SEM Analysis Pipeline v0.1.0

Version 0.1.0 is the first stable release of the SEM Analysis Pipeline.

Archived software DOI: https://doi.org/10.5281/zenodo.21845096

The released capability is paired-acquisition external-aggregate projected
aspect ratio. It uses automatic multiscale silhouette segmentation, explicit
abstention, and paired physical-mask IoU quality control. It does not measure
isolated-crystal morphology or three-dimensional sphericity.

Verification

• All 27 repository tests pass on macOS and Ubuntu with Python 3.11, 3.12, and
  3.13.
• A clean 64-image rerun reproduced all 32 manuscript field dispositions, all
  16 accepted primary and secondary aspect ratios, and all 22 available
  paired-mask IoUs exactly.
• The blinded real-envelope synthetic challenge accepted 33 of 60 images, with
  minimum accepted-mask IoU 0.9566 and maximum absolute aspect-ratio error
  2.69%.
• All eight complete accepted synthetic pairs passed paired-mask QC, with
  maximum cross-acquisition aspect-ratio difference 1.56%.
• An independent public-only audit regenerated the challenge corpus, reran the
  tagged candidate package, and reproduced the published synthetic results.
  A separate Claude Code audit independently reproduced the released
  single-image score report byte-for-byte.

Frozen validation assets

• sem_paired_real_envelope_validation_frozen.zip:
  7ea6207e8f622917202a7565d6a09ae8d6a42d27822b26636e0f3749073a3541
• sem_real_envelope_predictions_frozen.zip:
  0ac1166024c34826f88b40d98ae7f24d27da5d4f74fe131fe2175643844c423e
• sem_real_envelope_scores_single.json:
  154e3552eff82db1720167f0d9686cc6f0949620e757ca18b5ba57e3b21a2758
• sem_real_envelope_scores_paired.json:
  21d6ed9869d14621aaef1fadb8f0dee09a0aa914fac03af0340ed76cc36ebf9e

Regenerating the private manifest in an independent environment changed nine
harmonic-amplitude decimal serializations by at most 6.94e-18; all 60
rendered images and all 60 truth masks remained byte-identical.

Synthetic validation is limited to one isolated, fully visible, bright,
star-convex aggregate within the documented acquisition envelope. It is not
experimental ground truth for touching, truncated, charging-affected, or
otherwise unmodeled SEM fields. The public manuscript-reference tables are
internally consistent, but the 64-image experimental segmentation cannot be
independently rerun from this software release because the source images and
exact experimental run manifests are handled in a separate research-data
deposit.