diff --git a/scripts/validate-metamodel.rb b/scripts/validate-metamodel.rb
index c497aff..c236f7f 100755
--- a/scripts/validate-metamodel.rb
+++ b/scripts/validate-metamodel.rb
@@ -496,17 +496,22 @@ class ArtifactIndexGenerator
       output: '11-risks-and-technical-debt/generated/doc-11001-risks.adoc',
       anchor: 'risks',
       title: 'Risks',
-      cols: '1,3,1,1,1,3',
-      columns: ['ID', 'Risk', 'Probability', 'Impact', 'Priority', 'Mitigation/action'],
+      cols: '1,3,1,1,1,2,2',
+      columns: ['ID', 'Risk', 'Probability', 'Impact', 'Priority', 'Affects', 'Mitigation/action'],
       row: lambda do |artifact, helper|
         fields = helper.definition_table_fields(artifact)
         [
           helper.artifact_link(artifact, label: helper.short_id(artifact.metadata['id'])),
           helper.cell(artifact.metadata['title']),
-          helper.cell(fields['Likelihood']),
-          helper.cell(fields['Impact']),
-          helper.cell(fields['Priority']),
-          helper.relation_targets(artifact, 'affects')
+          # The assessment fields are prose in the source artifact; only their
+          # verdict fits a register column. See #summary_cell.
+          helper.summary_cell(fields['Likelihood']),
+          helper.summary_cell(fields['Impact']),
+          helper.summary_cell(fields['Priority']),
+          # What the risk endangers, and what is being done about it. These are
+          # opposites and had shared one column headed "Mitigation/action".
+          helper.relation_targets(artifact, 'affects'),
+          helper.incoming_relation_sources(artifact, 'mitigates')
         ]
       end
     }
@@ -1030,6 +1035,27 @@ def relation_targets(artifact, type)
     matches.map { |relation| artifact_ref(relation['target']) }.join(" +\n")
   end
 
+  # The mirror of #relation_targets: who points *at* this artifact with the given
+  # relation type. Relations are only ever authored outgoing, so a question like
+  # "what mitigates this risk?" can only be answered from the other side.
+  def incoming_relation_sources(artifact, type)
+    id = artifact.metadata['id']
+    return '-' unless id
+
+    sources = @artifacts_by_id.values.select do |candidate|
+      next false unless candidate.metadata
+
+      Array(candidate.metadata['relations']).any? do |relation|
+        relation['type'] == type && relation['target'] == id
+      end
+    end
+    return '-' if sources.empty?
+
+    sources.sort_by { |source| source.metadata['id'].to_s }
+           .map { |source| artifact_link(source) }
+           .join(" +\n")
+  end
+
   def derived_from_cell(entries)
     origins = Array(entries).compact
     return '-' if origins.empty?
@@ -1057,6 +1083,21 @@ def cell(value)
     text.gsub('|', '\|').gsub("\n", ' ')
   end
 
+  # For narrow index columns fed from prose fields. An assessment reads
+  # "Medium. It takes one lapse of attention while transcribing." — the verdict
+  # belongs in the register, the reasoning belongs in the artifact. Without this,
+  # writing a normal sentence in the source silently wrecks the generated table,
+  # and the author only finds out by rendering it.
+  #
+  # Only a sentence boundary counts: "Medium (rises when hosted)" survives whole.
+  def summary_cell(value)
+    text = value.to_s.strip
+    return '-' if text.empty?
+
+    verdict = text[/\A(.+?)\.(?:\s|\z)/, 1]
+    cell(verdict && !verdict.empty? ? verdict : text)
+  end
+
   private
 
