# Evidence notes: robotic finger co-design

Source study dated 26 Oct 2026. This page curates supplied computational outputs; it does not run a scientific solver in the browser. No hardware calibration or fabrication result was supplied.

## Final source hierarchy

Use REPORT.md, final v3; analysis_corrected.json; campaign_table.csv; confound.json; manifest.json; and analysis_bundle.tar.gz/ranking_corrected.npz. The final report and corrected analysis match their copies in the supplied final bundles. The original analysis.json and ranking.npz retain a superseded 55-design frontier. The corrected frontier has 72 entries.

## What was checked during curation

- The raw CSV has 4,099 unique candidate IDs, 515 unique skin keys, and every row is tagged fem. The manifest contains 12,297 seed entries.
- The corrected ranking arrays contain 20,000 four-task predictions and 72 true Pareto-mask entries. The six final shortlist predictions match their corresponding pool indices.
- Twelve raw grasp scores are NaN; the CSV contains no floored grasp zeros. The source floors those scores only in memory during surrogate fitting. The displayed 200-row campaign sample uses finite task scores and omits those rows.
- The supplied confound study has 48 runs. Recomputed group means show insertion changing from 7.26 to 4.44 mm over 25 to 400 Hz with equal iterations, and from 3.97 to 7.18 mm with equal episodes. Neither budget isolates a causal sampling-rate effect.

These are file and count checks. The finite-element mechanics, policy learning, surrogate fitting and cloud execution were not independently rerun.

## Interpreting the comparison

Campaign scores come from task simulations using FEM-derived contact responses. Frontier and shortlist scores are corrected surrogate predictions for unsimulated candidates, not physical tests. Reported corrected skin-grouped R² is 0.5382 for grasp, 0.5163 for pivot, 0.3681 for discrimination and 0.4466 for insertion. Residual SD describes model prediction spread, not a formal confidence interval or a calibrated hardware success probability.

Taxel count is a weak modeled predictor, but physical sensor density was not supplied as a distinct metric. No general causal claim that density is unimportant is established. Manufacturing cost and areal density are unavailable rather than zero. The construction drawing is a conceptual schematic, not a recorded full three-dimensional deformation.

## What changed

The initial low-sampling-rate advantage was confounded by unequal experience under fixed training iterations. The final analysis drops the rate predictor and regenerates the frontier and shortlist. The analytic-stub learning_curves.png also predates the final FEM swap and insertion-task repair, so it is not used as final campaign evidence.

## Important model limits

REPORT.md identifies the leading ridge-geometry lever as an unvalidated concentric-ring approximation to parallel ridges. It reports roughly 25–40% of contact increments as extrapolated rather than converged solves, provisional shallow sensor-transfer kernels, placeholder rate-dependence inputs, and a finite 60-iteration learning budget. The reference named prototype_v1 is a placeholder, not a measured fabricated finger. The six prototype predictions overlap substantially and should be tested as a set.

## Figures

- candidate-frontier.png is the unaltered fig_pareto.png. It explicitly shows the corrected 72-design frontier.
- design-lever-attribution.png is the unaltered fig_sensitivity.png. Sampling rate is absent from the corrected attribution plot.

The downloadable replay contains curated finite data and labels for these distinctions.