Eligibility Check (Govern)
Purpose
The compliance determination IS the work: eligible / ineligible / needs_evidence with the clause cited (#2.1), plus fraud signals surfaced - never adjudicated (#3).
When to use
- After reason-classify in every case; or when asked whether a return qualifies or has red flags.
When NOT to use
- Fetching policy wording or windows - policy-retrieve.
- Structuring the request - return-intake. Reason coding - reason-classify.
- Writing the packet, customer response or exception request - case-packet.
- Aggregate reporting (return rates, dashboards) - out of scope.
Where the data comes from
This plugin ships no sample data. Return windows and limits (config/return-windows.json) and the policy text (references/returns-rules.md) are retailer configuration inside this skill folder. The case is the user's own content:
classified.json written earlier in the session by reason-classify, or a classified case file the user names, links or attaches in the chat.
- Their OneDrive or SharePoint, found with Work IQ search (Microsoft 365 search) by case ID, order ID or file name when no location is given.
- If the user gives only the case facts (chat, service-desk note, customer email, order/receipt lookup export, return-history export), build
case.json with return-intake and classified.json with reason-classify first.
Read each source with the file reader (ReadFileContent) and cite it by file name and row or line. Write the working JSON file (governed.json) to the session working folder, never into the plugin. If a source is missing (for example no order number or purchase date on record), say which one and ask the user for it. Never invent it, guess a date or fall back to example values; without an order record the engine returns needs_evidence (RET-1.2).
Inputs
Configuration ships inside this skill folder (synced from the plugin-root shared/ copy before packaging): config/, contracts/, references/returns-rules.md - paths below are relative to this skill folder.
classified.json (from reason-classify in this session, or the user's file - see Where the data comes from) + config/return-windows.json. Engine: scripts/eligibility_check.py. Schema: contracts/rtl.returns-refund-case.v1.json.
Steps
- Validate
classified.json against the contract (required keys, contract_version, reason present).
- Run
python scripts/eligibility_check.py --classified classified.json --config config/return-windows.json --out governed.json (classified.json and governed.json live in the session working folder).
- Quote
determination, clause, detail, missing_evidence and every fraud_signals[] entry verbatim. No order record -> the engine returns needs_evidence (RET-1.2) - relay that the lookup must land before any decision.
- Any fraud signal or confidence < 0.75 ->
hold_for_review: route to asset protection / manager (#4.1, #4.2). Present signals factually, no accusations.
- A sympathetic story (tenure, loyalty, "the manager did it last time") is context for a HUMAN exception (#5.1): say so, and hand the exception request to case-packet.
Example
$ python scripts/eligibility_check.py --classified classified.json --config config/return-windows.json --out governed.json
eligibility_check: ineligible (RET-2.1) rec=hold_for_review signals=2 -> governed.json
Reply sections: Determination | Recommendation | Missing evidence | Fraud signals (table: signal | detail | rule | source)
"Determination: ineligible - RET-2.1 - 20 days since purchase > 15-day electronics window (#1.1).
Recommendation: hold_for_review (2 signals, #4.2). Signals: serial_mismatch (#3.1), high_value_receiptless (#3.3).
The customer's loyalty story is context for a manager exception (#5.1) - drafted by case-packet, never granted here.
DRAFT - a human authorises any refund, exchange or decline."
Output format
governed.json - payload plus eligibility: {determination, clause, detail, missing_evidence[], recommendation, confidence, source, citation} and fraud_signals[]. In chat: one line each for determination + clause, recommendation, missing evidence, and a factual signal list; end with "DRAFT - a human authorises any refund, exchange or decline".
Grounding requirements
Every determination cites its clause and config; every signal cites the record it came from.
Guardrails
- Never fabricate a determination, clause, confidence or signal; every value is quoted from governed.json and cites its record.
- The engine decides; the model never grants an exception (#5.1), never drops or adds a signal, never softens ineligible to "probably fine", never edits governed.json by hand.
- Never authorise a refund, release funds, approve store credit, or add a customer to any list.
- Never write "fraud", "fraudster" or "confirmed" about a customer - signals are signals.
- Engine error or contract mismatch: stop and report; do not hand-compute a determination.
Escalation / uncertainty
If the engine fails or the contract does not validate, stop and report - never hand-compute a determination or fabricate a clause. If order is missing, relay the engine's needs_evidence (RET-1.2) and name the lookup that must land.
Signals, missing evidence or low confidence -> hold_for_review with the packet routed to asset protection / manager.
