The path is not straight

Toward one setting that works for every bottle in this deliberately tidy training set. The path there is less tidy.

Here is the same loop with the pressing taken care of. Each press of one pass runs all eight corrections in sequence and adds a bar to the list. Run several and watch the bars come down.

The total absolute-error bars get shorter on every pass in this run. The table underneath tells a less tidy story, and note what its columns mean now: not before training, but before and after the single epoch you just ran.

Most bottles move closer. Some move away, and those are marked with a small arrow beside the new price. Press again and the marked rows change, because the bottle temporarily worsened this time may not be the one worsened last time.

Press through the first several epochs slowly and watch those arrows. After the second epoch, two bottles have gone backwards. By the fourth and fifth, half the set is going backwards on any given pass, while the total keeps falling the whole time.

That is not a contradiction. Every bottle was priced from the same hidden four-number rule, so one exact shared setting does exist.

The bottles share the same four weights, so a correction made for one lands on all the others. Fixing the bottle in front of you can knock a bottle you fixed three turns ago back out of place. That is what the arrows are showing, and it has a name: interference.

Which is why a bottle going backwards is not a failure. It is the cost of the weights being shaped by eight wines instead of one, and the total still falls because the pulls agree more than they disagree.

Keep pressing and the arguing dies down. Somewhere around the ninth epoch the marks stop appearing: every bottle ends the pass a little closer, because the weights are nearer the common setting and each correction is smaller. The interference was loudest while the model still had a long way to move.

No single bottle reveals the full rule. The pattern appears only because different bottles keep correcting the parts they activate.

The path is not straight

Toward one setting that works for every bottle in this deliberately tidy training set. The path there is less tidy.

Here is the same loop with the pressing taken care of. Each press of one pass runs all eight corrections in sequence and adds a bar to the list. Run several and watch the bars come down.

That is not a contradiction. Every bottle was priced from the same hidden four-number rule, so one exact shared setting does exist.

No single bottle reveals the full rule. The pattern appears only because different bottles keep correcting the parts they activate.