02_​MONITOR

The monitor decides and the monitor calls. 02_MONITOR sees the drift and starts 01_TRAIN itself. The registry/ folder is the only thing the two of them share.

1 Which model is in production

Never told which model to use: it reads registry.csv, takes the highest version and loads the file that row points at.
Red here means the registry is not there yet — nobody has trained a v1. Run 01_TRAIN once, then press this one again.

Ask the registry, do not assume

2 Which batch is next

Never told which batch to score either. It takes the newest row of monitor_log.csv: RETRAIN → score that same batch again, anything else → move on one.

Ask the log, do not be told

3 Score it

acc is how often the model was right; lost is that against the accuracy this version was accepted at.

Two numbers, and only one of them is new

4 Decide and log

Lost more than 20 → RETRAIN. A newer model than the one on the last row → REVIEW: the retrain's report card, which decides nothing.

One threshold, one row in the log

5 Retrain, from here

The monitor does not report the retrain, it performs it. The IF Switch deactivates the branch not chosen, so on HOLD the Call node never runs.

The monitor decides and the monitor calls

The loop closes inside KNIME. On RETRAIN this workflow calls 01_TRAIN itself, with train_upto set to the batch that just failed. A new version lands in the registry, and the very next press — same nodes, nothing edited — scores that same batch again with it and writes a REVIEW row. That second row is the retrain's report card, measured on the data that triggered it.

So a batch appears twice in monitor_log.csv, and that is correct. The rule fires at batches 2, 7, 8 and 9, so the log is 13 rows for 9 batches. Press once more after batch 10 and the Breakpoint says the data has run out.

Batch 8 is the row to read twice: there the retrain made it worse, 48.6 → 47.3. The log says so instead of hiding it — README.md.

The fast clock — two numbers that need no labels, and decide nothing

data_shift = mean over the 16 sensors of |mean now − mean ref| / sd ref pred_shift = |yes-share now − yes-share on the reference window|, in points

Both are read against that version's own profile_v<N>.csv, and both are written into monitor_log.csv. Neither one touches the IF Switch. The decision is still lost > 20, and lost needs the labels.

And that is the lesson, not a shortcut left out. On this run each fast signal's RETRAIN range OVERLAPS its own non-RETRAIN range, so neither gives a threshold; the numbers are in README.md. A fast signal says something changed, only the slow one says the model got it wrong.