Graph-safe reconstruction benchmarks for schematic routing

1 Introduction

Learned engineering systems require labels that describe the intended task. A noisy label can reward an electrically incorrect route or punish a valid alternative. This risk is high when training data comes from edited CAD files.

Schematic reconstruction offers a practical source of supervision. A human route is removed, and a system must restore the connection. The original geometry can then act as a preference reference.

Removal is not automatically safe. One wire segment can participate in a larger net tree, a junction, or an implicit attachment. A geometric cut can therefore change the electrical meaning of the task.

We address this problem with a graph-safe admission rule. A reconstruction task is accepted only after reference restoration reproduces the original connectivity graph. The rule makes benchmark construction conservative and reproducible.

2 Benchmark objective

Each task contains a scene with one route removed, a set of required anchors, and a human reference route. The system generates several orthogonal candidates for the missing connection. A deterministic validator labels every candidate before any preference score is applied.

The benchmark has two different labels. Validity states whether a candidate preserves required electrical and geometric constraints. Preference states how closely a valid candidate follows useful human routing structure.

This separation prevents a short route from receiving a strong target after it changes a net. It also permits several valid shapes for the same electrical relationship. The human route is a reference, not the only acceptable answer.

3 Corpus construction

The source corpus contains public KiCad schematics with recorded repository revision and license metadata. The collection contains 120 schematics. All 120 source files parsed in the recorded corpus build.

The parser extracts symbols, pins, wires, junctions, labels, text, and sheet geometry. A connectivity builder converts these primitives into a canonical graph. The graph represents pin membership and resolved electrical nets.

The extractor searches for direct-wire reconstruction cases. It does not assume that every wire is a safe training example. Unsupported multi-net or ambiguous structures are skipped.

4 Task admission procedure

4.1 Controlled removal

The extractor records the route anchors before it removes geometry. Anchors can be symbol pins or deliberate points on an existing wire. Mid-wire anchors also carry their required junction semantics.

The route is removed from a copy of the scene. The remaining graph becomes the input state for candidate generation. The original scene stays unchanged as the expected reference.

4.2 Restoration control

The extractor inserts the human route back into the reduced scene. It rebuilds the full graph and compares the result with the original graph hash. A mismatch rejects the task before evaluation.

This control catches hidden junction effects and unintended net partitions. It also detects parser or writer behavior that changes connectivity during a round trip. The task set therefore measures the router, not extraction damage.

4.3 Task-level split

All candidates from one reconstruction task remain in one data split. Candidate-level splitting would leak the same anchors and reference geometry across training and test sets. Task-level splitting reduces this source of dependence.

5 Candidate labels and metrics

Candidate generation produces Manhattan variants and obstacle-aware paths. It also supports multi-anchor trunks when the task contains more than two anchors. Each candidate records its family and complete geometry.

The validator inserts a candidate and rebuilds the graph. It checks anchor connectivity, graph equality, junction requirements, grid alignment, bounds, symbol clearance, and source round-trip behavior.

Preference features include length ratio, segment overlap, bend structure, directness, and local clearance. Invalid candidates receive no positive preference target. This rule prevents aesthetic similarity from hiding an electrical defect.

6 Evaluation results

The extraction procedure admitted 342 graph-safe tasks from the 120 parsed schematics. The initial evaluation used 240 tasks. Candidate generation produced 5,668 labeled records.

The selected-route pass rate was 95.8 percent. The valid candidate rate was 45.6 percent. This gap is an important property of the benchmark.

A high selected pass rate does not mean that candidate generation is solved. More than half of the candidate pool failed at least one hard check. The selector succeeded because it could identify useful options inside a difficult pool.

The candidate pool also defines the ranking ceiling. A ranker cannot recover a strong route when no strong route is present. We therefore retain candidate validity and oracle quality as separate diagnostics.

7 Threats to validity

The benchmark measures direct-wire reconstruction. It does not decide whether a relationship should use a wire, label, bus, or power symbol. Those representation choices require a broader authoring benchmark.

Public schematics contain repeated conventions and project-specific styles. Task-level splitting reduces local leakage, but it does not remove all source-family dependence. Repository-level holdouts are a useful next control.

Human geometry is not always optimal. The preference target can encode historical clutter or inconsistent layout practice. Electrical validity must therefore remain independent of human similarity.

The current corpus size supports engineering comparison, not broad statistical claims. Future reports should include repeated splits and confidence intervals. The present results are descriptive.