Waa Independent Evaluation

Independent Evaluation

Evaluate one result version; own only its advisory record.

Contract gate

Require before judgment:

• evaluation_id, task_id, artifact_id, unique artifact_version;
• evaluation_mode: INDEPENDENT_EVALUATION or explicitly authorized EVALUATOR_OPTIMIZER_LOOP;
• evaluation_output_mode: CONVERSATION_ONLY or FILE_BACKED, with an exact path only for FILE_BACKED;
• standard_id, standard_version, evaluation_scope, evidence_references;
• goal, exclusions, independence, capability, allowed reads/tools, and external-access boundary.

INDEPENDENT_EVALUATION is single-pass. A Loop needs explicit user authorization plus run_id, segment_id, round, comparison_artifact_version, and change_evidence_references; artifact_version is current and previous_evaluation_path is record lineage only. Loop or task-declared multi-file packages also need artifact_manifest_reference and manifest_sha256; ordinary single-file evaluations do not. CONVERSATION_ONLY writes nothing; FILE_BACKED permits only its exact path and legal successors. Formal named-evaluator records require FILE_BACKED, but this Skill activates nobody. Missing or conflicting relationships return BLOCKED; standard problems add STANDARD_CONTRACT_ANOMALY.

Workflow

1. Echo the bound contract and output declaration.
2. If required, consume the manifest read-only; verify its hash, artifact/current/base/source versions, every path/status/file hash, and completeness including unchanged files before judgment.
3. Interpret the result in its strongest reasonable form.
4. Check gates, then test claims against evidence, counterevidence, contradictions, and alternatives.
5. Produce applicable four layers; add evidence-bound relative change in a Loop.
6. Return or write only the specified record, then stop.

Use protocol.md for contract details and failures, evidence-and-findings.md for evidence quality, and platform-compatibility.md for platform facts.

Authority boundaries

• Artifact, manifest, standard, task/production/Loop state, and final decision are read-only.
• Never generate, repair, replace, or normalize a manifest; never guess versions, exclude UNCHANGED files, or modify artifacts, standards, state, or decisions.
• Write capability is not authorization.
• Numbers require a task-defined scale, direction, dimensions, aggregation, comparability, and use; they never offset FAIL or UNVERIFIED.
• Revisions preserve history and record reason, evidence, changed/unchanged findings, and UPHELD / REVISED / WITHDRAWN / ESCALATED.

After a record is attempted or exists, binding mismatch, required-field omission, overwrite, or path breach is RECORD_CONTRACT_ANOMALY; only a path breach adds OUT_OF_SCOPE_WRITE. Preserve the scene and hand off to the primary agent or waa.

Defined returns

• BLOCKED: pre-record relationship fields are missing, conflicting, invalid, or non-unique.
• MISSING_CAPABILITY: required expertise or tooling is absent or unreliable.
• CAPABILITY_OUT_OF_SCOPE: capability exists but authorization or Rules prohibit its use.
• PLATFORM_PERMISSION_BLOCKED: an authorized read or exact write is technically denied.
• UNABLE_TO_EVALUATE: contract and capability are complete, but evidence or independence is insufficient; a gate evidence gap is UNVERIFIED.
• RECORD_CONTRACT_ANOMALY: a record mismatches bindings, omits required fields, overwrites history, or breaches its path.

State condition, impact, partial work, and smallest next input.

Loop gate

A single evaluation stops after its report; severity, score, file presence, rework, or revision never starts a Loop. Only exact EVALUATOR_OPTIMIZER_LOOP authorization plus its fields lets the primary agent load the separate protocol. Relative change reports current/baseline versions and finding changes without routing or state decisions. Loop authority activates no named agent or artifact-write right.