diff --git a/MARS language/MARS-003-model-errors-with-coeffs.md b/MARS language/MARS-003-model-errors-with-coeffs.md
new file mode 100644
index 0000000..77c5c9b
--- /dev/null
+++ b/MARS language/MARS-003-model-errors-with-coeffs.md	
@@ -0,0 +1,105 @@
+# MARS Language Decision Record 003: Model Error Coefficients Index 
+
+## Status
+[<s>Proposed</s> | **Accepted** | <s>Deprecated</s> | <s>Superseded by [ADR-XXX]</s>]
+
+## Last Updated
+2026-01-09
+
+## Context
+
+Initial request from RD
+
+> As we are developing weak-constraint 4D-Var further, I would like to update the way we define and archive the model error fields.
+> At the moment, `type=me` is available for class LWDA and `type=eme` is available for class ELDA. They are used to archive 3D-fields for a deterministic or an ensemble of weak-constraint 4D-Var.
+>
+> We are developing a new formulation where we would need to archive a set of 3D-fields, instead of a single 3D-fields per physical variable. I think the best way to implement this would be to have a new GRIB variable in the template (i.e. a metadata that contains an integer). This should be then linked to a new MARS key, similar to what is done with the "member" key.
+> Let me know what you think, happy to discuss this further and give more details/motivations if needed.
+
+Assessment by Data Governance
+
+Essentially, instead of having a single value for a model error (of temperature for instance, but it could be any other parameter),
+the model error is now a series E1*sin(...) + E2*cos(...) + E3*sin(...) + E4*cos(...) + ..... instead of being a single value E.
+We need to store the coefficients E1, E2, E3, E4, etc. For now the truncation is at E3 but it could be anything.
+We can't use MARS number  to index these coefficients because number  is rightfully used for ensemble number with type=eme.
+
+This development is needed for 50r1.
+
+### Options Considered
+
+Option 1
+
+We propose a new MARS keyword, called `coefficient` or `coeffIndex`  to keep it generic so that it could be used in other context than to capture the n-th model error coefficient. 
+An immediate implementation is available by extending the currently used "model error" local section 39 (and 25 for non deterministic).
+
+```bash
+# Local definition 39: 4DVar model errors for long window 4Dvar system (inspired by local def 25)
+ 
+unsigned[1] componentIndex  : dump;
+alias mars.number=componentIndex;
+unsigned[1] numberOfComponents  : dump;
+alias totalNumber=numberOfComponents;
+unsigned[1] modelErrorType  : dump;
+ 
+alias local.componentIndex=componentIndex;
+alias local.numberOfComponents=numberOfComponents;
+alias local.modelErrorType=modelErrorType;
+ 
+# Hours
+unsigned[2] offsetToEndOf4DvarWindow : dump;
+unsigned[2] lengthOf4DvarWindow : dump;
+alias anoffset=offsetToEndOf4DvarWindow;
+```
+
+We could create a new local section from 39 by adding 2 new keys and alias it to a mars keyword:
+
+```bash
+unsigned[1] fourierCoefficientIndex : dump;
+unsigned[1] numberOfFourierCoefficients  : dump;
+alias mars.coeffindex=fourierCoefficientIndex;
+```
+
+For now, this new keyword will be solely used in the context of model errors (type=me and type=eme).
+Those are expert fields needed for restarting model runs but are of little use for the vast majority of people.
+
+This needs to be implemented in metkit as well and a new layout axis needs to be implemented in the MARS server.
+These coefficients are not planned to be disseminated or made available to external users. They will likely be plotted and consumed by evaluation tools.
+
+### Analysis
+
+In terms of the information stored in the GRIBs, we did not consider options outside of option 1.
+
+We had a wide ranging discussion about the options of `coefficient`, `coeffindex`, `component`, `componentindex` and other options. This settled on `coeffindex` fundamentally because:
+
+* The value is not a coefficient. 
+* `component` is already used in a related context
+
+## Decision
+
+Introduce a new MARS keyword, `coeffindex`, to identify which coefficient is being
+stored as part of a series.
+
+Actions required
+
+ * Implement encoding in eccodes
+ * New MARS keyword support in metkit
+ * New layout axis, and updated buildRules in MARS server
+ * Updated schema required for output into FDB
+ * review MARS client rules + metkit language rules (remove from inherited requests)
+
+### Related Decisions
+
+## Consequences
+
+No wider consequences expected.
+
+## References
+
+## Authors
+- Manuel Fuentes
+- Sebastien Villaume
+- Robert Osinski
+- Mohamed Azhar
+- Simon Smart
+- Emanuele Danovaro
+- Paul Dando