diff --git a/docs/REVIEW_LOG.md b/docs/REVIEW_LOG.md
index 610e0c7..29fb2aa 100644
--- a/docs/REVIEW_LOG.md
+++ b/docs/REVIEW_LOG.md
@@ -7453,3 +7453,66 @@ out loud by whatever replaces these two than quietly depended on.
 
 **No merge performed.** Both guards remain in the chain, both green, and the decision sits
 with the humans.
+
+---
+
+## R-304 — bold means two different things, and that decides the merge
+
+**Recomputed from the corpus at `e2a474b` rather than quoted.** The concurrent session
+measured the bolding of table columns after R-303 flagged it, and offered the numbers for this
+log; the house rule is to re-derive a peer's figure before restating it, so here is the scan,
+run over every measurement column in the playable corpus with two or more numeric cells.
+
+```
+measurement columns:  18        bolded INCONSISTENTLY:  8   (44%)
+   CLEAN_GROUND.md:1119  "Wiped"          4 bold of 9
+   CLEAN_GROUND.md:1119  "Median rounds"  4 bold of 9
+   CLEAN_GROUND.md:1143  "Wiped"          2 bold of 4
+   CLEAN_GROUND.md:1143  "Median rounds"  2 bold of 4
+   CLEAN_GROUND.md:1198  "Wiped"          2 bold of 5
+```
+
+⚠ **One correction, and it matters because the rate is the rhetorical number.** 18 columns and
+44% is what **`check-figures`'s** vocabulary yields. Under **`check-unmarked`'s** own
+vocabulary the same scan gives **14 columns and 8 inconsistent — 57%**. The count of bad
+columns is 8 either way and is robust; **the rate is not**, because the two guards do not
+define "measurement column" the same way. Anybody quoting a percentage here must say whose
+vocabulary produced it.
+
+### Why the columns are inconsistent, which is the finding
+
+Four consecutive rows of the mixed-force table, same two columns:
+
+| What they actually fight | Wiped | Median rounds |
+|---|---|---|
+| Six hollow men, column stands off | `0%` | `13` |
+| **+ two of the column** | **`8.3%`** | **`21`** |
+| **+ three of the column** | **`30.1%`** | **`25`** |
+| + six of the column | `93.8%` | `16` |
+
+**That is not carelessness. It is a second convention.** The two bolded rows are exactly the
+two the prose under the table singles out — *"Two of forty-one taking a swing moves the
+encounter from nobody dying to one wipe in twelve. Three takes it to nearly one in three."*
+The plain rows are the bounds nobody expects to run.
+
+**So in prose the corpus bolds what it PUBLISHES, and in a table it bolds what it wants READ.**
+Bold is a publication mark in one place and emphasis in the other, and a strict class keyed on
+boldness silently assumes the first everywhere.
+
+### What that settles
+
+The merge divides on evidence rather than on either session's preference, and it says where
+each rule is authoritative **and where each must stay silent**:
+
+| | Rule | Because |
+|---|---|---|
+| **Prose** | bold-as-published | in running text this corpus has no other use for bold on a number |
+| **Tables** | column inherits header, **bold ignored entirely** | inside a table bold carries no information about whether a figure is measured |
+
+**It also explains the asymmetry R-303 found from both sides**, which until now was only an
+observation: `check-unmarked` missed the unbolded cells because bolding in tables is optional,
+and `check-figures` missed the bold prose figures because they are not in columns. One cause,
+two symptoms.
+
+**No merge performed, and neither session will start one on the other's say-so.** The decision
+is with the humans; this entry is the evidence for it.