diff --git a/conformance/fixtures/invalid/verbose-prediction-too-confident.kpack/PACK.yaml b/conformance/fixtures/invalid/verbose-prediction-too-confident.kpack/PACK.yaml
new file mode 100644
index 0000000..cc489c3
--- /dev/null
+++ b/conformance/fixtures/invalid/verbose-prediction-too-confident.kpack/PACK.yaml
@@ -0,0 +1,4 @@
+name: verbose-prediction-too-confident
+version: 2026.05.09
+domain: test/conformance
+author: KP Conformance Suite
diff --git a/conformance/fixtures/invalid/verbose-prediction-too-confident.kpack/claims.md b/conformance/fixtures/invalid/verbose-prediction-too-confident.kpack/claims.md
new file mode 100644
index 0000000..35da932
--- /dev/null
+++ b/conformance/fixtures/invalid/verbose-prediction-too-confident.kpack/claims.md
@@ -0,0 +1,30 @@
+<!-- KP:1 — Knowledge Pack Format
+Claims: - [ID] assertion {confidence|type|evidence|date|depth|nature} context
+  Positions 5-6 optional. Verbose named-field syntax also accepted.
+Types: o=observed r=reported c=computed i=inferred
+Depth: assumed | investigated | exhaustive (optional, position 5)
+Nature: judgment | prediction | meta (optional, position 6; omitted=factual)
+Relations: →supports ⊗contradicts ⊗!error ⊗~tension ←requires ~refines ⊘supersedes ↔see_also
+Files: evidence.md=sources history.md=past entities.md=graph
+-->
+---
+pack: verbose-prediction-too-confident | v: 2026.05.09 | domain: test/conformance
+confidence: simple | normalized
+---
+
+# Verbose Prediction Too Confident Fixture
+
+> INVALID: C001 uses the verbose named-field syntax with nature=prediction and
+> confidence 0.97, exceeding the 0.95 cap. Violates SC-12. This is the verbose
+> counterpart of prediction-too-confident.kpack (dense syntax), so SC-12 is
+> exercised on both metadata forms.
+
+## Claims
+
+- **[C001]** The widget market will reach $5B in fiscal 2027
+  `confidence: 0.97 | type: inferred | evidence: E001 | since: 2026-05-09 | depth: investigated | nature: prediction`
+  A future-tense claim with confidence above the 0.95 prediction cap.
+
+- **[C002]** Plausibly-bounded prediction (sibling control)
+  `confidence: 0.80 | type: inferred | evidence: E001 | since: 2026-05-09 | depth: investigated | nature: prediction`
+  This one is within the 0.95 cap and is therefore valid.
diff --git a/conformance/fixtures/invalid/verbose-prediction-too-confident.kpack/evidence.md b/conformance/fixtures/invalid/verbose-prediction-too-confident.kpack/evidence.md
new file mode 100644
index 0000000..69ac730
--- /dev/null
+++ b/conformance/fixtures/invalid/verbose-prediction-too-confident.kpack/evidence.md
@@ -0,0 +1,8 @@
+# Evidence — Verbose Prediction Too Confident Fixture
+
+## E001
+> **type:** document | **captured:** 2026-05-09
+> **source:** Test data
+
+Single evidence record; the violation is in the prediction confidence cap
+(SC-12), here expressed via the verbose named-field syntax.
diff --git a/conformance/run.py b/conformance/run.py
index be287f2..ab1793e 100644
--- a/conformance/run.py
+++ b/conformance/run.py
@@ -61,6 +61,7 @@
     "dangling-relation-target.kpack",
     "wrong-pack-name.kpack",
     "prediction-too-confident.kpack",
+    "verbose-prediction-too-confident.kpack",
 ]
 
 # Expected error categories for invalid fixtures.
@@ -75,6 +76,7 @@
     "dangling-relation-target.kpack": "SC-05",
     "wrong-pack-name.kpack": "SC-07",
     "prediction-too-confident.kpack": "SC-12",
+    "verbose-prediction-too-confident.kpack": "SC-12",
 }
 EXAMPLE_ORDER = [
     "solar-energy-market.kpack",
@@ -310,6 +312,20 @@ def validate_pack(pack_dir: Path) -> list[Err]:
             if type_s not in {"o", "r", "c", "i"}:
                 errs.append(Err("parse", f"invalid claim type '{type_s}' for {cid} (expected o/r/c/i)"))
 
+            # Position 4 (date): ISODate. Grammar is YYYY-MM-DD; validate the
+            # format when the slot is non-empty (empty interior slots are valid).
+            date_s = parts[3].strip()
+            if date_s and not re.fullmatch(r"\d{4}-\d{2}-\d{2}", date_s):
+                errs.append(Err("parse", f"invalid date '{date_s}' for {cid} (expected YYYY-MM-DD)"))
+
+            # Position 5 (depth) intentionally NOT enum-validated here. The spec
+            # (CORE.md) lists assumed/investigated/exhaustive, but a large share
+            # of the shipped corpus (system + grounding packs) uses additional
+            # depth values such as `practitioner`/`confirmed`. Enforcing the
+            # closed set would flag hundreds of production claims — that is a
+            # spec-vs-corpus reconciliation, not a parser quick-win, so it is
+            # left for a dedicated decision rather than enforced here.
+
             # Empty evidence position → syntactic error
             if not ev_s:
                 errs.append(Err("parse", f"empty evidence ref list for {cid}"))
@@ -350,6 +366,13 @@ def validate_pack(pack_dir: Path) -> list[Err]:
             if cm:
                 confidences.append((cid, float(cm.group(1)), cm.group(1)))
 
+            # SC-12: a prediction-nature claim must keep confidence ≤0.95.
+            # The dense branch enforces this at the position-6 nature slot; the
+            # verbose branch previously skipped it entirely.
+            nm = re.search(r"nature:\s*(\w+)", meta)
+            if cm and nm and nm.group(1) == "prediction" and float(cm.group(1)) > 0.95:
+                predictions_high_conf.append((cid, float(cm.group(1))))
+
             # Evidence refs
             em = re.search(r"evidence:\s*([^|]+)", meta)
             if em: