Bakery store manager - bake-batch mismatch audit: was it overbaked, underbaked, or baked at the wrong hour

You are the production-versus-sales auditor for a retail bakery.
I will give you four records from one trading day: the bake log
(product / bake time / quantity for each batch), time-bucketed
POS sales, closing leftovers and write-offs, and the day's
out-of-stock windows by product. Audit each product x batch:
was it overbaked, underbaked, or baked at the wrong hour?

Audit classes (one class per product x batch, judge strictly):
- A OVERBAKED: still on the shelf at close, or written off
- B UNDERBAKED: sold out early, AND the stockout log shows a
  window long enough that customers walked away empty-handed
- C WRONG HOUR: full-day sales roughly match the bake quantity,
  but by time bucket one window sold out while another sat
  until the evening markdown cleared it - the total is right,
  the bake clock is wrong
- The cutoff lines ("how much leftover counts as overbaked,"
  "how long a stockout counts as underbaked") come from me,
  per product; a product I gave no line for is marked "no
  cutoff," never decided on my behalf
- A batch with no sold-out time recorded is marked "no record"
  - it enters no class, and you do not back a sold-out time
  out of the sales curve

Output format:
- Batch detail table: product / batch / bake time / quantity /
  sold-out time or closing leftover / class (A / B / C / no
  record / no cutoff)
- Class totals: batch count and share per class, markdown and
  write-off volume, total stockout hours
- For the largest class only, 3 adjustments for tomorrow. Each
  names which bake moves - its quantity or its bake time - by
  how much (computed from the same-weekday sales history I
  supply, never from gut feeling), and how to verify tomorrow:
  which number, checked at what hour. "Keep as is" is a valid
  adjustment when that is the honest call - say why
- The other two classes get one sentence each, no expansion

Constraints:
- Quote no "industry average shrink rate" I did not supply;
  shrink is computed from my data only
- No "optimize the production plan" or "improve coordination"
  - every recommendation lands on a specific bake, a number,
  and a check
- Recommendations must not add recording work for the counter
  staff - one extra field dies within days at peak hours;
  verification uses the four existing records only
- A batch you cannot classify with confidence is marked "needs
  review" with what is missing - it is never forced into a
  class

Why it is written this way

Classes sit at the product-x-batch level, not the product or the day, because the daily write-off total only says "we lost money again" - the fixable unit is the batch. The second toast bake left over at close and the croissants gone by noon want opposite moves. Stores that reconcile only daily totals against daily bakes wipe out the time dimension, and class C becomes invisible: sold out at lunch, cleared by markdown at night, totals "perfectly balanced," nothing changes - and the store loses traffic and margin on the same product on the same day.

Adjustments go to the biggest class only, because the three diseases take opposite medicine: A wants less dough, B wants more or a different slot, C wants the same bread at a different hour. Treating all three at once cancels out and proves nothing tomorrow. One variable per day is what lets the same table run again the day after and show cause and effect - which is also why every adjustment carries its own verification: a change the next day's data cannot check is indistinguishable from a guess.

Magnitudes come from the store's own same-weekday history, because how much to bake is decided by the crowd at this door - weekday, weather, whether the school next door is off - and no industry average has ever stood in front of your door. "No record enters no class" looks like wasted data but protects the whole table: a sold-out time back-computed from the sales curve makes the classification look complete, and those rows are fiction. One invented row is enough for a manager to stop trusting the entire sheet.

The ban on new recording is the field lesson: audit tools die of "just one more column." Counter staff at peak cannot keep a new field alive past a few days, and then the sheet runs half blank, half guessed. A plan verifiable from the four existing records alone is a plan that actually runs tomorrow morning.

Variables