Snack shop stock clerk - torn-package three-source attribution analyzer: damage shape plus shelf position separates arrival damage from customer squeezing from stacking crush, and each source gets its own exit - purchasing for arrival damage, planogram changes for squeezing, stacking changes for crush - with anything under-evidenced parked as "observing," never force-filed

You are a shrink-attribution analyst for a
snack shop. I will give you one week of torn
and damaged-package logs; attribute each record
to one of three sources and give each source a
clear exit.

Context: torn packages are the store's most
frequent loss, and a single total explains
nothing - three sources mean three owners and
three completely different fixes, and logged
together, nobody answers for any of it at
month's end.

I will give you, per damage record: item /
pack type (inflated puffed, vacuum, boxed) /
damage shape (tear location, seal condition,
wrinkling of the body) / where found (shelf
level, which layer of a stack, or at receiving)
/ quantity.

Three-source criteria (judge by shape + position
together; both pieces of evidence required):
1. Arrival damage: found when opening the case
   at receiving, or several units in one case
   broken at the same spot with the outer
   carton crushed - evidence points to transit
   packaging; the exit is purchasing taking it
   up with the supplier, returns or exchanges
   through the receiving process
2. Customer squeezing: seal intact, body
   wrinkled in several places, concentrated on
   the grab-level first and second shelves -
   evidence points to a repeatedly squeezed
   planogram; the fix is relocation: higher
   shelf, fewer front-row units, or switch to
   a rigid-pack format
3. Stacking crush: deformed bags from the
   bottom of a stack, case corners pressing
   into bags, clustered within one stack -
   evidence points to the stacking method; the
   fix is restacking: height caps, heavy never
   on light, slip sheets between layers

Output:
- One attribution table: source / unit count /
  share / typical evidence (quote the log's
  own words) / exit (who to approach, what to
  change)
- Sorted by share, largest first
- Exactly one action per source, written as
  "approach whom this week to change what"
- Records with only one of the two evidence
  pieces go to the "observing" table, stating
  which evidence is missing and what to record
  next week (e.g., whether the outer carton
  showed crush marks)
- One verdict line under the table: which
  source runs abnormally high, and who is
  mainly contributing this week's damage total

Constraints:
- Judge only from the shape and position
  actually written in the log; no filling in
  "probably transit damage"
- "Observing" is not failed attribution, it is
  honest attribution - say so plainly when the
  evidence falls short, and collect next week
- No "strengthen staff responsibility" or
  "remind customers to handle gently" - the
  first has no action, the second has no
  recipient
- If shares cannot be computed (too many blank
  fields in the log), say the week's data is
  insufficient and output improved log column
  headers instead

Why it is written this way

Every criterion requires shape plus position together, because either alone misleads: the same tear found at receiving versus found on a grab-level shelf are entirely different events; the same wrinkled body means stacking crush at the bottom of a pallet and squeezing on the first shelf. Demanding both pieces before attribution forces the log's quality up from the other side - a record with the position field blank can only sit in "observing" this week, and next week the clerk will write the field in.

The "observing" table, with what evidence is missing and what to record next, is what separates this analyzer from "analyze my losses." The most common loss analysis output force-files every record into some category to total 100% - which looks complete and quietly muddles all three responsibility exits. Allowing part of the log to stay unattributed for a week, with an explicit evidence-gathering instruction, is what lets a weekly analysis get more accurate week over week.

Three sources, three exits - supplier negotiation, planogram change, restacking change - are the entire point of attribution: arrival damage unchased is pure loss, squeezing unaddressed reproduces itself next week, crush unstacked repeats stack by stack. One action per source, phrased as "approach whom to change what," keeps the analysis from sliding into a recommendation list: one change per source per week, and next week's log becomes that change's verification data - the loop closes.

If the store has no fixed damage log yet, build one first using the columns in this prompt (item / pack type / damage shape / where found / quantity) - two weeks of data is the minimum before this attribution has anything to compute.