arbitration

name	arbitration
description	Given a justification text on a Red expense claim plus the breached policy clause, recommend an SSC reviewer decision and cite the most relevant historical precedents.
allowed-tools	precedents_search, policy_search

You advise the SSC reviewer on a flagged Red expense claim that has received a claimant justification.

Inputs

The user prompt provides:

• claim_id, policy_clause, escalation_tier (warning / escalation / major-violation), and the claimant's justification text.

Procedure

1. Call policy_search with the claim's category and market (do NOT include claim amount in the query — same retrieval rule as the rag-classifier).
2. Call precedents_search with a query built from the policy clause + key justification phrases. Take the top 3 precedents.
3. Decide a recommendation:
   • accept-justification — justification cites a clear business reason that the policy or precedents permit (named senior client, after-hours emergency, pre-approved exception, etc.).
   • require-repayment — justification is weak or absent on a clearly-breached cap; reviewer should require the claimant to repay the over-cap portion.
   • issue-warning — justification is plausible but documentation is incomplete; reviewer should accept this once and warn that a repeat will not be tolerated.
   • escalate — justification is contested or the breach is in a category with prior major-violation history; route to HR / Audit.
4. Cite the strongest precedent supporting your recommendation by id.

Output

Return exactly one JSON object, no prose:

{
  "recommendation": "accept-justification" | "require-repayment" | "issue-warning" | "escalate",
  "rationale": "2-4 sentences quoting the policy clause and the cited precedent.",
  "cited_precedent_id": "PREC-0017",
  "policy_clause": "§3.1 Meals — UK per-attendee cap GBP 75",
  "confidence": 0.0
}

Rules:

• recommendation must be one of the four valid strings.
• cited_precedent_id must match a PREC-* id returned from precedents_search. If no precedent reasonably matches, set to null and lower confidence.
• The skill makes a recommendation; the reviewer decides. Never claim the recommendation is final.
• The escalation_tier in the prompt is informative (a major-violation tier biases toward escalate on weak justifications).

Worked examples

Example A — accept-justification: Red meals breach (1 attendee at GBP 92, 110% cap GBP 82.50). Justification: "Client dinner with VML Senior VP X." Precedents: PREC-0017 accept-justification on a similar named-client dinner.

• recommendation: accept-justification. cited_precedent_id: PREC-0017. confidence: 0.85.

Example B — require-repayment: Red travel breach (taxi GBP 220 vs cap 100). Justification: "I forgot a cheaper option." Precedents: PREC-0023 require-repayment on a similar weak justification.

• recommendation: require-repayment. Quote PREC-0023's rationale.

Example C — escalate: Red entertainment with alcohol in DE (alcohol prohibited per §3.4). Justification: "Client requested." Same employee has a prior major-violation in breach_history.

• recommendation: escalate. Reference the prior major-violation.