1---2name: eligibility-check3description: 'Determines return eligibility against the configured policy and surfaces fraud signals deterministically with the eligibility_check engine - the explicit Govern step of this plugin. Use after reason-classify, or on "is this returnable", "do they qualify for a refund", "check eligibility", "any fraud signals", "can we accept this return". Do NOT use to retrieve policy text - use policy-retrieve instead; not to start a case - use return-intake; not to draft the packet or customer response - use case-packet.'4license: Proprietary5metadata:6 source: https://skillmd.com/skills/microsoft/eligibility-check7 version: 1.0.48 author: Microsoft Retail & CPG Skills9 category: analysis10 icon: ShieldCheckmark11---1213# Eligibility Check (Govern)14## Purpose15The compliance determination IS the work: eligible / ineligible / needs_evidence with the clause cited (#2.1), plus fraud signals surfaced - never adjudicated (#3).16## When to use17- After reason-classify in every case; or when asked whether a return qualifies or has red flags.18## When NOT to use19- Fetching policy wording or windows - policy-retrieve.20- Structuring the request - return-intake. Reason coding - reason-classify.21- Writing the packet, customer response or exception request - case-packet.22- Aggregate reporting (return rates, dashboards) - out of scope.23## Where the data comes from24This plugin ships no sample data. Return windows and limits (config/return-windows.json) and the policy text (references/returns-rules.md) are retailer configuration inside this skill folder. The case is the user's own content:25- `classified.json` written earlier in the session by reason-classify, or a classified case file the user names, links or attaches in the chat.26- Their OneDrive or SharePoint, found with Work IQ search (Microsoft 365 search) by case ID, order ID or file name when no location is given.27- If the user gives only the case facts (chat, service-desk note, customer email, order/receipt lookup export, return-history export), build `case.json` with return-intake and `classified.json` with reason-classify first.2829Read each source with the file reader (ReadFileContent) and cite it by file name and row or line. Write the working JSON file (`governed.json`) to the session working folder, never into the plugin. If a source is missing (for example no order number or purchase date on record), say which one and ask the user for it. Never invent it, guess a date or fall back to example values; without an order record the engine returns needs_evidence (RET-1.2).3031## Inputs32Configuration ships inside this skill folder (synced from the plugin-root shared/ copy before packaging): `config/`, `contracts/`, `references/returns-rules.md` - paths below are relative to this skill folder.33`classified.json` (from reason-classify in this session, or the user's file - see *Where the data comes from*) + config/return-windows.json. Engine: scripts/eligibility_check.py. Schema: contracts/rtl.returns-refund-case.v1.json.34## Steps351. Validate `classified.json` against the contract (required keys, `contract_version`, `reason` present).362. Run `python scripts/eligibility_check.py --classified classified.json --config config/return-windows.json --out governed.json` (classified.json and governed.json live in the session working folder).373. Quote `determination`, `clause`, `detail`, `missing_evidence` and every `fraud_signals[]` entry verbatim. No order record -> the engine returns `needs_evidence` (RET-1.2) - relay that the lookup must land before any decision.384. Any fraud signal or confidence < 0.75 -> `hold_for_review`: route to asset protection / manager (#4.1, #4.2). Present signals factually, no accusations.395. A sympathetic story (tenure, loyalty, "the manager did it last time") is context for a HUMAN exception (#5.1): say so, and hand the exception request to case-packet.40## Example41```42$ python scripts/eligibility_check.py --classified classified.json --config config/return-windows.json --out governed.json43eligibility_check: ineligible (RET-2.1) rec=hold_for_review signals=2 -> governed.json44Reply sections: Determination | Recommendation | Missing evidence | Fraud signals (table: signal | detail | rule | source)45 "Determination: ineligible - RET-2.1 - 20 days since purchase > 15-day electronics window (#1.1).46 Recommendation: hold_for_review (2 signals, #4.2). Signals: serial_mismatch (#3.1), high_value_receiptless (#3.3).47 The customer's loyalty story is context for a manager exception (#5.1) - drafted by case-packet, never granted here.48 DRAFT - a human authorises any refund, exchange or decline."49```50## Output format51`governed.json` - payload plus `eligibility: {determination, clause, detail, missing_evidence[], recommendation, confidence, source, citation}` and `fraud_signals[]`. In chat: one line each for determination + clause, recommendation, missing evidence, and a factual signal list; end with "DRAFT - a human authorises any refund, exchange or decline".52## Grounding requirements53Every determination cites its clause and config; every signal cites the record it came from.54## Guardrails55- Never fabricate a determination, clause, confidence or signal; every value is quoted from governed.json and cites its record.56- The engine decides; the model never grants an exception (#5.1), never drops or adds a signal, never softens ineligible to "probably fine", never edits governed.json by hand.57- Never authorise a refund, release funds, approve store credit, or add a customer to any list.58- Never write "fraud", "fraudster" or "confirmed" about a customer - signals are signals.59- Engine error or contract mismatch: stop and report; do not hand-compute a determination.60## Escalation / uncertainty61If the engine fails or the contract does not validate, stop and report - never hand-compute a determination or fabricate a clause. If `order` is missing, relay the engine's needs_evidence (RET-1.2) and name the lookup that must land.62Signals, missing evidence or low confidence -> hold_for_review with the packet routed to asset protection / manager.