Augustus v0.6.0

Released 2026-09-22. This release restores the skill’s core workflow: evidence → bounded judgment → explicit policy → checked action → observed outcome. TypeSafe Jev remains the default hosted exemplar; the method applies across model families and software, business, organizational, and everyday decisions.

Highlights

• Focused guidance. The entry point shrank from approximately 192 KB to 9 KB, with 17 task-routed references. Historical research is preserved by exact Git revision and a verified 30-file hash manifest.
• More precise decision design. Clear boundaries distinguish ranking, probability, confidence, calibration, abstention, causal interventions, optimization, and authority. Current research is attributed and bounded by its assumptions; unrun experiments remain explicitly unrun.
• Meaningful verification. Structural lint, content budgets, 53 regression tests, numerical self-tests, independent behavioral review, and CI replace repeated-prose locks and tautological assertions. Twelve realistic skill scenarios include no-model and non-trigger cases.
• Safer research maintenance. New promotion guidance requires a changed design decision. Refresh collection is bounded and read-only; it never stages, commits, pushes, clones projects, or overwrites a receipt.
• Better installation and onboarding. Fixed Claude marketplace metadata, Codex skill presentation, six canonical placements, two worked examples, feedback intake, social previews, sitemap, and mobile layout checks.

These checks establish repository and installation behavior, not measured model superiority, real-world decision improvement, search rank, or growth.

Upgrade notes

Reinstall or update Augustus through your existing installation method; avoid installing duplicate copies through multiple methods. Claude Code uses /augustus:augustus; Codex uses $augustus. The skill itself needs no API key. Hosted model calls remain an optional, separately configured integration.

If you use the offline helpers directly:

• Pass --cost-fp and --cost-fn explicitly for cost reports; add --cost-abstain for selective-policy cost. Use --lower-threshold and --upper-threshold for an abstention band. Fit thresholds on calibration data and evaluate the frozen policy on a separate split.
• Complete binary predictions use action_rate; selective coverage is a different measure. All-abstain selective error is undefined, not zero. Impossible observed probabilities have infinite log loss. Duplicate IDs and invalid probabilities are rejected. Partial baselines are accepted but excluded from comparison; comparative reports require a complete baseline.
• Refresh wrappers now emit JSON receipts to standard output. Use --output for a new receipt file; existing files are never overwritten. Update any scheduler expecting appended logs or automatic Git writes.
• Run make check from a source checkout for the complete check suite. The old uniqueness-gate entry point runs only the structural checker; historical repeated research passages are no longer required.

Evidence and earlier releases

Read the integration review, primary-source research, and discoverability audit. These are dated pre-release evidence; their descriptions of unperformed publication or deployment refer to the audit time, not the current release. Native-host implicit activation and real-user outcome/growth measurements remain follow-up work.

Earlier release: v0.5.1.