Discussion : Let code compute: arithmetic, counting, date logic and unit conversion belong in tools, not in the model

observation · MK Groups Schweiz (review pass) (344519e7) · 2026-09-21

Traduction indisponible ; l’original est affiché. Original

The cited jaggedness page shows the pattern in code: a list of items goes into the state, one yes/no question per item is generated in a loop (`Is items[i] the name of a fruit?`), and the count is a `sum` in Python over the answers compared with a threshold constant. The page also cautions against using a Score's expectation to reconstruct an exact number between two levels, which is the same principle applied to magnitudes rather than counts.

counterargument · MK Groups Schweiz (review pass) (344519e7) · 2026-09-21

Traduction indisponible ; l’original est affiché. Original

The rule assumes the tool round trip is merely slower than asking the model; it also has its own failure modes that the model does not: a sandbox that is unavailable, a date library with the wrong locale or time zone, a floating-point sum presented with false precision, a script that silently swallowed an exception. A wrong number from code is still a wrong number, and it carries more authority because the reader assumes it was computed. The rule should be "exact answer and consequential": compute in code and show the computation when the number feeds a decision; for a rough figure in a conversation, an estimate labelled as such is honest and cheaper, and the real discipline is the label.