Where in a codebase would you allow the conversion to an unlabelled rectangle, and what would you require at that boundary?

## Why this deserves a rule rather than a habit Every crossing from a labelled table to a positional rectangle moves a contract out of the data and into the codebase's memory of a field ordering. Any one crossing is defensible. What is not defensible is having no position on where they are allowed, because the cost is not paid by the line that does the conversion. It is paid by whoever later adds a field upstream, or reorders a selection for readability, and gets arithmetic over the wrong pair of values with nothing raised anywhere. That is a standing cost, spread thin across a team and across time, which is exactly the shape of thing a lead is expected to own. ## A defensible default State it as: **convert late, convert narrow, convert in one declared place.** Concretely, three requirements at the seam: 1. **The conversion takes an explicit ordered list of fields**, never the whole table. A field added upstream then cannot join the rectangle by accident, and a reordering upstream cannot silently change what a given position means. 2. **The rectangle does not travel.** It is produced inside the routine that needs it and turned back into named output there. A positional holding passed across a module boundary hands an ordering contract to callers who never agreed to it and cannot see it. 3. **The ordering is recorded where the conversion happens.** If the rectangle must be returned at all, return the ordered field names alongside it so the receiver can check rather than assume. The rule costs a little ceremony at one seam. What it buys is that the number of places in the codebase where position carries meaning is a number somebody can state. ## The opposite default, and when it is right This is where the judgment actually lives, and a candidate who presents the rule above as universal has missed it. In a codebase whose bulk is uniform numeric computation — where names matter at the edges, when data is read in and when results are reported, and carry no weight in between — **the rectangle is the natural primary holding and the labelled table is the adapter.** Forcing every intermediate step through a labelled table there adds a second coordinate system nothing uses, and pays for names on values that are never looked up by name. | codebase shape | primary holding | where the conversion lives | | --- | --- | --- | | fields added, renamed and matched on throughout | the labelled table | one narrow seam, late, explicit field list | | uniform numeric work between a read and a report | the positional rectangle | at the two edges, named on the way in and out | | both, in different subsystems | whichever each subsystem is | at the subsystem boundary, ordering recorded | The rule is the same rule in all three rows: **there is one declared place where the two holdings meet, and it is written down.** What varies is which side is the default, and that is decided by how often the code refers to a value by name. ## What to argue about, and what not to - **Worth arguing:** whether a given subsystem is numeric enough to earn the rectangle as its default; whether a returned rectangle may cross a module boundary at all; whether the explicit field list belongs in code or in configuration. - **Not worth arguing:** whether the field ordering is "obviously stable". Nobody promised it, and treating it as a promise is precisely how the failure arrives. - **A trap:** banning the conversion outright. Code that genuinely needs a rectangle will convert anyway, in an unreviewed corner, without the field list — so the ban has produced the worst version of the thing it banned. - **A second trap:** allowing the conversion but keeping it early, at the top of a pipeline. Everything after it then works positionally, which is the maximal version of the cost with none of the benefit. ## Making the rule stick A rule that lives in a document is checked when somebody remembers it. A rule expressed as code is checked on every run. The cheapest version is a single named routine that takes a table and an ordered list of field names and returns the rectangle; everything else calls that, and the field list is then visible at every call site. It is not enforcement, but it turns "did they convert correctly?" from a question about discipline into one you can answer by reading the call sites. ## What an interviewer is listening for Not a rule. A rule **with its precondition**, an honest account of what it costs, and a case where the opposite rule is better. "It depends" with nothing after it fails the question; so does a universal ban. The signal is a candidate who can name which property of a codebase decides the default, say what the rule costs the people who follow it, and state what would change their mind.