The shortcut only one model took
Given a StockTable that draws the whole stock page in one node, the strong model used it every time. The weak one mostly built the page by hand, or picked the wrong shortcut.
Date: 2026-11-10. Context: the warehouse's own components in generated views (journal ch. 67).
What happened
The catalog got five warehouse components, each answering a whole part of a page in one node:
StockTable, PickTable, ShipmentStatus, ReceiptLines, QuantityField. One short run per
model on the five warehouse tasks, against the same tasks with the shared components only.
Beam used the right component in all five tasks, first try, and wrote 37% less (PickTable took
it 862 output tokens). Nemotron used one in two tasks. For the goods receipt it chose
QuantityField, the one-quantity field, instead of ReceiptLines, the one made for order lines,
and dropped the supplier's name on the way. For the shipment it wrote the status by hand and the
page never said "In transit". It finished two pages of five, against three with the shared
components only, while writing half as much.
What we learned
- A shortcut helps the model that recognises it. The components were described in one line each, among twenty-four; the strong model matched them to the tasks, the weak one mostly did not.
- Less writing is not more finishing. Both models wrote far less; only one finished as much.
- A near miss is worse than no shortcut.
QuantityFieldin place ofReceiptLineswas accepted, rendered, and lost data a primitive layout would have kept.
What it means for Vektos
The same lesson as the data binding (anecdote 25), from another side: richer help is a bet on the model. For a design system that serves several agents, the composite components are worth having, but the shared primitives stay, and the catalog's words decide whether a weak agent finds the right composite.
Follow-ups
- Name the components after the tasks they answer ("ReceiptLines: the lines of a goods receipt against a purchase order") and run Nemotron again.
- Count, per run, which components a view used: the pattern above came from reading the specs.
The help that hurt
Letting views point at their data made the strong model perfect and the weak one worse, three reruns in a row. The new feature changed how the weak model wrote everything else.
The label that pointed nowhere
The first behaviour checks failed a model that had done the right thing: the design system's date picker had no label inside a form field, and axe had never noticed.