aoa-eval

Intent

Route one evaluation need to the narrowest procedure that preserves its real owner, input contract, output, and proof limit. Consolidation here is a shared front door, not a collapse of five different capabilities into generic prose.

Trigger boundary

Use this skill when:

• an eval surface must be selected or applied
• an established no-fit needs a local intake packet or a bounded suite design
• researched eval-trigger classes and supplied or retrieved session hits need

eval-owned candidate classification

Do not use this skill when:

• the task is an ordinary test run, the invariant is still unknown, the real

need is generic session harvesting, or the requested effect is central proof promotion

Inputs

• one mode intent plus its owner, source ref, acceptance target, relevant

evidence or inventory, effect authority, and required environment details

Outputs

• one mode-specific result defined by references/contract.yaml, including

owner, evidence class, proof limit, next route, and stop line

Procedure

Record <bundle_dir> as the directory containing this loaded SKILL.md.

Choose exactly one mode

Mode: select

Read and follow references/select.md.

Mode: apply

Read and follow references/apply.md.

Mode: local-need

Read and follow references/local-need.md.

Mode: design

Read and follow references/design.md.

Mode: session-mining

Read and follow references/session-mining.md.

1. Read references/contract.yaml and the selected mode reference completely

before acting. Do not load unrelated mode references.

1. Read the target owner's route law and preserve source, environment,

freshness, evidence class, effect posture, and proof authority.

1. If a required mode input cannot be obtained from permitted sources, stop as

blocked_missing_input; do not guess, substitute a broader action, or relabel absence as an owner-boundary deferral. In apply, the mode reference's complete dotted-field preflight must be observed before reading or executing the selected target; a summary claim that the contract is complete is not a preflight.

1. Execute each chosen procedure as one task-local node and return its typed

output. A request that explicitly asks to find and run may form select -> apply only after an exact fit provides the complete apply ABI. Other later modes remain explicit handoffs, never automatic continuation.

1. When the task asks what an applied observation actually constrains, hand the

evaluation-observation to aoa-verification; do not reinterpret command success as proof inside this bundle.

1. When the task explicitly asks for central proof meaning, a central verdict,

admission, or lifecycle interpretation, preserve the observation and its source, claim class, environment, and proof limit in a central-proof-review-request, then hand it to aoa-evals-skills. Do not invoke that owner merely because an eval ran successfully.

1. When prior-session evidence still must be found, let the session-memory

evidence route retrieve and ground it, then consume the bounded packet here. The evidence route owns retrieval; session-mining owns eval-trigger classification and its local eval handoff.

Contracts

• selection, execution, intake, design, session evidence, and proof promotion

are different effects

• eval discovery/application and interpretation of invariant coverage are

distinct but composable nodes

• local evidence stays local until a stronger proof owner accepts it
• Eval Forge health, command success, and generated dashboards are not result

proof

• .aoa evidence is optional candidate evidence and never overrides owner

source truth

• technique records may explain provenance but are not runtime dependencies

Risks and anti-patterns

• choosing by keyword or replacing the selected action with a broader green

gate

• collapsing local pressure and executable design into one vague proposal
• mining sessions before trigger classes and owner routes are known
• creating a permanent validator before manual cases establish a durable

invariant

Verification

• confirm one mode, exact owner/source/environment posture, required inputs,

output type, effect, and termination condition

• inspect commands and artifacts manually against the declared acceptance

target

• state what the result proves, what it does not prove, skipped checks, drift,

and the next owner route

Adaptation points

Owners supply local eval ports, schemas, command runners, artifacts, acceptance criteria, privacy rules, session providers, and proof-promotion workflows.