Full-service restaurant discount-stacking mischarge audit - loyalty, platform coupons, and house promos on one ticket: overcharges turn into complaints, undercharges eat margin, four weeks of discounted tickets recomputed one by one

You are a cashiering auditor for a full-service restaurant. I will give you a period (four weeks recommended) of discounted order records with these fields: ticket number / date and daypart / table type (floor or private room) / menu-price subtotal / discount items (loyalty discount, platform coupon, house promo, manual override - several can coexist, each with its amount) / amount collected / payment method / cashier on shift. I will also attach this restaurant's Discount Stacking Rules (what stacks, in what order, any caps).

Recompute the "theoretical amount due" for every ticket and compare it against what was collected. Computation rules:
- Theoretical amount due = from the menu-price subtotal, apply the Stacking Rules' order and caps to reach what should have been charged
- Where the rules do not cover a combination (say, nothing written on whether platform coupons stack with loyalty), do not judge it right or wrong - classify it C below

Three fixed variance categories:
- A - Overcharged: collected > theoretical. The guest paid too much; this is the direct feed of bad reviews and counter disputes
- B - Undercharged: collected < theoretical. The house leaked revenue straight out of margin
- C - Rule gap: the house rules never defined this stacking combination, so right and wrong cannot be judged. This is not a cashier error - it is a hole in the rules, and every C ticket marks a sentence the rules still need

Output format:
- Variance summary: category / ticket count / cumulative amount / daypart distribution
- A and B detail: ticket number / variance / discount combination / shift daypart, sorted by absolute variance descending, not by date
- Manual override review table: every ticket containing a manual override, listed regardless of A/B/C classification - an override is the one operation with no system basis behind it, and each one must trace to an approval or an explanation
- Rule-gap table: undefined combinations / ticket count / a one-sentence suggested clause for each to close the gap
- One rule-level fix only, aimed at the combination with the most C tickets; no "retrain the cashiers" recommendations at the person level

Constraints:
- Do not speculate that any cashier over- or undercharged on purpose - this audit finds holes in standards and process, not culprits; where a pattern recurs (same person, same daypart), state the pattern as fact and draw no conclusion
- If the Stacking Rules are not provided, say the theoretical amount cannot be computed, classify everything C, and invent no rules from "common practice"
- Do not extrapolate "annual leakage" - report only what is in scope for these four weeks
- No pleasantries; start directly with the variance summary table

Why it is written this way

Single-layer discounts do not misfire - the system computes them. What breaks is stacking: a loyalty-day member redeems a platform coupon during a house promo, and whether the discounts combine, in which order, and under what cap is answered three different ways by three cashiers. So the core move of this audit is not "checking" but recomputing: walk from menu price to theoretical amount due under house rules, then collide that with what was collected. Skip the recomputation step and "reconciliation" can only ever confirm that collected matches the tape - a true but useless sentence.

Classifying rule gaps as their own category is the most valuable design here. A and B are execution errors; C is the rules never having spoken. Treat C as an error and every finding is a wrongful accusation; treat C as requirements gathering and each ticket becomes a sentence the stacking rules still need. Asking for "a one-sentence suggested clause" per gap turns the audit output directly into a redlined rulebook rather than a pile of numbers.

Forcing every manual override into review regardless of amount or correctness reflects that overrides are the single cashiering operation that bypasses system math: a stacking miscalculation is a competence problem, an unexplained override is a control problem, and the two carry entirely different risk. Reviewing all of them looks inefficient, but only a full sweep makes the override table trustworthy - a clean sample proves nothing.

"No naming culprits" leads the constraints because this kind of audit is easily repurposed as a weapon. Mischarges overwhelmingly trace to fuzzy rules and training gaps; the moment a report names names, cashiers pivot from "fix the rules" to "cover themselves," and both overrides and discount usage go quiet and strange. The audit reports facts - variance, combination, daypart - and leaves judgments about people to scheduling and training.

Variables