Isolated Evaluator

Single-option evaluator skill for delegated subagent runs. It evaluates exactly one target alternative and returns structured output that a parent workflow can merge into comparative scoring input.

When to Use

• A parent workflow needs one evaluator per alternative.
• You need strict evaluator isolation for auditability.
• You need a deterministic handoff artifact for later harness scoring.

Inputs

Use references/input-schema.md and references/input-schema.json as the source of truth.

Required:

• decision
• criteria_confirmed
• criteria_confirmation_source (when criteria are confirmed)
• current_platform
• criteria
• target_alternative
• evaluator_id

Optional:

• major_platforms
• score_scale
• known_evidence
• assumptions

Workflow

1. Validate input and isolation boundaries.
   • Confirm exactly one target_alternative.
   • Reject or ignore peer alternatives, rankings, and recommendation hints.
2. Evaluate criterion-by-criterion and platform-by-platform.
   • Score with the declared scale.
   • Use null for missing evidence, never zero-fill.
   • Track evidence references for each meaningful claim.
3. Write isolated rationale.
   • Produce option-level justification.
   • Produce evaluator summary focused on the single target.
4. Emit output using references/output-schema.md and references/output-schema.json.
   • Include evaluated_alternative (scores + justification).
   • Include independent_evaluation with isolation_confirmed=true.
5. Report evidence gaps and assumptions for parent-level aggregation.

Output Contract

evaluated_alternative must preserve the comparative workflow alternative shape:

• id
• name
• effort
• risk
• feasible
• scores
• justification

independent_evaluation must preserve the comparative workflow evaluator-record shape:

• alternative_id
• evaluator_id
• isolation_confirmed
• summary
• evidence_refs

Guardrails

• No cross-option comparison language.
• No ranking or final recommendation output.
• All missing evidence stays explicit (null scores + gap notes).
• Every score and summary claim should be traceable to evidence.
• Apply a brief agentic-eval reflection on the evaluator summary (max 2 iterations) to verify evidence coverage.
• Keep critique output structured (JSON) and do not emit separate analysis files.

Apply references/quality-gates.md for pass/fail checks.

Optional Validation

python3 -m jsonschema -i <isolated-eval-input.json> skills/workflow/isolated-evaluator/references/input-schema.json
python3 -m jsonschema -i <isolated-eval-output.json> skills/workflow/isolated-evaluator/references/output-schema.json

References

• Input contract: references/input-schema.md
• Input schema: references/input-schema.json
• Output contract: references/output-schema.md
• Output schema: references/output-schema.json
• Quality gates: references/quality-gates.md
• Output template: assets/isolated-evaluation-template.v1.json