The package fixture is a pedagogical guardrail and does not replace those witnesses or claim that a passing boundary check preserves internal limits, protection, complete feasible sets, objectives, or solver results. All three audience routes make knowledge:PSK-000008 mandatory when the existing Kron misconception is triggered; the deterministic packet and source-hash checks remain unchanged.

Ninth trace: positive-sequence collapse applicability

The ninth slice guards a controlled specialization onto the positive-sequence invariant subspace:

The passing companion in the contract tests is intentionally narrower than a balanced-model claim: it leaves phase-specific limits, neutral/earth, negative/zero-sequence, protection, internal devices, complete feasible sets, objectives, and solver behavior unassessed. Existing sequence witnesses remain the scientific evidence; the package check is the refusal boundary that keeps those claims from being silently promoted.

Tenth trace: fixed versus state-dependent equivalents

The tenth slice guards the promotion of an operating-point map into a reusable state-domain equivalent:

The package pass is declaration evidence only: it does not validate a nonlinear update law or establish feasible-set, objective, protection, or solver equivalence. The book's state-dependent grounding and continuation witnesses remain the scientific evidence.

Eleventh trace: floating references and singularity

The eleventh slice makes reference/rank preservation explicit before a transformed model is accepted:

The pass does not replace equation-level Ybus/Jacobian or solver validation; the package checks declaration consistency only. Existing numerical and grounding evidence remains the scientific basis.

Twelfth trace: terminal/conductor ordering and permutation

The twelfth slice turns a typed coordinate-action obligation into a narrow metamorphic primitive check:

The package pass is not a complete coordinate-transformation theorem. It leaves downstream limits, nonlinear state, network feasible sets, decisions, objectives, recovery, and solver behavior unassessed; the book retains those claims and the pedagogical permutation examples. The seeded suite covers reciprocal, strictly diagonally dominant complex series matrices with one through six conductors. It checks exact permutations, deliberate one-entry corruptions, deterministic replay, and one-conductor shrinking. The injected failures are classified as expected contract rejections, not newly discovered scientific counterexamples.

Thirteenth trace: complete solved-network feasibility

The thirteenth slice separates solver status from an independently computed feasibility witness:

The package validates the supplied witness; it does not calculate residuals or prove its completeness, global optimality, or full decision equivalence.

Fourteenth trace: unit/base and serialization invariance

The fourteenth slice binds declared conversion metadata to a canonical payload:

The package pass binds declared serialization evidence only; parser/writer round-trip behavior, hash computation, complete equations, limits, and study decisions remain separate validation dimensions. The seeded suite generates positive finite SI base maps, performs real JSON encode/decode round trips with reordered keys, and independently mutates the unit system, one base value, and the semantic hash. Every injected failure is reduced to a one-base witness and classified as an expected contract rejection, not a scientific counterexample or proof of physical SI-to-per-unit conversion.

Fifteenth trace: a negative retrieval result without an executable package contract

LLM-RETRIEVAL-001 and PSK-000015 record a different kind of knowledge object. The pinned compact neural retriever and generic cross-encoder reranker each failed at least one of the benchmark's predeclared not-worse-than-hybrid retrieval gates. Neither was promoted, and the failed experiment remains available as a source-bound negative result rather than disappearing behind the retained deterministic hybrid route. The scientific record hashes an immutable generated snapshot of the benchmark evidence; the separate mutable compatibility marker can then track later corpus revisions without creating a corpus-hash cycle.

The record contains the question, hypothesis, motivation, setup, cases, software and model provenance, attempted methods, observed result, failure criterion, interpretation, limited conclusions, possible changed conditions, reproduction commands, and review status required by the negative-result quality standard. Its scope is deliberately the archived prior corpus named by the benchmark artifact. It does not claim that neural retrieval is generally inferior or that the recorded rankings describe the current corpus.

This trace has no BMOPFTools contract, Finding, or fixture. Its executable relationship is explicitly not_applicable, and the federated pair records an empty execution link instead of inventing package ownership. The book's misconception router still makes the claim and PSK-000015 mandatory for questions that assume embeddings or a generic reranker must improve retrieval. Context packets expose the structured record under negative_results while leaving executable_checks empty.

Sixteenth trace: iteration failure is not an infeasibility certificate

LOAD-CONTINUATION-001 and PSK-000016 expose the first structured numerical-pathology. On the recorded scalar two-bus fixture, a damped fixed-point continuation of the constant-power branch converges through demand scale 1.7 and fails the declared iteration and residual test at 1.8. The other three declared load-law branches remain converged through 3.0. A separate standard-library Python implementation reproduces the sampled convergence flags, converged rows, and CP failure scale.

That reproduction establishes the behavior of the declared algorithm on the declared grid. It does not show that the equations are infeasible at 1.8, that 1.8 is a saddle-node, or that constant-power loads are generally less appropriate. The pathology record therefore keeps observed_behavior and algorithmic_boundary separate, lists the tempting invalid inferences, and requires discriminating checks such as step/initialization variation and a mathematically appropriate continuation method.

