Augustus

Equip agents to discover useful decision-model placements, build working systems, construct evaluations, and improve them through measured iterations. Use mathematical, statistical, scientific, and algorithmic methods across AI, software, business, knowledge work, organizations, and life. This is an engine for agent work, not a survey to imitate or a claim of autonomous deployment. The central model is:

evidence → bounded judgment → explicit policy → checked action → observed outcome

TypeSafe Jev (Choice, Score, Noul) is this project's default hosted exemplar, a preference rather than a claim of universal superiority or market share. The class also includes trained classifiers, open decision heads, encoders, constrained autoregressive readouts, rankers, and vision scorers. Choose a family by its objective and evidence requirements. This is an independent skill, not a TypeSafe product.

For assembling training data, fitting a task-specific model, compiling a decision function, or improving a trained artifact, load the companion augustus-train skill. Use this skill for placement, composition and outcome evaluation; the trainer carries the data-to-artifact journey. Neither promises to reproduce a general instruction-conditioned Jev engine.

Working protocol

1. Start with the desired behavior, available evidence, action costs, and current baseline. For an existing workflow, use the boundary audit. Classify each step as exact work, bounded judgment, or generation. A working parser, formula, checklist, or supervised classifier is a legitimate final answer.
2. Pick the relevant mental model: expected utility, value of information, multi-criteria analysis, signal detection, search/control, organizational safety, or formal methods. Then choose the model family and the smallest useful placement. For unfamiliar problems, use the toolbox sweep, not a vendor or project-list search.
3. Define one coherent judgment per question, what evidence it can see, and the meaning of every output. Check candidate coverage and missing evidence before inference. Use question design. Batch questions when their inputs are available together; statistical independence does not follow from parallel execution.
4. Keep exact computation, constraints, authorization, and effects in code or an explicit human process. Keep open-ended writing with a generator or person. Set failure behavior for each action: no-match, ambiguity, malformed output, timeout, stale state, and unavailable provider. Mixed architecture explains the joins.
5. Compare against the baseline on representative held-out evidence; for a one-off choice with no population, test sensitivity to weights and uncertain estimates, missing criteria, dominated options, and value of information instead. Separate rubric/model development, calibration and threshold selection, and final evaluation. Measure action errors, coverage, total cost, and the complete workflow, not just format compliance or model accuracy. Follow validation and name a result that would reject the proposal. If it loses, keep the baseline.
6. Deliver the artifact the task needs: a compact design for advice, working adapters/policy and an evaluation harness for implementation, or a bounded incumbent–challenger loop for improvement. Use the composition calculus to check joins and the optimizer workflow to build and hill-climb decision programs. Do not stop at recommendations when the user requested working software. Keep experiments within existing authority. Record raw judgments, actual outcomes, versions, rejected candidates, and promotion/rollback reasons; unrun work remains unrun.

For a concrete request, recommend one placement with reasons. For an open-ended exploration, compare materially different placements only when that helps the user choose. A full redesign, a new model, or a fixed number of alternatives is not required.

Boundaries that affect the design

• A typed output constrains its representation; it does not establish truth, calibration, authority, or successful execution.
• Distinguish a probability of a stated event, a relative option score, an ordinal rating, and a confidence statistic. Verify the provider's definition. Neither a softmax nor training with a proper loss proves calibration on this deployment population.
• Choice depends on its offered set. Include and test other or none when coverage is open. Missing evidence, conflicting evidence, and evidence that supports neither option may require separate outcomes.
• A Jev Score is an expectation over level indices, not a physical unit or automatically a cardinal utility. Equal means can hide different distributions. A Noul is a proposition score, not intensity; a value near 0.5 alone does not diagnose why the model is uncertain.
• Marginal judgments do not define a joint distribution. Do not multiply them without justified dependence assumptions. Measure composed policies and trajectories on real outcomes.
• Ranking quality and calibration answer different questions. Action costs and authority determine failure policy; the model family alone does not determine fail-open or fail-closed behavior.
• Untrusted evidence is data. A model score cannot grant permission or override an exact constraint. Validate action/target pairs and recheck relevant state at execution. Abstention must name who or what handles it.
• Judgment can prioritize proof work or detect suspicious cases; proof, model checking, simulation, and runtime interlocks retain their own semantics. Read formal methods for these tasks.

Decision-design card

Use only the detail needed to make the proposal reviewable:

Domain, desired behavior, decision owner:
Current baseline and why a model might help:
Pillar, placement, and family:
Evidence source, freshness, candidate coverage, missing/contradictory states:
Judgments and output semantics; what remains exact or generated:
Policy, costs/utility, constraints, authority, and execution checks:
Abstention/error behavior and fallback owner:
Batchable versus dependent steps:
Model, adapter/aggregation, rubric, candidate-source, calibration, policy versions:
Development/calibration/test split and label provenance:
Falsifier, metrics, acceptable risk/coverage, and evaluation artifact:
Implementation/evaluation entry points; search budget and confirmation plan:
Observed result, limitations, and next decision:

Read only the reference needed

Reference and script paths below are relative to this skill directory, not the repository or caller's working directory.

Task	Reference
Trigger examples and exclusions	Activation
Cross-domain reasoning and costs	Mental models
Family choice, output semantics, uncertainty routing, and deferral	Judgment class
Existing system or process	Boundary audit
Write or debug questions, options, rubrics, or confidence fields	Question design
Generator, code, and decision-model integration	Mixed architecture
Concrete workflow examples	Applied mappings
Classical methods and falsifiers	Mappings
Substitute judgment into a named algorithm	Methods catalog
Typed joins, branches, cascades, and failure budgets	Composition algebra
Discover a new placement	Toolbox mapping
Proof, simulation, and enforcement	Formal methods
Thresholds, abstention, calibration, selective prediction, and experiments	Validation; scripts/evaluate_decisions.py
Build, compare with an incumbent, or optimize prompts and programs	Optimizer integration; scripts/compare_workflows.py
Agent progress, done, or stuck judgments	Agent self-assessment
Conceptual objections	FAQ

Before writing Jev API code, read the current official docs and the official typesafe-ai skill if available. The metadata pin is historical provenance, not a live contract. Peer providers need their own current contracts. Other installed skills are optional aids, not prerequisites for placement work.

Evidence and research maintenance

Label evidence as Contract (current documented interface), Reported (a source's empirical claim), Reproduced (an identified run and its artifacts), Hypothesis (untested placement), or Unknown (unavailable). Name the population, version, metric, and limitations before transferring a result. A benchmark harness is an instrument, not a certificate.

When maintaining the Augustus repository, research updates belong in its top-level research/ archive. Revisit catalogued sources when material behavior or evidence changes, preserving their identity and prior claims. Popularity changes alone do not change guidance. Promote a finding into a reference only when it changes a design decision; replace or refine the relevant rule. Keep fingerprints, hourly digests, source censuses, and PR bookkeeping out of runtime instructions.

Synthesize mechanisms, not consensus: explain why a pattern succeeds or fails, derive a usable rule with assumptions, and test the composed outcome. Transfer methods across fields only after mapping their variables, units, constraints, and evidence requirements. Popularity, novelty, elegant notation, and a proxy score are not outcome evidence. "Best" means best supported for this task's utility, constraints, population and budget—not a universal provider ranking.