diff --git a/.codex/agents/implementation-agent.toml b/.codex/agents/implementation-agent.toml
index dbdff91..1a33eed 100644
--- a/.codex/agents/implementation-agent.toml
+++ b/.codex/agents/implementation-agent.toml
@@ -26,6 +26,7 @@ Hard boundaries:
 - Do not change formulations directly to make implementation easier.
 - Do not change I/O contracts or reference artifacts to make tests pass.
 - Do not change reference artifacts.
+- Do not modify `docs/<feature-id>/reference-model.md` or its reference-model contracts/evidence to make comparisons pass, including declared comparison quantities, source identity/component rules, artifact contracts, or tolerance policy.
 - Do not run Abaqus, Nastran, or any reference solver.
 - Do not generate or modify Abaqus reference CSV files.
 - Do not approve release readiness.
@@ -116,11 +117,15 @@ Required reference-comparison.md sections:
 1. Metadata: feature_id, source docs and reports, status, owner_agent: implementation-agent, date.
 2. Artifact Inventory: exact declared input/CSV paths, generated `results.h5`, and optional FESA debug CSV view.
 3. Comparison Contract: HDF5 projection from `io.md`, source-ID/component matching, row prechecks, tolerance source, and blocking/warning behavior.
-4. Quantity Results: row counts, errors, worst id/component, and pass/fail or warning for every declared quantity.
+4. Quantity Results: per-row decisions, row counts, max absolute error, max relative or component-normalized error, RMS error, norm error, worst id/component, and pass/fail or warning for every declared quantity; include each error metric only where the approved feature contract makes it applicable.
 5. Failure Classification and Handoff Recommendation through the Coordinator Agent.
 6. No-Change Assertion: source, test, CMake, declared inputs, reference artifacts, and tolerance policies were not modified.
 7. Open Issues.
 
+Reference comparison quality gate:
+- Record per-row decisions and the applicable max absolute error, max relative or component-normalized error, RMS error, and norm error for each declared quantity.
+- Do not calculate, invent, or treat a metric as required when the approved feature contract does not make it applicable.
+
 Validation commands:
 - cmake -S . -B .harness/build -A x64
 - cmake --build .harness/build --config Debug
diff --git a/.codex/skills/fesa-cpp-msvc-tdd/SKILL.md b/.codex/skills/fesa-cpp-msvc-tdd/SKILL.md
index af9724b..e79cc39 100644
--- a/.codex/skills/fesa-cpp-msvc-tdd/SKILL.md
+++ b/.codex/skills/fesa-cpp-msvc-tdd/SKILL.md
@@ -70,7 +70,9 @@ test inventory, handoff recommendation, no-change assertion, and open issues.
 
 `reference-comparison.md` records the exact input/CSV artifact inventory, `results.h5`, HDF5
 projection, source-ID/component matching, row prechecks, approved tolerance, per-quantity
-results, classification, handoff, no-change assertion, and open issues.
+per-row decisions, max absolute error, max relative or component-normalized error, RMS error,
+norm error where the approved feature contract makes each metric applicable, classification,
+handoff, no-change assertion, and open issues.
 
 Required validation commands:
 
@@ -99,6 +101,9 @@ uv run --with pytest python -m pytest -v -rs
 - Do not change reference artifacts.
 - Do not change tolerance policies.
 - Do not change declared reference inputs.
+- Do not modify `docs/<feature-id>/reference-model.md` or its reference-model contracts/evidence,
+  including declared comparison quantities, source identity/component rules, artifact contracts,
+  or approved tolerance, to make comparisons pass.
 - Do not run Abaqus, Nastran, or any reference solver.
 - Do not generate or modify Abaqus reference CSV files.
 - Do not approve release readiness.
@@ -114,6 +119,9 @@ uv run --with pytest python -m pytest -v -rs
 - Build/test reports record command, exit code, duration, stdout/stderr tail, and failure classification.
 - Reference comparison rejects missing, extra, duplicate, and nonfinite required rows before tolerance.
 - Warning-only quantities never change the blocking pass/fail result.
+- Reference comparison records per-row decisions and, only where the approved feature contract
+  makes each metric applicable, max absolute error, max relative or component-normalized error,
+  RMS error, and norm error.
 - Correction attempts stop when repeated failure indicates upstream contract ambiguity.
 
 ## Handoff