Allura-Ecosystem Policy-Replay Review
Use this skill only when the employee supplies an approved policy version, deterministic replay output, or a system-produced expected-versus-actual report.
Safety boundary
- The supplied policy/replay system remains authoritative; do not replace it with model reasoning.
- Do not make a credit decision, recommend a loan action, or determine a legal notice reason.
- Do not invent policy rules, thresholds, or interpretation when the source is absent.
- Escalate unclear results to the authorized underwriter, auditor, or compliance owner.
Workflow
- Identify the policy version, replay run identifier if supplied, and source materials.
- For each reported rule, compare the expected result, actual result, and cited evidence.
- Classify the comparison as
match,mismatch,insufficient evidence, orsource unclear. - Preserve exact source wording for rule IDs and thresholds; do not normalize or repair identifiers. Cite provenance per
references/evidence-metadata.md. - Produce this table:
| Rule / check | Expected | Actual | Evidence cited by source | Comparison | Human review question |
|---|
- State explicitly that the output summarizes a supplied replay and does not determine the loan outcome.
Handoff
Use audit-packet-draft to create a human-review packet from the evidence and replay comparison.
Additional resource
references/safety-boundaries.md— fixed authority and data-handling boundaries.references/evidence-metadata.md— provenance fields and rendering rules.