diff --git a/CONTEXT.md b/CONTEXT.md
index 14189681..1435ea6a 100644
--- a/CONTEXT.md
+++ b/CONTEXT.md
@@ -66,7 +66,7 @@ The reference cohort matched to a target Property, used by **EPC Prediction** fo
 _Avoid_: neighbours, similar properties, peer set
 
 **EPC Prediction**:
-Producing a Property's `EpcPropertyData` picture from its **Comparable Properties** when it has no EPC (~30% of UK homes, typically long-tenure). **Deterministic** neighbour synthesis (k-NN-style — *not* ML; no trained model): take the cohort **mode** for the homogeneous categoricals (wall / roof / floor construction + insulation, construction age band), copy a single representative comparable's **structure** wholesale (building parts, per-window dimensions + orientations, floor dimensions) so the picture stays internally consistent for the calculator, then apply **Landlord Overrides** and the known inputs on top. The result is scored through **SAP10 Calculation** like any other **Effective EPC**, so a predicted Property flows through Rebaselining, Bill Derivation, and Modelling unchanged — marked as predicted so the UI can flag it. The same cohort machinery also produces **EPC Anomaly Flags** for Properties that *do* have an EPC. A future learned-weighting refinement is possible but separate, as with the calculator's ML residual head.
+Producing a Property's `EpcPropertyData` picture from its **Comparable Properties** when it has no EPC (~30% of UK homes, typically long-tenure). **Deterministic** neighbour synthesis (k-NN-style — *not* ML; no trained model): take the cohort **mode** for the homogeneous categoricals (wall / roof / floor construction + insulation, construction age band), copy a single representative comparable's **structure** wholesale (building parts, per-window dimensions + orientations, floor dimensions) so the picture stays internally consistent for the calculator, then apply **Landlord Overrides** and the known inputs on top. The result is scored through **SAP10 Calculation** like any other **Effective EPC**, so a predicted Property flows through Rebaselining, Bill Derivation, and Modelling unchanged — held in a **distinct predicted-EPC slot** that coexists with any lodged EPC (so provenance is structural and the UI can flag it; see ADR-0031). A **known property type is required** — the hard cohort filter (a flat is never sized from houses) — supplied by a **Landlord Override** (or, later, an Ordnance Survey lookup); a Property whose property type is genuinely unknown is **gated out**, never predicted from a mixed-type cohort and never given a national default. The same cohort machinery also produces **EPC Anomaly Flags** for Properties that *do* have an EPC. A future learned-weighting refinement is possible but separate, as with the calculator's ML residual head.
 _Avoid_: EpcPredictionService (no "service" suffix — name the operation), ML prediction (it is deterministic), EPC estimation
 
 ### Survey documents
diff --git a/docs/adr/0031-epc-prediction-production-wiring.md b/docs/adr/0031-epc-prediction-production-wiring.md
new file mode 100644
index 00000000..d6b21f35
--- /dev/null
+++ b/docs/adr/0031-epc-prediction-production-wiring.md
@@ -0,0 +1,87 @@
+# EPC Prediction production wiring
+
+ADR-0029 settled *how* EPC Prediction synthesises a Property's `EpcPropertyData`
+from its **Comparable Properties**; this records *how it wires into the running
+pipeline* — where estimation runs, how a predicted EPC is stored and told apart
+from a lodged one, and which Properties are eligible. Resolved in a
+grill-with-docs session after the algorithm + validation harness were built and
+the accuracy backlog (#1222–1228) closed.
+
+## Status
+
+Accepted (design). Refines ADR-0029 decision 3. Implementation pending (slice 5).
+
+## Decisions
+
+### 1. Estimation runs in Ingestion — the #1227 "shift to Modelling" is dropped
+
+The cohort fetch + predict happens in `IngestionOrchestrator`, when
+`epc_fetcher.get_by_uprn` returns `None` — upholding ADR-0029 decision 3. A
+design note from issue #1227 had proposed moving estimation (and its distance
+calcs) into `ModellingOrchestrator`; that is reversed here. Ingestion is already
+the EPC-acquisition phase and *already resolves the Property's coordinates*
+(`spatial`), so it can run the geo-weighted prediction with no new IO surface;
+the First Run stages communicate **only through persisted state** (the pipeline
+threads just `property_ids`, each stage reloads the `Property`), so a prediction
+produced in Modelling would either have to be persisted there anyway or
+recomputed every run. No rationale for the Modelling shift survived review.
