Hotpot restaurant trainer - training effectiveness behavior audit: surveys measure satisfaction; the dinner rush measures behavior

You are the trainer at a hotpot restaurant. Design a "training
effectiveness behavior audit" plan: two weeks after training
ends, replace the questionnaire with on-the-floor observation
and audit whether training became behavior.

Observation targets and action lists (by post; I may add or
remove items):
- Servers: offer the three-item kit on seating (apron / phone
  bag / hair tie); proactively ask heat level and allergies
  when taking the order; announce the base name when it lands;
  top up broth before the line drops below [standard] without
  waiting to be flagged
- Broth station: dilution ratio measured with the standard
  vessel (eyeballing does not count as execution); self-check
  against the reference sample before sending; send within
  [standard] of firing
- Runners: announce each dish on landing; bases and soup dishes
  take the steady route; placement orientation per table
  standard

Audit method:
- Sampling: peak hours, each person observed [N] times minimum
  per item (items under N observations are tagged "sample
  insufficient")
- Execution rate = observed executions / total observations
- Grading: >=80% habituated; 50-79% half-formed, added to
  pre-shift spot checks; <50% not landed, goes to attribution

Attribution for not-landed items (the critical step, two
branches):
- Re-demonstrate once on the spot and they immediately do it
  right -> "can't": a training problem, the trainer's
  responsibility, retrain that action on a deadline
- Re-demonstrated and still not done -> "won't": an execution-
  environment problem (unmanageable peak load / no checking
  mechanism / no difference between doing and not doing),
  hand to the store manager and scheduling supervisor -
  training does not take it back

Output format:
- Person x action execution-rate table: trained actions only,
  no commentary on other performance
- Half-formed and not-landed lists
- Attribution table for not-landed items: action / "can't" or
  "won't" / basis (the on-the-spot reaction after the
  demonstration)
- Training-side actions: retrain list (the "can't"s) + handoff
  list (the "won't"s, naming the recipient and the reason)

Constraints:
- Behavior only, never attitude - "bad attitude" appears
  nowhere in the output
- Survey scores and test results do not appear as effectiveness
  evidence in any conclusion
- "Won't" items get no "more training" / "attitude education"
  recommendations - state the handoff; that is training's
  boundary
- During observation, do not interrupt operations and do not
  correct on the spot (correction contaminates subsequent
  observations)
- No preamble; start with the execution-rate table

Why it is written this way

"Two weeks later" is the hardest parameter in the plan. Behavior right after training is memory; behavior still present two weeks out is habit - the audit point must sit past the forgetting curve, because audited on the day, every training session ever run shows "significant effect." Observing at peak isn't only about sample density either: peak is when actions fall apart first, and a standard action performed in a lull only counts as known once it survives the rush.

Rates over impressions is the foundation of the whole method. "They've mostly got it" is the classic cover story of training failure - how much "mostly" hides is unanswerable, while an execution rate turns "mostly" into a numerator and denominator, giving the problem a scale. The N-observation floor and the "sample insufficient" tag guard against the spot-check reflex of concluding from two or three sightings.

The "can't / won't" cut is the heart of the audit. The trainer's most common situation is being handed non-training problems to solve with training: with peak staffing stretched to the bone, nobody can execute "proactive top-ups"; with no checking mechanism, doing and not doing pay the same, and ten retrainings change nothing. Without this cut, the training function becomes the landfill for every floor problem, and every load dilutes what actually needed training. Naming the handoff recipient turns "training's boundary" from a complaint into a process.

No on-the-spot correction during observation is a quietly methodological rule: behavior after correction is corrected behavior, and every later observation measures "prompted execution" rather than "spontaneous execution." Swallowing that one correction is what buys clean data.

Variables