diff --git a/CHANGELOG.md b/CHANGELOG.md
index 3729529..55da00d 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -15,6 +15,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   - The `supports_snp_labels` and `supports_secondary_axis` boolean properties are removed. Optional capabilities are negotiated with `@runtime_checkable` protocols (`SupportsRegionHighlight`, `SupportsSNPLabels`, `SupportsSecondaryAxis`) checked by `isinstance`. `supports_hover` stays a boolean.
 - **Legend styling is uniform across plot families.** The five per-family legend styles are replaced by one. Visible differences from 1.x: the LD lead-SNP marker is 7pt rather than 6pt; eQTL, fine-mapping, and effect legend text is 9pt rather than 8pt, making those boxes slightly taller; and the effect legend now carries an "Effect" title. Legend content, colours, and ordering are unchanged.
 - **`schemas.py` validators are expressed on `DataFrameValidator`** instead of a parallel hand-rolled implementation. Which frames are accepted or rejected is unchanged; error wording differs where the engine phrases a check differently (`Missing columns: [...]` for `Missing required column`, `values <= 0` for `non-positive`, `null values` for `missing values`). A frame with a missing column now also reports its dtype and range faults instead of stopping at the first phase.
+- **BREAKING: load-time validation rejects nulls consistently.** `validate_gwas_dataframe` already required non-null positions and p-values; `validate_eqtl_dataframe`, `validate_finemapping_dataframe`, and `validate_genes_dataframe` did not, so a file with a blank `p_value`, `pip`, `start`, or `end` loaded silently and lost those rows later at plot time. All four now reject nulls in the numeric columns they range-check, naming the offending column. Plot-time intake is unchanged and still filters nulls with a warning, which is the intended two-tier split.
 - **BREAKING: two never-read parameters are removed**: `validate_gwas_dataframe(rs_col=...)` and `validate_finemapping_dataframe(cs_col=...)`. Both were accepted and documented but never used.
 - **Internal**: coordinate liftover moved behind a `CoordinateLifter` port with a pure `liftover_positions` (`_liftover.py`), and recombination header detection was extracted as the pure `ensure_recomb_header`. `liftover_recombination_map` keeps its signature.
 
diff --git a/CONTEXT.md b/CONTEXT.md
index d4b01b8..d68ed87 100644
--- a/CONTEXT.md
+++ b/CONTEXT.md
@@ -57,6 +57,12 @@ visualization concepts the library renders. Settled decisions live in
 - **eQTL `gene` requirement** — required at load (multi-gene GTEx files keyed by
   gene) but optional at plot (frames may already be gene-filtered). Same
   intentional two-tier shape as GWAS.
+- **Nulls are a load-time failure, not a plot-time one** — every `schemas.py`
+  validator rejects nulls in the numeric columns it range-checks, so a malformed
+  file fails at load with the offending column named. Plot-time intake stays
+  lenient and filters nulls with a warning (`prepare_pvalue_data`,
+  `prepare_eqtl_for_plotting`), because a frame reaching the plotter may have
+  been assembled by the caller rather than loaded from a file.
 
 ## Deepenings from the design grill
 
diff --git a/src/pylocuszoom/schemas.py b/src/pylocuszoom/schemas.py
index 0eeaa86..3a13828 100644
--- a/src/pylocuszoom/schemas.py
+++ b/src/pylocuszoom/schemas.py
@@ -80,6 +80,7 @@ def validate_eqtl_dataframe(
         _loader_validator(df, "eQTL")
         .require_columns(["pos", "p_value", "gene"])
         .require_numeric(["pos", "p_value"])
+        .require_not_null(["pos", "p_value"])
         .require_range("pos", min_val=0, exclusive_min=True)
         .require_pvalue("p_value")
         .validate()
@@ -110,6 +111,7 @@ def validate_finemapping_dataframe(
         _loader_validator(df, "Fine-mapping")
         .require_columns(["pos", "pip"])
         .require_numeric(["pos", "pip"])
+        .require_not_null(["pos", "pip"])
         .require_range("pos", min_val=0, exclusive_min=True)
         .require_range("pip", min_val=0, max_val=1)
         .validate()
@@ -140,6 +142,7 @@ def validate_genes_dataframe(
         _loader_validator(df, "Gene annotation")
         .require_columns(["chr", "start", "end", "gene_name"])
         .require_numeric(["start", "end"])
+        .require_not_null(["start", "end"])
         .require_range("start", min_val=0)
         .require_ordering("start", "end")
         .validate()
diff --git a/tests/test_validation_contract.py b/tests/test_validation_contract.py
index 5d19c8b..989b764 100644
--- a/tests/test_validation_contract.py
+++ b/tests/test_validation_contract.py
@@ -1,10 +1,13 @@
 """Prose-independent contract for the loader schema validators.
 