   def derived_from_entry(entry)
diff --git a/skills/risk/SKILL.md b/skills/risk/SKILL.md
index ae2db4e..9d775b5 100644
--- a/skills/risk/SKILL.md
+++ b/skills/risk/SKILL.md
@@ -132,7 +132,9 @@ Use these templates:
 - The impact is tied to architecture outcomes, quality scenarios, delivery,
   operations, security, compliance, or stakeholder value.
 - Likelihood, impact, priority, timeframe, and confidence are qualitative and
-  reviewable.
+  reviewable, and each leads with its verdict. The generated risk register
+  carries only the first sentence of likelihood, impact and priority; the
+  reasoning belongs in a paragraph below the assessment table.
 - Mitigations, monitoring, or review actions are realistic and owned.
 - Assumptions and open questions are explicit.
 - Risk acceptance remains a human decision.
diff --git a/templates/risk.adoc b/templates/risk.adoc
index 268b2c5..a19e95a 100644
--- a/templates/risk.adoc
+++ b/templates/risk.adoc
@@ -32,6 +32,11 @@ Because <cause>, <event> may occur, leading to <impact>.
 
 === Assessment
 
+Likelihood, impact and priority are lifted into the generated risk register,
+which has one narrow column each. Only the first sentence is carried over, so
+lead with the verdict and put the reasoning in a paragraph below the table —
+`Medium. It only takes one lapse of attention.` reaches the register as `Medium`.
+
 [cols="1,3", options="header"]
 |===
 | Field | Value
@@ -42,6 +47,8 @@ Because <cause>, <event> may occur, leading to <impact>.
 | Confidence | Low, medium, or high.
 |===
 
+Why the rating is what it is, in as many sentences as it needs.
+
 === Mitigation Options
 
 * Mitigation option.
diff --git a/test/validate_metamodel_test.rb b/test/validate_metamodel_test.rb
index d259bce..9262009 100644
--- a/test/validate_metamodel_test.rb
+++ b/test/validate_metamodel_test.rb
@@ -572,6 +572,41 @@ def test_artifact_index_output_is_deterministic
     assert_includes first, '| xref:q-arch-001[Should the valid fixture demonstrate ADR provenance?]'
   end
 
+  def test_risk_index_separates_what_a_risk_affects_from_what_mitigates_it
+    # Given: the toolkit's own risks, which R-001 affects a quality scenario and
+    # ADR-007 mitigates
+    validator = MetamodelValidator.new(
+      root: ROOT,
+      docs_dir: ROOT.join('src/docs'),
+      relations_schema: SCHEMA
+    )
+    artifacts = validator.validate
+    generator = ArtifactIndexGenerator.new(root: ROOT, docs_dir: ROOT.join('src/docs'))
+    definition = ArtifactIndexGenerator::INDEX_DEFINITIONS.fetch('Risk')
+    output_path = ROOT.join('tmp/test-doc-11001-risks.adoc')
+
+    # When: the risk register is rendered
+    rendered = generator.render(artifacts, definition, output_path)
+
+    # Then: the two are separate columns, and the mitigation column is fed from
+    # the incoming mitigates relation rather than from what the risk endangers
+    assert_includes rendered, '| ID | Risk | Probability | Impact | Priority | Affects | Mitigation/action'
+    assert_includes rendered, 'xref:qs-004-ai-suggestion-reviewability[QS-004-ai-suggestion-reviewability]'
+    assert_includes rendered, 'xref:adr-007-artifact-provenance-metadata[ADR-007-artifact-provenance-metadata]'
+  end
+
+  def test_risk_index_reduces_a_prose_assessment_to_its_verdict
+    # Given: an assessment field written as a normal sentence, and one written
+    # with a parenthetical that is not a sentence boundary
+    helper = ArtifactRenderHelper.new(ROOT.join('tmp/test-summary-cell.adoc'), [])
+
+    # When/Then: only a sentence boundary truncates; everything else survives
+    assert_equal 'Medium', helper.summary_cell('Medium. It takes one lapse of attention')
+    assert_equal 'Medium (rises when hosted)', helper.summary_cell('Medium (rises when hosted)')
+    assert_equal 'Low, medium, or high', helper.summary_cell('Low, medium, or high')
+    assert_equal '-', helper.summary_cell('')
+  end
+
   def test_open_questions_index_output_is_deterministic
     # Given: the arc42 source tree with open questions
     generator = OpenQuestionsIndexGenerator.new(