Release notes v0.5.1

Patch on the 0.5.0 package surface. v0.5.0 (2026-09-20) shipped class-wide recipes and Pages. Hourly folds after that stayed pinned at 0.5.0. This cut is the human-facing pin.

Place typed probabilistic judgment for the decision-model class. TypeSafe Jev (Choice, Score, Noul) is the dominant exemplar most users will call. Peers are in the class. They are not equal in adoption. Exact work stays in code or policy. A soft score is not a proof.

Pegged against typesafe-ai/skills v0.5.7 (65a39f3, 2026-09-12). Live HEAD of that repo is still this commit. Checked 2026-09-21.

Prior class recipes: v0.5.0 notes. Diagram: with vs without Augustus.

Added

• GEPA domain-adapt. Praneeth ADE on a Jev Choice. Schema-valid is not the same as correct. API confidence is not P(correct). GEPA revises Choice instructions and criteria with weights fixed. Brier and F1 stay theirs. A review queue is not an F1. A soft score is not a gate.

Changed

• Package pin 0.5.0 to 0.5.1 (skill metadata, marketplace, README, Pages kicker, CITATION.cff).
• README still ends at License. The uniqueness gate fails if a dump wall returns after that heading.
• Skill YAML description is a short class-first blurb. The trigger keyword wall is references/activation-triggers.md.

Catalog densifies since 0.5.0 stay in research/notes.md and research/changelog-hourly.md. They are not release-note walls.

No invented metrics. No pipeline() install recipes.

Recipes

Backend-agnostic. Problem, without Augustus, with Augustus, what to measure. Numbers below are theirs, not a new bench. Quote the source card. Do not treat this page as a leaderboard.

Domain adapt: GEPA on Jev

• Problem. A schema-valid Choice is treated as a correct label, API confidence is treated as P(correct), or F1 is treated as the review-queue policy.
• Without. Ship the adapted prompt. Treat GEPA as a new scoring-table species.
• With. schema-valid is not the same as correct. API confidence is not P(correct). GEPA revises Choice instructions and criteria with weights fixed. Brier and F1 stay theirs. A review queue is not an F1. A soft score is not a gate. Not an 18th scoring-table species.
• Measure. Brier 0.1357→0.0747 theirs. F1 69.1%→79.7% theirs. FN 4→6 theirs. Retention of positives under the review cutoff is a different ledger from F1.

Earlier class recipes (encoder locate vs decide, open heads, NanoJev, llm-to-jev, jcr, prompted JSON, measurement honesty) stay on the v0.5.0 notes.

Homepage: https://24601.github.io/Augustus/

Install

claude plugin marketplace add 24601/Augustus
claude plugin install augustus@augustus

npx skills add 24601/Augustus --skill augustus

Full notes: CHANGELOG.md