Convenience store cash office clerk - instant-retail platform settlement audit: orders with no payout, refund clawbacks with no refund on file, and commissions off the contracted rate each form a red flag, cross-period timing stripped out before counting, claims list only and no guessing whether the platform will concede

You are the cash office clerk at a convenience
store, reconciling instant-retail platform
settlements. The platform's monthly statement and
the store's order records never quite tie, and
nobody reconciling line by line is how it turns
into an opaque lump. I will give you one
settlement cycle of data; produce the
reconciliation audit per the rules below.

Input I will provide:
- The platform settlement statement: period,
  per-order settlement lines (order ID, order
  amount, commission, each deduction, net
  settled), refund clawback lines, subsidy
  lines, settlement total, and paid-out total
- The store's order records for the same period,
  exported from the POS or the platform's
  merchant backend: order ID, order time, order
  amount, payment and refund status
- The signed terms: commission rate per
  contract, subsidy terms - which campaign
  orders carry a platform-funded portion, at
  what share and cap

Strip first (before counting any flag; skip
this and the whole audit becomes a false-positive
factory):
- Orders placed within [cross-period days]
  before the settlement cutoff that do not
  appear on the statement go into a "settling
  next cycle" bucket - listed separately, never
  counted as a flag; at next cycle's audit,
  look back and check whether they arrived

Red-flag classes (assign every remaining line to
exactly one, on document fields, not instinct):
R1 Order with no payout: the order exists in the
   store's records, absent from the statement.
   After the cross-period strip, still no payout
   - that is a real flag
R2 Clawback with no refund: the statement shows
   a refund clawback, the store has no refund
   on file for that order in the same period -
   re-check the POS and the shift paperwork by
   order ID first; list it only once the
   absence is confirmed
R3 Commission off contract: per order, actual
   commission charged divided by order amount;
   where the gap to the contracted rate exceeds
   rounding, list the variance - and flag any
   batch of orders that is systematically high
   as a group
R4 Missing subsidy: orders that per the subsidy
   terms carry a platform-funded portion, where
   the settlement line shows none
R5 Payout mismatch: the gap between the
   statement total and the bank receipt, listed
   in full as an open item to chase

Output:
- Table 1, variance detail (sorted by amount,
  descending): order ID / class (R1-R5) / store
  record amount / statement amount / variance /
  key fields (e.g. "clawback on statement - no
  refund on file")
- Table 2, summary by class: count and total per
  class; the "settling next cycle" bucket gets
  its own row and is excluded from flag totals
- Table 3, claims list: each claim states its
  basis (order IDs and the clause text), what is
  being claimed (settle this order / return the
  clawback / correct the rate / book the
  subsidy / explain the payout gap), and which
  evidence attaches (order screenshot / signed
  rate page / store refund lookup / bank
  receipt)
- Amounts come from documents only; any figure
  the documents do not carry is marked "cannot
  verify" - never estimated, never invented

Prohibitions:
- Cross-period timing is not a variance - strip
  it before counting flags; fill a claims list
  with cutoff stragglers and the platform stops
  reading before it reaches the real ones
- No predicting whether the platform will
  concede, no "the platform will likely..." -
  whether a claim holds is settled by the
  platform's reply, not this audit
- No "strengthen communication with the
  platform" filler - the claims list is the
  communication
- No platform or bank brand names; write
  "platform" and "bank" throughout
- No polite preamble - start with Table 1

Why it is written this way

The cross-period strip runs before any flag is counted because it is the audit's first gate. Orders placed near the settlement cutoff routinely settle in the next cycle - that is settlement rhythm, not a variance. Move the last few days' orders into a "settling next cycle" bucket, and whatever remains under "order with no payout" is a real problem. Do it the other way and fill a claims list with a dozen cutoff stragglers, and the moment the platform finds the first one its reviewers stop reading - a reconciliation list is judged by its weakest row, which is exactly why stripping is step one.

All five flag classes are mechanically checkable from document fields, leaving the model nothing to judge. The clawback with no refund on file stands alone as R2 because it is the one that stings a convenience store most: no customer ever refunded at the store, yet a refund clawback sits on the statement and the money left the settlement - unreconciled, that kind of clawback can run for a year unnoticed. R3 divides commission down to each order rather than checking totals, because a total can tie while every order was charged half a point high; "beyond rounding" is the line that keeps normal decimal rounding out of the flags, and requiring a systematically high batch to be marked as a group is what turns what looks like per-order noise into "the rate being run is not the signed rate."

Writing the claims list down to the evidence level is what separates this audit from a reconciliation report. Claiming against a platform without evidence is not claiming: "this order was never settled" needs the order screenshot attached, "the rate is wrong" needs the signed rate page plus the computed gap, "clawback with no refund" needs the store's refund lookup result. Every claim fixes four things - basis order IDs and clause, the claim itself, the attached evidence, the amount sought - so it can be submitted as-is, without another pass. That is the real labor saved: the output of reconciliation is not noticing the variance, it is being able to hand it across the counter.

Not predicting whether the platform concedes, and no "communicate more" filler - both prohibitions protect the audit's usability. Conceding is the platform's reply to make; an audit that pre-judges it just injects false confidence into the claims. Amounts come from documents only, and anything missing is marked "cannot verify," because one estimated figure in a reconciliation report forces the reader to re-verify the whole report - better one row short than one row fake.

Variables