Skip to content

Augustus 0.7.0 — build, evaluate, and improve decision-model systems

Choose a tag to compare

@24601 24601 released this 22 Sep 18:13
· 197 commits to main since this release
ef8e035

Build, evaluate, and improve decision-model systems

Augustus is an agent skill and working method for finding useful decision models,
implementing compositions, building evals, and iteratively improving decision-driven
and Software 3.0 systems. Research informs mechanisms and falsifiable experiments;
it is not a vendor catalog or a substitute for working outcomes.

This minor release makes that mission explicit in the skill, README, Codex UI,
Claude marketplace and website, and adds executable outcome-comparison capability.
The planned 0.6.1 patch was not published. TypeSafe Jev remains the default hosted
exemplar; exact rules, classical models, human processes and no new model remain
valid design choices.

Added and improved

  • Typed composition contracts and conditional laws for branch/cascade risk,
    failure budgets, bounded-loss decision regret, information substitution and
    feedback trajectories—with assumptions and counterexamples.
  • A standalone paired-workflow evaluator separating descriptive search from
    fixed-sample confirmation. Unknown costs remain unknown; fixtures and proxies
    stay labeled; separate constraints and deployment authority remain external.
  • A bounded build/evaluate/improve loop: runnable incumbent, explicit adapter and
    policy, outcome capture, independent confirmation, regression checks and rollback.
  • A full primary-source refresh including Jev-Omni, decision-model-testing, pijev,
    official contracts, selected implementations and recent research. Traversed
    400 tracked Jevusers entries and 1,299 app rows; traversal is not a code audit.
  • Fresh re-derivation of earlier delegated scopes, corrected overgeneralized
    requirements, stronger provenance/revisit/release guidance, and targeted
    positioning-update drafts for existing listings. No third-party submissions
    were made by the listing audit.

Reliability and upgrade notes

The offline suite grows from 53 to 106 tests. Astra reviews exposed four
numerical failure modes that passing initial tests missed: maximum-float mean
overflow, mixed selective-cost overflow, log-loss cancellation and premature
normalization of paired deltas. Exact-ratio arithmetic and stable logarithms now
have boundary, randomized-oracle and fixed-sample null regressions.
Fable's independent review additionally led to strict-bound/report consistency,
order-stable Brier/ECE aggregation, explicit exclusions, fixture-preserving
violation labels and build/improve activation examples.

Existing helpers reject duplicate JSON keys, conflicting known source identities
and invalid schema-version types. Malformed JSON, huge numbers and interrupted
HTTP responses have explicit failure handling. Structural checks cover installed
skill containment, recursive reference budgets, strict versions, fragments and
accidental noindex directives. Structural lint is not semantic acceptance.

Update through your existing installation method; avoid duplicate installations.
The skill needs no API key, and its offline helpers make no model calls. No new
provider integration is bundled and valid-input interfaces remain compatible.

Evidence and limits

Forty actual behavioral answers across recorded runtime snapshots cover 31 catalog
scenarios and two additional requests, including a runnable synthetic eval build.
Repeated answers are not independent benchmark units. No provider benchmark was
reproduced and no native activation, user-growth or deployment benefit is claimed.
Recurring reassessment targets are documented; the external research scheduler
remains unverified. Public listing presence is distinct from useful adoption.

Dated pre-publication audits retain their historical state; final review,
installation, CI, tag and deployment receipts are recorded on the release PR.