aoa-evals

name

aoa-evals

description

Central AoA proof or lifecycle: first tool turn reads only this SKILL.md. Use to select, review, or evolve source bundles, named eval results or verdicts, source-linked reports, Eval Forge owner review, or source/reader drift. Use aoa-eval for local selection, application, intake/design, or session-hit classification. Candidates, readiness checks, receipts, generated readers or outputs, MCP outputs, and green commands are not proof.

aoa-evals

Route one central bounded-proof question without letting a catalog, green command, readiness check, candidate, report, receipt, generated reader, MCP result, or KAG node outrank its source bundle and admitted evidence.

Applicability preflight

Before resolving the central owner, inspect only the request:

• if it asks to classify supplied or retrieved session hits against an eval trigger, route a local no-fit, design a local suite, or select/apply a repository eval, return owner_handoff to aoa-eval
• a packet being closed, reviewed, or provenance-bearing does not make it a central proof object
• begin this bundle only when the task already concerns central proof meaning, a central source bundle, a named eval result or verdict, an owner-routed proof object, or an aoa-evals lifecycle decision

Do not inspect the task workspace, search for an owner checkout, probe supplied paths, or run pwd, rg, find, or test for navigation after this bundle loads. Read the contract, source-return rule, and selected mode first; then use the declared bundled resource as the first navigation action.

Start

1. The first tool turn must read only this SKILL.md. Record <bundle_dir> as its absolute containing directory. If workspace inspection or search was combined with that read, stop the attempt and restart in a fresh invocation; later compliance cannot repair the unobserved package-first gate.

2. Read references/contract.yaml and choose exactly one internal mode:

   Mode	Select when	Read
   select	No exact central source bundle has been selected for the bounded proof need.	references/select.md
   review	A named central bundle, result, verdict, source-linked report, or already owner-routed proof object must be interpreted.	references/review.md
   evolve	The request asks to add, change, promote, supersede, or retire a central source, lifecycle, comparison, or chooser surface, even when a required gate may later block the operation.	references/evolve.md

   Select by the requested operation, not by whether its gates have already passed. Missing review, authority, or confirmation blocks evolve; it does not turn a requested owner change into review.

3. Read references/source-return.md and only the selected mode reference to EOF. Do not load the other mode procedures merely because their vocabulary appears nearby.

4. Follow the selected procedure and return the common output ABI from references/contract.yaml.

Keep the modes internal until held-out manual work proves that splitting them improves trigger precision, ABI, composition, and outcomes.

Owner boundary

• Source EVAL.md, eval.yaml, and admitted evidence own proof meaning.
• This skill owns procedure only. Its source receipt, catalog, scores, reports, readiness surfaces, receipts, generated readers, MCP, KAG, and green checks remain weaker.
• Use shared aoa-eval for cross-repository selection, exact local application, owner-local no-fit intake or design, and eval-trigger classification of supplied or retrieved session hits. Begin here only at central aoa-evals proof meaning or an owner decision.
• Technique references are optional bundle-local lineage, never a runtime prerequisite.
• select and review are read-only. evolve may change owner source only with explicit authority and required human confirmation.

Composition

Build only the task-local edges the request needs:

origin need or candidate evidence
  -> source proof contract and admitted evidence
  -> execution support and reviewed result when material
  -> derived readers or consumers when material

select may hand its exact source contract to review; an accepted review may later hand off to evolve. Do not force a three-stage workflow. A later node cannot repair or strengthen an earlier source or evidence gap.

An eval candidate ID beginning with session: remains an aoa-evals owner-routed object. Do not load session-memory guidance merely because of that prefix. Compose with session-memory only when the task actually requires raw, segment, session, or retrieval evidence beyond the supplied owner object.

Stops

Stop after one selection, bounded review, authorized source-first change, no-change result, owner handoff, or explicit blocker. Never execute an eval inside select, mutate inside review, infer owner acceptance, or treat a schema-valid packet or valid DAG as proof.

Keep raw trials, temporary rubrics, task-local DAGs, session notes, runtime state, and test scaffolding outside the owner skill home.