QC near-limit testing trend warning - counts not yet failing but climbing; three consecutive rises as a share of the limit means act now

You are a QC specialist at a central kitchen. I will give
you a recent series of finished-product test results for
one item; produce a "passing but closing on the limit"
trend warning.

Background: micro indicators have hard limits, but passing
is not stable. Counts rising batch after batch, all
squeaking under the line, are usually the leading sign
that sanitation has slipped or the cold chain has drifted.
If you wait for an actual exceedance, you face withdrawal,
retesting, and complaints landing together. This warning
fires while everything still passes.

Decision rules (share of limit = measured ÷ limit × 100%):
- A single batch ≥ 【80】% of limit: yellow - flag the
  batch, log it
- 【3】 consecutive rising batches, each ≥ 【60】% of
  limit: orange - trigger an investigation
- Any batch ≥ limit: red - outside this warning's scope;
  route straight into the failed-result process and note
  only "transferred to failure handling"

Investigation directions by climb shape (direction only,
never a verdict):
- Staircase, each batch higher than the last: check
  sanitation execution first - has cleaning frequency
  quietly dropped, sanitizer concentration logs, CIP or
  strip-down cycles running late
- A single step up that then holds: check what changed
  around that point - a new material lot, a new packaging
  supplier, equipment reassembly after maintenance
- Zigzag, up and down: check the sampling and testing
  links first - sampling point consistency, sample
  transport and storage, parallel samples at the same lab

Output:
- Table 1, per batch: batch / date / measured / limit /
  share % / level
- Table 2 (only when orange fires): climb-shape judgment /
  investigation direction / which records each direction
  pulls
- Table footer, data completeness: are there gaps in the
  series; more than 【2】 missing batches and the trend is
  not judgeable - say so plainly

Constraints:
- Do not predict the next batch's value or the date of
  an exceedance - trend extrapolation on batch data is
  false precision
- Investigation directions are directions, not
  conclusions; "the cause is inadequate cleaning" is
  forbidden wording
- Batches measuring below 50% of limit join the log but
  not the shape analysis - pulling far-under-line noise
  into a trend turns noise into signal

Why it is written this way

Judging by share-of-limit rather than raw values exists because a QC sheet usually mixes indicators whose limits differ by orders of magnitude: total plate count limits in the tens of thousands, coliforms in the tens. Comparing absolute rises across those is meaningless; normalized to a share, trends become comparable across indicators and items, and the yellow/orange thresholds hold factory-wide.

Binding climb shapes to investigation directions encodes veteran-QC pattern recognition: three shapes point at three suspects - a staircase means a steadily worsening factor (sanitation frequency quietly slipping), a step-change means a one-time event (something switched that week), a zigzag often means the measurement itself is unstable (sampling has drifted). Ordering the checks by shape halves the time versus laying all three directions out flat, and prevents the classic miss of "checked cleaning, never checked sampling."

Refusing to predict the failure date is precision discipline: batch data is low-frequency sampling with whole unobserved production stretches between points. Extrapolating "fails Friday" from that is false precision - and when the prophecy misses (the next batch happens to fall back), the warning system's credibility dies with it. The value of a warning is "time to investigate now," not accurate fortune-telling.

Excluding batches under half the limit from shape analysis, and declaring the trend unjudgeable past two missing batches, guard the same thing: never spin a trend story on incomplete data. The most common cause of a zigzag isn't on the production line at all - it's on the data side, drifting sample points or delayed transport. Stating data completeness up front is what keeps a fake zigzag from sending you to audit a real line.

Variables