Release notes v0.5.0

Place typed probabilistic judgment. Jev is the exemplar, not the monopoly. The same with/without split applies to GLiNER/GLiClass, Laya, SemIf, NanoJev, kev, Jeff-1, localjev, llm-to-jev shaped prompts, and other Choice/Score/Noul-style tools. 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.

Diagram: with vs without Augustus (README SVG + Pages). Merged #39 hygiene is on main. Open #41 is not in this cut.

Added

• Pages / onboarding. Custom layout. With vs without comparison.
• README. Scannable skill map. No uniqueness wall.
• Class / migration. llm-to-jev on-ramp. SemIf rename densify. NanoJev unified-games densify. jcr lookup does not execute.
• Measurement honesty. Hysteresis is policy. Instruct-tuning ECE. hop-ECE invariance. Harbor-jevals / Verdict / DecisionOps as themes. ranking ≠ calibration. soft Noul ≠ hard gate.

No invented metrics. No pipeline() install recipes. heuristic conversion ≠ calibrated Noul.

Recipes

Backend-agnostic. Problem → without Augustus → with Augustus → what to measure. Numbers below are shapes, not new benches. Quote theirs on the source card; do not treat this page as a leaderboard.

Encoder locate vs decide — GLiNER / GLiClass

• Problem. A locate or categorize encoder is dropped in as if it were a Choice/Score/Noul head.
• Without. Swap GLiNER for the decision model. Hard-gate spans or label scores. Bake off against a chat LLM.
• With. Species map first (locate vs categorize vs safety-schema vs decide). Extractive spans stay extractive. Remainder judgment is a separate question. Soft scores ≠ hard gates.
• Measure. Span/F1 or assignment quality on the encoder task, separately from ECE/Brier on the decision. Never quote a softmax or span score as a calibrated Noul.

Open heads — Laya, SemIf, kev, Jeff-1

• Problem. Wire-compat, argmax agree, or a speedup is treated as a replica of a calibrated decision model.
• Without. Drop-in swap. Ship the systems timing. Skip OOD. Treat accuracy as calibration.
• With. Softmax over options ≠ calibrated Noul. Systems comparison ≠ semantic equivalence. Replica honesty: name the head (open, constrained-AR, encoder, specialist). Acc vs ECE. Fail polarity in code.
• Measure. Held-out ECE/Brier and accuracy/AUC. In-distribution vs OOD. Do not promote a speedup (even a large one) as semantic equivalence.

Gameplay specialist — NanoJev

• Problem. Game success is treated as a calibrated Noul.
• Without. Quote a win rate as if it licensed a production gate.
• With. Specialist gameplay S1 / open replica, not TypeSafe Jev. Local boolean ≠ TypeSafe noul. Soft scores ≠ hard gates.
• Measure. Held-out game metrics on one ledger. ECE/Brier on another. Do not mix the two.

Prompt conversion — llm-to-jev

• Problem. A decision-shaped LLM prompt is assumed to be an equivalent Choice/Score/Noul.
• Without. Paste, convert, ship. Treat the compiler as a replica.
• With. Heuristic on-ramp, not an LLM and not a guarantee. Review the generated Score rubric. Ordered criteria, not an arbitrary 0–1 range. Prose stays with the LLM. Conversion ≠ calibrated Noul.
• Measure. Suitability / compatibility labels and human review of generated criteria. Not “equivalent behavior.”

Domain adapt — GEPA on Jev

• Problem. A schema-valid Choice is assumed correct, or API confidence is treated as P(correct), or F1 is treated as a 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/criteria with weights fixed. jev-1.13.0 weights fixed. review-queue policy is not F1. Soft is not gate. Not an 18th scoring-table species.
• Measure. Brier 0.1357→0.0747 theirs. F1 69.1%→79.7% theirs. FN 4→6. Retention of positives under the review cutoff is a different ledger.

Capability lookup — jcr

• Problem. Finding a documented command is treated as permission to run it.
• Without. The agent executes whatever the tree returned.
• With. One tool, nested capability tree, returns context, does not execute. Routing ≠ permission. Docs ≠ authority to run. 0.6 band is application policy.
• Measure. Lookup+explain only. Input/cost/wall-time deltas are not Harbor task-execution.

Prompted JSON vs structured read — localjev

• Problem. Parsed JSON from a generator is treated as a Noul.
• Without. Schema-valid output is taken as the picked-right probability.
• With. Prompted JSON ≠ structured logit read. Wire-compat ≠ logit-equiv. Schema-valid ≠ picked-right.
• Measure. Schema pass rate separately from calibration and from “did code take the right effect.” Entropy-as-confidence is not ECE.

Measurement honesty — hysteresis, ECE, Harbor

• Problem. A single threshold, an equal-width ECE, or a schema-pass is treated as Harbor.
• Without. 0.85 as a hard gate. Ranking as calibration. Silent provider failure as a policy outcome. Instruct-tuned confidence as honesty.
• With. Hysteresis {enter, exit} is policy attached to a probability. Report equal-width and quantile ECE. hop-ECE must not move under permutation (if it does, the instrument is the story). DecisionOps: ACT / REVIEW / FALLBACK; a provider failure is not a policy outcome. VERIFY needs discriminating evidence.
• Measure. Accuracy@0.5 vs ECE (they can disagree). Schema-pass vs joint fields. Quality denominators include only valid scored answers.

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