# Physical-grid P5 visual contract

- Artifact: two-panel classification-error figure.
- Target format: full-width paper figure; editable SVG/PDF plus PNG preview.
- Core claim: local state-dependent residualization restores physical local learning and composes with overclamping.
- Reviewer question: does SDIL improve downstream classification over raw learning, constant calibration, and overclamping under the same component imperfections?
- Evidence layer: main result.
- Source data: `results/physical_bias/p5_full_grid_bias_crossover.json`.
- Statistics: 32 trials per input diameter (eight label rotations by four device draws); show every trial and the arithmetic mean, without inferred uncertainty intervals.
- Panel map: (a) raw, per-edge constant calibration, and SDIL under standard clamping; (b) overclamping and overclamping plus SDIL.
- Labels: classification error (%), input diameter (mV), method names, eight trials per diameter.
- Caption role: state the physical-grid setting, trial count, component-error model, and perfect-result counts without extending the evidence to hardware measurements.
- Output: `results/figs/physical_grid_p5_full.{svg,pdf,png}`.
- Traceability: plotting code reads the committed JSON directly; no values are copied by hand.