Hotpot restaurant field supervisor - broth consistency blind-tasting audit: the paste comes from the same bucket; what drifts is the assembly

You are a regional field supervisor for a chain hotpot brand.
Design and produce a "broth consistency blind-tasting audit" for
the stores I oversee; the deliverable is an audit plan plus a
results table.

Tasting protocol (fixed):
- Sampling: within the same week, order the same base type, same
  size, same ordering script at every store (dine-in or delivery,
  uniform across the round), no dipping sauces, and record the
  batch number of the base paste used
- Scoring: two tasters, blind and independent (no sharing scores
  during the round), four dimensions each 1-5:
  1. Visual yield: consistency of broth and oil volume at the
     table
  2. Aroma: intensity and character at first boil
  3. Initial salt and heat on the palate
  4. Late boil: salt drift and volume behavior after roughly
     forty minutes of cooking
- Drift inference patterns (outlier stores only, matched by
  profile): yield runs large and everything runs mild -> dilution
  ratio high; aroma weak and arrival temperature low -> pot sat
  too long after firing; front fine but late boil diverges ->
  simmer time short

Output format:
- Master table: store x four dimensions, averaged; any cell where
  the two takers differ by more than 1 point is flagged and that
  dimension voided for re-tasting
- Outlier rule: any store more than [X] points off the fleet mean
  on any dimension is an outlier
- Drift inference for outliers: matched against the pattern
  definitions, with the stated basis
- Verification plan: second tasting covers outliers only, and the
  supervisor attends one full base assembly on-site (recording
  the water measure, the timing, and the fired-to-table gap) to
  confirm or overturn the inference
- Graded conclusion: still outlier after verification ->
  "drifted," with the store's specific assembly deviations named;
  no longer outlier -> first round voided, possible causes noted
  (batch, daypart), and no remediation suggestions

Constraints:
- A single first round produces hypotheses only: no drift verdict
  on any store before verification - a one-round conviction is a
  verdict the store will never accept
- Scores live in tables; dimension-free descriptions like "felt
  a bit off" or "not aromatic enough" never appear
- No ranking of base types against each other - the audit's
  object is cross-store consistency of one base, not base
  criticism
- Stores running a different paste batch are listed separately
  and voided for the round - batch is an imported variable, not
  the store's fault

Why it is written this way

"Two independent tasters, cells differing by more than 1 point voided" is the precondition for a blind tasting to be a measurement at all. One supervisor scoring alone produces a personal palate preference; two scores that never met are a measurement, and one score is an impression. The voiding rule doubles as a built-in self-check on reliability - a dimension where the takers diverge isn't a store drifting, it's a thing that cannot be measured this way.

Late boil gets its own dimension because of the physics of hotpot broth: plenty of bases are perfectly normal at the table and turn salty and bitter forty minutes in as the broth concentrates. Guests attribute every bit of that to "the broth is bad," brand QA checks the paste and finds nothing, and the actual disease sits in dilution and simmer time. Without this dimension the audit misses the most common drift pattern in the format.

Drift inference is restricted to pattern-matching rather than open analysis to keep it verifiable: every pattern corresponds to actions directly observable during verification (the water measure, the timer, the firing-to-table gap) - verification is literally checking the match on-site. An open-ended "possibly management-related" inference cannot be verified, which makes it no inference at all.

"No verdict from a single round" protects the supervisor's own credit. Blind tastings run on inherently small samples; convicting a store on one round and then having verification overturn it collapses more than that conclusion - it takes the stores' trust in every future tasting with it. Hypothesis, verification, verdict: slower, and every step stands.

Variables