diff --git a/README.md b/README.md
index d209f11..d017560 100644
--- a/README.md
+++ b/README.md
@@ -25,6 +25,22 @@ package implements.
 - `USCOGDATA_CACHE_DIR` — optional override for the manifest cache directory
 - `USCOGDATA_MANIFEST_TTL_SECS` — optional manifest re-fetch TTL (default 3600)
 
+## Raw-parquet caveat: `survey_weight` is not an aggregation weight
+
+Users reading the corpus parquet directly (DuckDB, arrow) will see a
+`survey_weight` column (schema v5, col 26). It is legacy Census IndFin
+sample-design **metadata passed through verbatim** — the Census Bureau's own
+source documentation says it "is for informational purposes only and should
+not be used to derive any other statistics" (`_ReadMe_First_IndFin.txt`;
+likewise `UserGuide.xls` Data User Note 8: "Do not use the weight field to
+derive state or national totals"). The raw encoding is also inconsistent
+across vintages (reciprocal scale most years, direct scale in 2003, a `1`
+placeholder in 1967/70/71/73/2001, all-`0` in 2007–2012, `NA` for all
+modern-source rows), so `sum(amt * survey_weight/10000)`-style expressions
+produce silently wrong totals — including exact zeros for 2007–2012. Sum
+`amt` unweighted; no uscogdata function reads this column. Full evidence:
+`cog_pipeline/.superpowers/sdd/weight-semantics-findings.md`.
+
 ## Developer notes
 
 ### Testing