-Pins what callers can rely on across a change of validation implementation:
-which inputs are accepted, which raise, the exception class, and which column
-names the error names. Deliberately asserts nothing about error phrasing, so
-the same file passes before and after ``schemas.py`` moves onto
+Pins what callers can rely on: which inputs are accepted, which raise, the
+exception class, and which column names the error names. Deliberately asserts
+nothing about error phrasing, so a change of validation implementation is a
+checkable claim rather than a promise. It held unchanged across the move onto
 ``DataFrameValidator``.
+
+This is the strict load-time tier. Plot-time intake is deliberately more
+permissive; see the two-tier note in ``CONTEXT.md``.
 """
 
 import pandas as pd
@@ -55,21 +58,11 @@ def _genes(**overrides):
         validate_eqtl_dataframe,
         pd.DataFrame({"pos": [100], "p_value": [1.0], "gene": ["BRCA1"]}),
     ),
-    (
-        "eqtl-null-p-tolerated",
-        validate_eqtl_dataframe,
-        pd.DataFrame({"pos": [100, 200], "p_value": [0.05, None], "gene": ["A", "B"]}),
-    ),
     (
         "finemapping-pip-bounds-inclusive",
         validate_finemapping_dataframe,
         pd.DataFrame({"pos": [100, 200], "pip": [0.0, 1.0]}),
     ),
-    (
-        "finemapping-null-pip-tolerated",
-        validate_finemapping_dataframe,
-        pd.DataFrame({"pos": [100, 200], "pip": [0.5, None]}),
-    ),
     ("genes-minimal", validate_genes_dataframe, _genes()),
     ("genes-start-at-zero", validate_genes_dataframe, _genes(start=[0])),
     ("genes-point-feature", validate_genes_dataframe, _genes(start=[1000], end=[1000])),
@@ -212,6 +205,38 @@ def test_accepted_input_returns_the_same_object(validator, df):
         pd.DataFrame({"pos": [0], "pip": [0.5]}),
         ["pos"],
     ),
+    (
+        "eqtl-null-pvalue",
+        validate_eqtl_dataframe,
+        pd.DataFrame({"pos": [100, 200], "p_value": [0.05, None], "gene": ["A", "B"]}),
+        ["p_value"],
+    ),
+    (
+        "eqtl-null-pos",
+        validate_eqtl_dataframe,
+        pd.DataFrame({"pos": [100, None], "p_value": [0.05, 0.1], "gene": ["A", "B"]}),
+        ["pos"],
+    ),
+    (
+        "finemapping-null-pip",
+        validate_finemapping_dataframe,
+        pd.DataFrame({"pos": [100, 200], "pip": [0.5, None]}),
+        ["pip"],
+    ),
+    (
+        "finemapping-null-pos",
+        validate_finemapping_dataframe,
+        pd.DataFrame({"pos": [100, None], "pip": [0.5, 0.6]}),
+        ["pos"],
+    ),
+    (
+        "genes-null-start",
+        validate_genes_dataframe,
+        _genes(
+            chr=["1", "1"], start=[1000, None], end=[2000, 3000], gene_name=["A", "B"]
+        ),
+        ["start"],
+    ),
     (
         "genes-missing-all",
         validate_genes_dataframe,