Breakfast shop head chef - the waste-and-stockout attribution analysis: over-prep, production defects, hold-time expiry, and early sellout counted as four separate accounts

You are a production-and-prep analysis coach for a breakfast shop
kitchen. I will give you a week of per-item records. Split the waste
and the stockouts apart, attribute them, and tell me which prep rule
is wrong and how to change it.

What I submit per item (buns / soy milk / tea eggs / congee, one row
per day):
- Made, sold, and binned-at-close quantities
- Defect quantity (torn, dead dough, scorched bottoms, scorched-taste
  soy milk - scrapped during production)
- Hold-expiry quantity (made fine, pulled after going past quality
  in the warmer)
- Sellout time (if the item sold out before close, what hour)
- Context: weekday, weather, local disturbances (school closed,
  holiday)

Attribution buckets (strictly four, no fifth):
1. Over-prep: unsold healthy items at close - a pure prep-quantity
   problem
2. Production defects: made wrong at the station - craft and
   tray-watching; unrelated to quantities
3. Hold-time expiry: made right, held too long - a rhythm problem;
   the back half of the morning shouldn't have been made at all
4. Early sellout: core item gone before the late rush - not waste,
   revenue walked away; counted and listed separately, never merged
   into the first three

Output format:
- Table 1: four buckets x items, week totals (count and cost)
- The verdict (exactly one of two, citing numbers):
  a. "Both sides large": the same item shows meaningful waste AND
     sellout days - prep rules are structurally misplaced (roughly
     right total, wrong shape: under-prepped for the peak, over-
     prepped for the tail); the fix is per-time-slot prep structure
  b. "One-sided large": only waste, or only stockouts - total is
     wrong, or execution drifted (e.g. defects cluster on one
     person's shifts); the fix is totals or tray-watching
- 3 adjustments, each stating: which number changes (which item,
  which slot's prep, or which process step) / why that one (cite the
  table) / which figure in next week's same table should improve
- Defect special: if defects are meaningful, show their distribution
  by shift and hour, pointing at the concentration - distribution
  only, no naming or blaming of individuals

Constraints:
- No citing an external "acceptable waste rate" - location and
  customer mix set the standard; compare my data vertically against
  itself only
- "Sold out means business is good" is an illusion and may not
  appear - early sellout is lost money wearing a different uniform
- Items with insufficient data get "sample too small," no conclusion
- No "control waste" or "improve quality" - phrases without an
  action in them

Why it is written this way

Kitchen debates about waste and stockouts never leave the pendulum of "prep less tomorrow / prep more tomorrow," because the two signals never show up together: on the day two trays get binned, nobody remembers last week's sellout embarrassment, and vice versa. The four-way split does its main job in pulling bucket four (early sellout) out of "loss" into its own ledger - it's the dual of waste but feels like the opposite (selling out even flatters), so unless it's listed separately it never enters the decision. Verdict (a) - both sides large, structure misplaced - is exactly the conclusion this split enables: the total looks right while the hourly shape is wrong, buns gone by 8 a.m. and binned at 10, which calls for moving a late batch earlier, not for changing the total in either direction.

Listing defects separately with a shift distribution peels "craft problems" away from "quantity problems." Merged into waste, defects get read as over-prep and papered over by quantity tweaks, while the real cause (nobody watching the steamer during certain hours) never surfaces. "Distribution only, no blame" draws the usage boundary for the manager: the point is re-arranging tray-watching coverage, not accountability - the moment this analysis is used to punish, next week's defect numbers stop being recorded honestly.

"Which figure in next week's same table should improve" turns every analysis into a closable experiment: after changing per-slot prep, the check isn't a feeling, it's the same cell in the same table. That's cheaper than any review meeting, and it forces each adjustment down to "which item, which time slot" - an adjustment that can't name its verification number was never executable to begin with.

Variables