Make the frozen uuid2 candidate the default M checkpoint. Retain the original XS (1.045 MB) and S (9.167 MB) inference files as selectable variants, and package M at 24.575 MB with byte-verified FP32 Core ML coefficients. Update the model card with all three same-corpus scores versus Gitleaks, repository uncertainty, reused-development-data caveats, paper links and Mill badges. These are distinct authored checkpoints, not quantizations. Native M parity remains untested. Keep the repository public as explicitly requested.
9 days ago
Release frozen 9.17 MB Blot Mini precision winner
Replace the default r3a inference files with the exact precision-a-2026-11-04 candidate, update examples and checksums, and verify Core ML/NumPy parity. Align the model card with the released weights, retain historical r3a evidence, explain the separate CredData and issue-sample comparisons, and add the Weight-Space Programming paper section and Mill badges. Keep the repository private.
10 days ago
Document precision research results and add Built with Mill badges
Add the frozen 2,000-case precision-candidate comparison, confidence intervals, evaluation scope, overlap-audit correction and aggregate evidence. Identify the newer candidate separately from the unchanged published r3a weights. Add Mill attribution and branded method, task, checkpoint, parameter and format badges.
Publish Blot Mini authored-weight Core ML research checkpoint
Private release of frozen r3a with runnable reference and native inference, weight builder, model card, aggregate evaluation and curated fork-and-flush provenance.