The experiment lives in the book repository rather than BMOPFTools, so its execution relationship is not_applicable. This does not contradict the package's separate solved-network feasibility guardrail: that guardrail deliberately refuses a residual-witness conclusion when no solved status is available. The federation does not pretend that this generic refusal behavior reproduces or certifies the continuation experiment. Context packets expose the record under numerical_pathologies with no package execution link.

Seventeenth trace: exact coupled-corridor lowering has a declared boundary

COUPLED-CORRIDOR-001, COUPLED-CORRIDOR-002, and PSK-000017 expose the first structured scope-boundary. The established result is deliberately narrow: two reciprocal fixed-linear scalar series sections with a nonsingular joint impedance admit an exact six-edge weighted-lattice realization of the four-terminal equation, with recovery of the two source currents. The executable witness also checks orientation covariance, signed generated weights, per-unit round trips, and refusal of inverse-based lowering for a singular joint primitive.

The boundary record makes the adjacent non-results machine-readable. The generated cross edges are equation objects rather than conductors or galvanic connections, and the scalar proof does not automatically cover block-valued or full-pi factors, physical limits, states, protection, measurements, feasible decisions, or objectives. Extending the result requires an appropriate block or tableau derivation plus separate recovery and study-preservation evidence.

The witness is a book experiment and has no BMOPFTools counterpart. Its federated execution status is therefore not_applicable, while the existing mutual-coupling misconception makes PSK-000017 mandatory for all three audiences. Context packets expose the structured record under scope_boundaries; they do not present the result as a counterexample or invent an executable package link.

Eighteenth trace: the CS1035 source construction remains an open question

AU-CARSON-001 and PSK-000018 expose the first structured open-question. The repository has an independent Australian CS1035 four-conductor reference matrix and a separately identified UGHV cable fixture, but the available source history does not map that fixture to the reference. The generated reproduction therefore keeps the matrix out of its model inputs and retains the mapping as unresolved.

The record distinguishes the evidence already in hand from the information still missing. OpenDSSDirect and an independent linear-algebra solve agree for the declared UGHV fixture, while changing the Carson diagnostic probe from 50 Hz to 60 Hz increases rather than resolves the underground matrix mismatch. Those checks validate the fixture computation and reject one possible explanation; they do not recover the CS1035 conductor, screen, geometry, earth-return, ordering, or provenance fields.

The new provenance-shortcut misconception makes the record mandatory for student, software-engineer, and power-engineer questions. Context packets expose it under open_questions, including the attempted discriminators and resolution criteria. The record is book-owned and has execution status not_applicable: the book experiment is evidence for the open question, not a BMOPFTools contract or a license to relabel the UGHV fixture.

What generalizes

The federation pattern generalizes to other scientific guardrails:

• one stable PSK identity can link claims, misconceptions, artifacts, executable contracts, APIs, Findings, fixtures, and recipes;
• scientific and executable exports can evolve independently while a pair manifest detects incompatible releases;
• retrieval can make a PSK mandatory through the existing misconception graph;
• implementations can report passed, failed, inapplicable, or indeterminate without turning absence of evidence into a pass; and
• source hashes make context packets and cross-repository links auditable.

The numerical formula used by the first contract does not generalize automatically. The scalar voltage-drop reduction, lack of shunts, fixed linear admittances, common coordinates, and explicit current ratings are case-specific. The second contract likewise does not turn relation matching into electrical equivalence: its explicit-neutral and scalar neutral-only grounding domain is case-specific. The third contract establishes that the contract/result/refusal machinery generalizes to post-solve evidence, but its initial bus-result coverage is not a universal feasibility or optimality validator. The fourth shows that an existing package diagnostic can be promoted into an auditable contract without duplicating its inference machinery, but declaration consistency is not importer fidelity or solved-model validation. The fifth separates preservation of an adjustable domain from equality at one initialized snapshot, while keeping broader transformer physics and decision equivalence outside the initial package check. The sixth shows how a typed coordinate convention can be checked without pretending that a compact declaration check proves the completed factor or a subtype-changing reversal. The seventh generalizes the routing and refusal pattern to an umbrella evidence manifest while deliberately leaving the truth of every cited dimension to narrower checks. Multiconductor coupling, nonlinear devices, explicit-earth networks, protection quantities, full equation residuals, device constraints, recovery, and global optimality require separate PSK objects or broader contracts with their own evidence and fixtures.

Reproduction and release pairing

From the book repository with BMOPFTools checked out beside it:

python3 scripts/generate_scientific_knowledge.py --check
python3 scripts/check_federated_knowledge.py --check --bmopf-root ../BMOPFTools.jl
python3 scripts/check_llm_reproducibility.py

From BMOPFTools:

python3 scripts/generate_executable_knowledge.py --check
julia --project=test --startup-file=no -e \
  'using Test, BMOPFTools; include("test/scientific_contract_tests.jl"); include("test/executable_knowledge_tests.jl")'

Changing either export requires regenerating and reviewing the pair manifest. A normal book-only check validates its pinned snapshot; the sibling-aware check proves that the local BMOPFTools checkout still matches that snapshot.