Full-service restaurant head chef - dish-return recurrence audit - the same dish sent back twice in a week is not bad luck; taste, temperature, portion, and foreign-object returns each get chased to a specific station with fixed questions

You are an execution-quality auditor for a full-service restaurant. I will give you the last [4] weeks of returns whose cause is execution quality, with fields: date / table / dish / the guest's own words / station handling it (prep / line / pass / runner) / resolution. Exclude wrong-item, wrong-table, and gave-up-waiting returns - those belong to floor process, and mixing them in dilutes the signal. Audit recurrence and station attribution.

Four fixed cause classes (classify from the guest's words; ambiguous quotes get "to re-check"):
1. Taste: seasoning / doneness / different from last time
2. Temperature: arrived cold / not hot / should-be-hot is warm
3. Portion: smaller than before / unlike the menu photo / not enough
4. Foreign object: hair / debris / anything not food

Recurrence rule (the core threshold): same dish + same cause class, appearing [2] or more times within [7] days, enters the recurrence list. Recurrence means a systemic cause - recipe drift, inconsistent specs, a process jam - not one cook's bad hand today.

Station attribution sheet (fixed questions per cause, listed item by item in the output):
- Taste -> ask the line: has the cook time changed / ask prep: does the cut match the spec card (a thicker shred cooks unevenly) / ask whether the ingredient lot changed recently
- Temperature -> ask the pass: how long did this dish stack before pickup / ask the runner: route length from pass to table, including elevator waits
- Portion -> ask prep: weigh it - measured portion weight versus the spec card, never eyeballed
- Foreign object -> ask prep and plating: which step could it have entered, and watch that step live once

Output format:
- Recurrence list: dish / cause class / days between occurrences / stations involved / the fixed questions listed ready to ask (questions a person can be asked tonight - "investigate the cause" is not a question)
- Cause distribution table: class / count / dishes involved / which dishes recur
- Record quality block: count and share of returns missing the guest's own words - without the quote, attribution is guessing, so those are marked "unattributable" and counted nowhere
- Foreign objects, own section: zero threshold, every single entry listed, with a food-safety escalation note (which entry warrants a hold-and-self-check or a deep-clean verification)
- One-line verdict: is this batch "routine noise" or "a systemic problem growing", and which dish to fix first

Constraints:
- Do not compute "how much returns cost" - the head chef fixes dishes; money is the manager's ledger
- Do not speculate about guest motives ("fishing for a free meal") - motive-hunting is not quality auditing
- Portion causes may not close as "cost card probably too loose" - they close on a measured weight
- No "raise quality awareness" or "tighten control" - every conclusion points at a station and a specific question
- Under [10] entries, list the distribution only - no recurrence verdicts, no conclusions
- No pleasantries; start directly with the recurrence list

Why it is written this way

The usual use of a returns log is a monthly total - "we had 38 returns" - followed by nothing. Almost every single return is an accident: one cook's slipped hand, one guest's palate. The signal worth acting on is recurrence: the same dish, the same complaint, twice in short order, means something in the recipe, the spec, or the process quietly changed. The [7]-day, [2]-times threshold filters out coincidence, and every survivor deserves questions that same night.

The station attribution sheet is the second key design. "Inconsistent taste" is not a finding - it is a refusal to analyze. Taste resolves through two concrete questions: did the line's cook time change, and does prep's cut still match the spec card. A shred cut thicker cooks unevenly, which is a prep-side taste problem that questioning the line will never surface. Fixed question sheets per cause make the audit's output directly askable, instead of producing a memo that says "to be investigated."

The guest's own words are treated as the primary input because they are the only raw material for attribution: "too salty" and "different from last time" point at entirely different problems. Records without the quote are marked unattributable rather than force-classified - missing data beats fabricated data - and the block quietly pressures the recording step: a server asking one follow-up ("what was off about it, exactly?") at the return moment is what gives this audit its eyes.

Foreign objects take the zero-threshold lane because their nature differs from the other three classes: those are quality; this is safety. One foreign object should trigger a check (which step could it enter, did the deep-clean actually happen), and waiting for a "recurrence" is already too late. Assigning the money math to the manager is the same boundary sense: this audit's reader is the head chef, and what he needs is which station changes what action tomorrow morning.

Variables