Augustus v0.7.0

Released 2026-09-22. This minor release equips agents to find, build, evaluate, and improve decision-model systems, with new paired-outcome tooling and a more explicit composition calculus. It adds no provider integration or model and makes no claim of measured model superiority. TypeSafe Jev remains the default hosted exemplar; exact rules, classical models, and human processes remain valid alternatives.

The important shift

Augustus is a skill and working method for agents to discover useful decision models, implement compositions, build evals, and hill-climb decision-driven or Software 3.0 systems. Research is input to that work, not its finished output. The goal is to derive and test the best-supported method for the actual task, not reproduce a consensus survey or chase a fashionable benchmark.

That capability is now named directly in the skill’s activation description, Codex UI, Claude marketplace, README, and website. Discoverability should match what users want to accomplish: build and improve working systems with outcomes and explicit constraints—not merely read about model placement.

What changed

• Composition-to-implementation loop. Typed evidence contracts, conditional cascade and branch risk, dependence-safe budgets, decision-regret bounds, information substitution limits, and trajectory-level evaluation. Cross-field analogies must supply their assumptions and a falsifier.
• Executable outcome comparison. A standalone paired-workflow evaluator separates descriptive search from fixed-sample confirmation, flags observed constraint violations, and keeps unknown costs explicit. Fixtures and proxy results are not promoted to outcome evidence; the helper never deploys a policy.
• Bounded improvement. Agents retain a runnable incumbent, build evals before hill climbing, protect confirmation data, preserve rejected trials, and use authorized rollout and rollback. Research informs mechanisms, not a median survey.
• Primary-source refresh. Reviewed Jev-Omni, decision-model-testing, pijev, official contracts, selected ecosystem changes, and recent decision research. Traversed all 400 tracked Jevusers entries and 1,299 app rows. Directory traversal and GitHub/Hub metadata discovery are not code audits.
• Precise score and evidence boundaries. Version wrapper and aggregation semantics; distinguish convex-loss averaging from calibration; measure actual media coverage, accepted-case support, injection controls, and complete workflow cost and latency.
• Independent re-derivation. Earlier delegated scopes were redone with requested GPT-6 Astra/xhigh agents. Corrected unnecessary training/filtering requirements, a blanket calibration prerequisite, post-action enforcement claims, Alloy temporal limits, and distillation overstatements.
• Regression protection. The offline suite grows from 53 to 106 tests. Skill scenarios grow from 12 to 31, with fresh answers reviewed by the coordinator. Structural checks and behavioral smoke tests are distinct. Adversarial review additionally caught maximum-float overflow, subnormal aggregation/delta errors and log-loss cancellation; exact arithmetic oracles, randomized boundary probes and fixed-sample null checks now cover those paths. Fable follow-up notes drove strict-bound/report consistency, stable Brier/ECE aggregation, explicit exclusions and fixture-preserving violation labels.
• Discoverability checks. Corrected About metadata, verified current directory presence and page layout, fixed encoded fragments, and added noindex checks. Audited existing awesome-list entries and drafted targeted positioning updates; no third-party PRs, issues or corrections were submitted by that audit. Installation, implicit activation, indexing, and real benefit remain separate claims; listings alone do not establish growth.

Upgrade notes

Update through your existing installation method; avoid duplicate installations. The skill needs no API key. No provider benchmark was reproduced; the bundled offline helpers make no model calls. Model-based code review is separate.

The existing offline helpers now reject duplicate JSON keys, conflicting known repository node IDs, and invalid schema-version types. Oversized numbers, decoder nesting failures, and interrupted HTTP reads produce explicit input or receipt errors. Large finite policy costs are averaged without multiplying counts into an avoidable overflow. Valid-input interfaces remain unchanged; partial baselines still carry no comparative claim.

Run make check from a source checkout. Runtime references must stay within the installed skill, including nested cards; budgets, links, and SemVer are checked. Rendered-site validation checks asset existence, not image decoding or CSS URLs. Use the explicit GitHub Pages gem activation in CONTRIBUTING for a matched build.

Evidence and limits

Read the refresh and source dispositions, replacement ledger and integration evidence, adversarial findings and corrections, discoverability audit, existing-listing update queue, and maintainer guidance. These dated artifacts preserve what was verified at each stage; final publication and deployment are recorded on the release PR. Requested model routing is not independent backend attestation. Public projections are not private benchmark replay; no source result was upgraded to Reproduced without a local run.

Earlier release: v0.6.0.