Barely-Existing Columns and Headers

A column whose body is entirely empty still exists. It MUST be kept as an empty column, and any marker on its label MUST be kept — a superscript such as New Program² is part of the header text, not decoration.

If the corner above the row labels is blank on the page, it MUST be rendered blank. "Category", "Item" or any other reasonable-sounding word MUST NOT be supplied. An unlabelled cell is data too.

| State | Premium | Losses |
| (1)   | (2)     | (3)    |   ← a header row: render it

A numbered tier running across the columns is a header row, not decoration, and MUST be rendered.

Dropping it costs twice. The tier also runs the full width of the grid, which is the evidence that every column below belongs to one row — the proof Group and Cohort Bands (§9.2.5) and Bands Inside the Grid (§9.2.1) rely on to show that banded blocks are a single table. Without the tier, both the header and the boundary signal are lost.

A band naming a section is a header row even when it is set in plain type. "Lending" over a block of lending figures, or a CA|PPO line above its rates, are headers; demoting them to body rows is non-conforming.

| Lending    |       |   ← plain-styled band naming a section: still a header
| Commercial | 120.5 |

Boldness MUST NOT decide what is a header. Headers are recognised from position, span and what the line does for the rows beneath it — Bold and Italic Inside Tables (§9.6.2) shows styling is applied unreliably in both directions and cannot carry this judgement.

| 34 | S |          | [rowspan=3]34     | S |
|    | M |    vs    | ROW_SPAN_CONTINUE | M |
|    | L |          | ROW_SPAN_CONTINUE | L |

A region that is visually blank beside or below a labelled cell MUST be rendered as separate empty cells. The label's rowspan or colspan MUST NOT be extended over it unless the cell is genuinely merged — when the print says so, not when it is inferred.

Telling a real merge from adjacent blankness

The stopped-divider test. A ruling line that stops is as informative as one that continues. On a bundle-credit rate table the divider beneath Number of Additional Policies stops before the Factor column: that stopped rule is positive evidence that the group covers Car and Other only, and that Factor is a genuinely merged cell spanning both header tiers. What looks like "blank space above Factor" is in fact inside a visibly merged two-row header cell. Where a divider stops, the cell beside it spans.

Most blank-adjacency cases have no such geometry. Absent a stopped divider or an equivalent printed signal, nothing proves a merge — which is why the default is do not extend.

Carve-out. A stub label that spans the header tiers (Multi-Row Header Tiers (§9.3.3)) still takes a rowspan across those header rows. That is a logical merge and this rule does not touch it.

Why this way. A span mirrors what is printed, never what is inferred. A blank cell asserts nothing; a span asserts "this label owns that region", which needs printed evidence. Making merge the default would teach the reflex when in doubt, rowspan — exactly the habit Repeated Identical Values (§9.5.4) exists to suppress. It is also the simpler question: "is this cell's printed content empty?" rather than a semantic judgement, the same test Blank and Spacer Rows (§9.5.1) uses in the table body.

Every case here is a drop or an invention. One principle covers all of them, the one also stated in Repeated Identical Values (§9.5.4): never emit less data than went in, and never invent what was not printed.