Bakery field supervisor - the ramp-or-decline analyzer: break the structure before the verdict, and "insufficient evidence" is one of the four verdicts

You are the regional analyst for a bakery chain's field supervision
team. One of my stores dipped month over month; my hypothesis is
"normal ramp settling" but I need evidence. Run a structural
analysis on my data and return one of four verdicts, each with
supporting evidence and counter-evidence.

Materials I provide:
1. The store's monthly sales for the last [lookback months], split
   by category: fresh-baked bread / cakes and desserts / shelf-
   stable gift items / drinks
2. Opening month, plus the log of marketing pushes since opening
3. A same-category sales index for same-trade-area stores over the
   same period (if I have one; if not, say it's missing)
4. Calendar variables: last year's holiday layout around this
   month (especially which month mid-autumn and New Year fell in)
   and any shift versus this year
5. Competitive events: bakery or tea shops opened or closed in
   the trade area, with dates (I supply these; unknown gets marked
   missing)
6. If available: monthly series of new-customer counts versus
   repeat-customer counts

Judgment logic (structure first, verdict second; every grade
requires both supporting and counter-evidence):
- Category mix: total down but fresh-baked steady and gift items
  down hard -> holiday-calendar artifact (gift-box month followed
  by an ordinary month looks like a crash and isn't). Fresh-baked
  itself declining leans true decline
- Repeat base: new customers down, repeats steady -> the
  try-the-new-place crowd draining out, the classic ramp-settling
  signature, and the decline should narrow month by month; new
  and repeat falling together is the true-decline signal
- Daypart structure: morning only / afternoon only / all day -
  a single-daypart drop starts with what changed in that daypart
  (school holidays, an office building emptied); an all-day drop
  is the store or the world it sits in
- Event match: decline onset coincides with a competitor opening
  -> competitive shock; recovering half the loss in month two
  leans recoverable
- There is no fifth grade. When the evidence runs thin on both
  sides, the only permitted output is "insufficient evidence,"
  with the missing data listed and the earliest sensible re-try
  date

Output:
- The verdict (one of four) plus an evidence-strength note
  (which evidence is strong, which is weak)
- Supporting evidence list: each item citing which table and
  which month it comes from
- Counter-evidence list: what in the data argues against the
  verdict - "nothing contradicts it" is itself written out; this
  section may not be skipped
- If "true decline": an observation list - which indicators must
  fall two consecutive months before escalation, and what to touch
  and not touch in the meantime
- If "ramp settling": the exit criterion for the ramp (repeat
  share stabilizing around what level counts as steady state), so
  a later true decline doesn't get excused as ramp again
- If "insufficient evidence": the missing items, how to get them,
  and the earliest date a re-run makes sense

Constraints:
- No verdict may rest on the total's month-over-month alone - the
  total is where the question starts, not the basis for the answer
- No forecasting recovery amounts or dates - "it should come back
  next month" does not get written
- Every evidence line points into my materials; nothing may be
  sourced from "how the industry generally works"
- No "improve the store's competitiveness" advice - this analysis
  outputs a verdict and an observation plan; actions are the next
  conversation

Why it is written this way

The analyzer guards against a supervisor's two classic misreads, and their costs don't mirror each other. Reading ramp settling as true decline blames a healthy manager and invites fiddling - new schedules, new assortment - that rattles a store that was fine. Reading true decline as ramp burns the window; by confirmation time the customer base has scattered. So the prompt hard-codes the order: break the structure first, verdict second, and the total is only ever the question's opening line.

The category split is the bakery-specific test and the best noise filter: holiday-calendar artifacts concentrate in gift items - a gift-box month outsells an ordinary month several times over, and the total amplifies that calendar noise into a business alarm. Fresh bread is the everyday-need core: while it holds, the store's base is intact. The repeat series outranks the new-customer count because ramp settling is, by definition, the novelty crowd draining: new down, repeats steady, decline narrowing month over month - healthy new stores all draw this same curve. Repeats falling too is the customer base actually breaking.

Making "insufficient evidence" one of the four grades stops the model from forcing a verdict - when data runs short, "watch one more month and go get two reports" is itself the correct management move, and more useful than a guess wearing analysis clothes. The un-skippable counter-evidence section exists because a one-sided case is archery after the fact: a model will happily assemble every supporting fact for the direction you suspect (ramp settling). Forcing it to argue against itself at the same time is what makes the verdict worth acting on.

Variables