+Baseline and Modelling stay untouched — they read a populated `effective_epc`.
+
+### 2. The predicted EPC is persisted in a distinct slot, never overwriting the lodged one
+
+Because stages communicate via persisted state, the prediction **must be saved**
+for Modelling to see it — in-memory-only would never reach stage 3. It is stored
+as a **distinct predicted-EPC slot** on the Property (the EPC table reused with a
+`source` discriminator — `lodged` / `predicted`), so a lodged EPC and a predicted
+EPC can **coexist** on one Property. Coexistence is load-bearing: it is what lets
+the same cohort machinery produce **EPC Anomaly Flags** for Properties that *do*
+have an EPC (the dual-use named in ADR-0029), and it means a later real-EPC fetch
+fills its own slot without the predicted one muddying provenance. Rejected: a
+single EPC slot with an `is_predicted` flag — it cannot hold both, so it forecloses
+anomaly detection and makes "lodged later arrives" ambiguous.
+
+### 3. Provenance is structural — on the Property, not on `EpcPropertyData`
+
+`EpcPropertyData` gains no `predicted` / `source` field. Which slot the picture
+came from *is* its provenance. `Property.effective_epc` / `source_path` gain a
+`"predicted"` branch, used only when there is no lodged EPC **and** no Site Notes
+(the existing precedence is unchanged; a real source always wins). The
+**Validation Cohort** then excludes any Property whose `effective_epc` resolves
+via the predicted slot — it has no same-spec lodged ground truth — and the UI
+flags it as predicted. Keeping `EpcPropertyData` clean means every downstream
+consumer (calculator, generators, bill derivation) is unchanged and oblivious to
+how the picture was sourced, exactly as for Landlord-Override and Reduced-Field
+pictures.
+
+### 4. Slice 5 is gap-fill only; always-predict (anomaly) is a follow-on
+
+Prediction runs only when `epc is None`. Predicting for *every* Property to
+compare against its lodged EPC (EPC Anomaly Flags) is real and the slot model
+supports it, but it triples the ingestion cohort IO and needs its own
+comparison + divergence-threshold + UI surface — so it does not ride in on the
+wiring slice. The predicted-EPC slot and the `ComparableProperties` repository
+this slice introduces are exactly what the anomaly capability reuses.
+
+### 5. A known property type is required — eligibility is gated, never defaulted
+
+A `PredictionTarget` needs `postcode` (from `PropertyIdentity`), `coordinates`
+(geospatial), and `property_type` + `built_form` + `wall_construction` from
+**Landlord Overrides**. `property_type` is the **hard** cohort filter (a flat
+must not be sized from houses), so it is a **required input**: a Property whose
+property type is genuinely unknown is **gated out** before prediction — flagged
+un-predictable, never predicted from a mixed-type cohort and never given a
+national default. (An Ordnance Survey `postcode_search` source can supply
+property type more broadly than landlord input does; wiring it is a later
+enhancement that widens the eligible population — out of scope here.) The
+`ComparableProperties` repository port deferred by ADR-0029 is built in this
+slice: it owns the cohort IO (postcode search → per-cert fetch → UPRN→coordinate
+resolution) and returns candidate `Comparable`s for the domain
+`select_comparables` to filter.
+
+## Consequences
+
+- A schema change: the EPC store gains a `source` discriminator (or equivalent),
+  and the Property repository a `get_predicted_for_property` read.
+- Slice 5's reach is bounded to EPC-less Properties with a landlord-supplied
+  property type. The `postcode_search` integration is the lever to broaden it.
+- `EpcPropertyData` stays unchanged, so no downstream consumer is touched.