Add ADR for temporal system runner

Author: marmurar

docs/architecture/README.md

@@ -48,3 +48,4 @@ Jano is organized as a set of layers that should remain separable:
 - [ADR 0001: Keep the splitter model-agnostic](adrs/0001-keep-splitter-model-agnostic.md)
 - [ADR 0002: Keep runner reporting data-first](adrs/0002-runner-reporting-data-first.md)
 - [ADR 0003: Keep manual fold iteration public](adrs/0003-manual-fold-iteration-remains-public.md)
+- [ADR 0004: Add a temporal system runner above the core](adrs/0004-temporal-system-runner.md)

docs/architecture/adrs/0004-temporal-system-runner.md

@@ -0,0 +1,52 @@
+# ADR 0004: Add a Temporal System Runner Above the Core
+
+## Status
+
+Accepted.
+
+## Context
+
+Jano originally focused on temporal partitioning, simulation and model execution
+through `WalkForwardRunner`. That covers sklearn-like estimators well, but some
+real systems update without fitting a classic model:
+
+- retrieval indexes can be refreshed,
+- prompts can be revised,
+- routing policies can change,
+- fine-tuning jobs can be wrapped in an operational update step.
+
+Those workflows still need fold geometry, retraining or refresh decisions and
+structured outputs, but they do not fit the `fit()` / `predict()` contract cleanly.
+
+## Decision
+
+Jano adds a `TemporalSystemRunner` layer for updateable systems.
+
+The runner:
+
+- consumes temporal folds produced by the existing simulation layer,
+- calls a user-provided system `update()` step when policy says to update,
+- calls a user-provided system `evaluate()` step on each fold,
+- returns structured, data-first results.
+
+The new layer is intentionally above the core splitter and simulation primitives.
+It does not change fold geometry, and it does not move update logic into the
+splitter or planning layers.
+
+## Consequences
+
+- Jano can model more than sklearn-style estimators without changing the splitter.
+- LLM, RAG and prompt-refresh workflows can reuse the same temporal machinery.
+- The runner surface grows, but the core temporal contract stays stable.
+- Agent-facing and notebook-facing workflows can consume structured update and
+  evaluation results without parsing HTML or bespoke logs.
+
+## Invariants
+
+- `TemporalBacktestSplitter` remains model-agnostic.
+- `plan()` still describes geometry only.
+- `TemporalSimulation` still materializes temporal folds.
+- `TemporalSystemRunner` stays above the core primitives and does not modify
+  partition semantics.
+- Updateable systems are modeled through user-provided callables or protocols,
+  not built-in LLM-specific assumptions.